Consultative Group on Risk

Management

Governance of AI adoption in
central banks

January 2025

BIS Representative Office for the Americas

This publication is available on the BIS website (www.bis.org).

© Bank for International Settlements 2025. All rights reserved. Brief excerpts may be reproduced or
translated provided the source is stated.

ISBN 978-92-9259-831-0 (online)

Table of contents

Foreword ............................................................................................................................................................................................... 1

Executive summary ........................................................................................................................................................................... 2

1 Use of AI ............................................................................................................................................................................. 3
1.1 AI benefits for central banks and use cases ................................................................................................ 4

2 Risks and their impact ................................................................................................................................................... 7

2.1 Risks associated with the adoption of AI ..................................................................................................... 7
2.2 Broader considerations and impacts ........................................................................................................... 10

3 Risk management for AI ............................................................................................................................................. 12

3.1 Risk management strategy .............................................................................................................................. 12
3.2 Information security, privacy and cyber security risks .......................................................................... 15
3.3 Specific actions to mitigate GAI risks ........................................................................................................... 16

4 Governance ..................................................................................................................................................................... 19
4.1 Current industry frameworks .......................................................................................................................... 20
4.2 Proposed actions for AI governance at central banks .......................................................................... 21

5 Conclusions ..................................................................................................................................................................... 24

References .......................................................................................................................................................................................... 25

Appendix 1: Artificial intelligence definitions ....................................................................................................................... 30

Appendix 2: Summary of AI use cases described in Section 1.1 .................................................................................. 31

Appendix 3: Risks associated with AI ...................................................................................................................................... 32

Appendix 4: Review of current AI frameworks..................................................................................................................... 33

Appendix 5: Members of the Artificial Intelligence Task Force ..................................................................................... 45

Governance of AI adoption in central banks i

Foreword

Artificial intelligence (AI) presents huge opportunities for central banks. At the same time, its adoption
entails complex risk management challenges. The use cases for AI span a broad range of critical functions
of a central bank including data analysis, research, economic forecasting, payments, supervision and
banknote production. The adoption of AI presents new risks and can amplify existing ones. The potential
risks are wide-ranging and include those around data security and confidentiality, risks inherent to AI
models (eg “hallucinations”) and, importantly, reputational risks. The potential risk exposure for central
banks can be significant, owing to the criticality and sensitivity of the data they handle as well as their
central role in financial markets.
This report on the governance of AI adoption in central banks provides guidance on the
implementation of AI at central banks and proposes a governance and risk management framework. A
comprehensive risk management strategy can leverage existing risk management models and processes,
in particular the well established three lines of defence model. In incorporating the specific issues around
AI and its use cases, risk managers at central banks can make use of the frameworks proposed by a number
of international bodies. A good governance framework is key for adopting AI. The report proposes an
adaptive governance framework and recommends ten practical actions that central banks may want to
undertake as part of their journey in adopting AI.
The report is the outcome of work conducted by Bank for International Settlements (BIS) member
central banks in the Americas within the Consultative Group on Risk Management (CGRM), which brings
together representatives of the central banks of Brazil, Canada, Chile, Colombia, Mexico, Peru and the
United States. The Artificial Intelligence Task Force that prepared this report was co-led by Alejandro de
los Santos from the Bank of Mexico and Angela O'Connor from the BIS. The BIS Americas Office acted
as the secretariat.

Claudia Álvarez Toca Alexandre Tombini

Chair of the CGRM Chief Representative for the Americas
Bank of Mexico Bank for International Settlements

Governance of AI adoption in central banks 1

Executive summary

The adoption of artificial intelligence (AI) technologies in the financial sector may usher in a transformative
era for financial services, offering unprecedented opportunities for innovation, efficiency and customer
focus. As with any groundbreaking new technology, AI also introduces complex risk management
challenges for banks and central banks. Risks from the use of AI include but are not limited to operational,
information security, privacy and cyber security risks; information and communication technology (ICT)
risks; third-party risks such as external dependency; and risks inherent in AI models (eg “hallucinations”).
The materialisation of these risks can have significant reputational and financial impacts.
The identification and management of such risks can be complex owing to their variety and
interactions between different types. The analysis and management of these risks should therefore be
approached with a holistic view that incorporates a wide array of expertise to consider the full range of
risks and complex interactions.
The Bank for International Settlements Consultative Group on Risk Management (CGRM) set up
a task force to provide guidance to central banks on how AI is being used in different functions, and how
to organise and govern risk management related to AI adoption. This report provides specific suggestions
on how central banks can identify, analyse, report and manage risk associated with the adoption of AI
models and tools in their organisations. The suggestions are based on risk scenarios and risk management
practices developed in the central banking community and the financial sector at large, such as the three
lines of defence model. The report does not focus on the impact of AI on financial stability or the broader
economy, which are covered by other forums.
The report argues that central banks need to find a balance between fostering innovation using
AI and mitigating the different risks that this technology may generate. Good governance schemes for the
adoption of AI in the organisation, with a holistic view beyond technology, might help to achieve such a
balance.
The report is organised as follows. Section 1 begins with a brief overview of AI models, highlights
use cases from central bank publications and complements the discussion with answers provided by
members of the CGRM task force to a questionnaire on AI usage. Some of the benefits and uses identified
by the group include the automation of processes, analysis of large data sets and solving complex
problems. Central banks adopt AI to enhance efficiency, improve operational robustness and inform
decision-making in different areas of the organisation. This includes core functions such as economic
forecasting, payments, supervision and banknote production. Central banks are also exploring the use of
AI to provide customer and corporate services, for instance by using chatbots to answer enquires from
regulated entities or assist their own researchers. These applications demonstrate the potential of AI to
address complex challenges and support central banking operations.
Section 2 describes, from a holistic perspective, risks related to the adoption of AI for central
banks, covering both new risks raised by AI and existing risks that are amplified by its use. The section
introduces a taxonomy into strategic; operational (legal, compliance, process, people and capacity);
information security and cyber security; ICT; third party; AI model; environmental, ethical and social; and
reputational risks. It also highlights some considerations about generative artificial intelligence (GAI) and
the unintended consequences of its adoption, which can result in significant risk exposure for central
banks, if the associated risks are not well managed.
Section 3 provides some guidance for the implementation of AI models in central banks. The
report suggests a comprehensive risk management strategy that leverages existing models and processes.
It suggests updating the three lines of defence model with some considerations specific to the use of AI.
The section also suggests a process that central banks can employ to identify and analyse use cases,

2 Governance of AI adoption in central banks

initiatives or projects related to AI before its deployment. This process can be based on the frameworks
suggested by various standard-setting bodies, especially those related to information security and cyber
security. If central banks decide to rely on third-party AI services, tools, components or algorithms, a
mature third-party risk management model is essential. The use of GAI poses additional risks, such as the
interpretability of the results of such models, possible biases, limitations and robustness. This means that
GAI requires more human supervision and checks for legal implications than other technologies, as well
as other specific controls based on a holistic and multidisciplinary perspective.
Section 4 proposes a governance and risk management framework for AI based on concepts
developed in previous sections. Good AI governance is important not only for complying with national
strategies, laws or regulations but also for ensuring the alignment of AI uses with the organisation’s
strategy and objectives. Effective AI governance boosts efficiency and stimulates innovation while also
identifying and mitigating associated risks. The proposal presented in this section aims to balance both
the risks and opportunities offered by AI.
As an initial guide, the section presents current industry standards and suggestions for risk
management and governance frameworks that reflect key features and concerns already identified as
pertaining to the use of AI. These frameworks share common elements that can help central banks to
implement AI processes that are ethical, secure, transparent, explainable, reliable, responsible and comply
with data privacy. The safe and proper usage of AI across the central bank functions may demand changes
to existing risk management and governance frameworks. In this sense, central banks will eventually have
to consider a careful review of their current governance structures and risk management practices, while
embracing AI models and tools safely and efficiently. An adaptive AI governance framework may prove
helpful in this regard. Finally, this section proposes ten actions that have proven useful for central banks
on their journey to adopt AI: (i) establish an interdisciplinary AI committee; (ii) define principles for
responsible AI use; (iii) establish an AI framework and update existing guidance; (iv) maintain an AI tools
inventory; (v) map AI tools and stakeholders; (vi) perform a detailed assessment of risks and controls; (vii)
perform regular monitoring; (viii) report anomalies and incidents; (ix) develop and improve workforce skills;
and (x) perform ongoing reviews and adaptations to the framework.

1 Use of AI

The concept of artificial intelligence (AI) can be broad and mean different things in different contexts. A
number of different sources were considered as part of the process of arriving at a common definition of
AI for this report (Appendix 1). The definition adopted here is based on the widely used definition from
the Organisation for Economic Co-operation and Development (OECD), according to which:
“An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the
input it receives, how to generate outputs such as predictions, content, recommendations, or
decisions that can influence physical or virtual environments. Different AI systems vary in their
levels of autonomy and adaptiveness after deployment.”1
AI applications rely on different models, which perform operations based on rules, code,
knowledge and data (inputs) to learn and generate the required outputs. According to the definition of AI
above, the outputs can be predictions, content, recommendations or decisions. Some of these models are

1
OECD (2024a).

Governance of AI adoption in central banks 3

able to learn or act without human involvement (autonomy) and to continue evolving after they are
deployed (adaptiveness).2 Box A describes different AI models.

Box A

AI models

Machine learning (ML) is a subfield of artificial intelligence (AI). This technique gives systems the ability to “learn”
from the input data without being explicitly programmed to do so (ISO/IEC (2022)). ML models are widely used to
identify patterns that are not otherwise readily evident, mainly through statistical learning algorithms. ML models are
capable of improving (Sharma et al (2021)).

Deep learning is a subfield of ML. Deep learning models learn from large and complex inputs, using neural networks
with multiple neurons and hidden layers (Amazon AWS (2024)). This architecture allows deep learning models to be
more autonomous, by automatically identifying the most relevant features from the data. This allows for deep learning
models to be generalised, meaning that they can be used with different types of inputs or in different applications
(Sharma et al (2021)).

Generative AI (GAI) models are capable of generating original content in response to user-supplied inputs called
prompts. The outputs of these models can be text, images, audio, video or code. General awareness of AI increased
with the rise of GAI technologies in 2022, beginning with image generation and later with chat-based AI interfaces.

Natural language processing (NLP) models enable machines to understand natural language (linguistics) and
identify meaning and context. At the same time, these models can also generate natural text, such as phrases,
sentences or paragraphs with meaning (Khurana et al (2023)).

Large language models (LLMs) are a type of NLP trained on a very large number of parameters or quantity of data
using deep learning. Generative LLMs understand how languages work based on very large volumes of text input and
can themselves generate text based on this training (Chiarello et al (2024)).

Generative pre-trained transformers (GPTs) are a type of LLM that are able to consider the context and relationships
between words in entire sentences. They are capable of executing a wide range of language-related tasks (Yenduri et
al (2024)).

1.1 AI benefits for central banks and use cases

Central banks have been exploring AI applications, developing proof of concept solutions and, in some
cases, deploying AI applications.
AI can create opportunities by supporting human activities with the following benefits:
• Automation of business processes, optimising the use of resources and time, improving the
efficiency of repetitive or highly manual tasks and thereby increasing productivity.
• Swift analysis of large volumes of data enabling improvements in decision-making.
• Execution of processes that require the involvement of many people, allowing employees to
undertake other productive tasks.

2
OECD (2024b).

4 Governance of AI adoption in central banks

• Solving complex problems through the analysis of large volumes of data and the recognition of
patterns that, with other methods, might go unnoticed or require significant time to identify.
• Innovation fostered by the use of AI technologies, which can be applied in different business
processes and sectors including central banks.
AI use cases relevant to central banks are set out below. These use cases were identified from
papers authored within the central banking community and from responses to a questionnaire from Bank
for International Settlements Consultative Group on Risk Management (CGRM) member central banks. In
the former case, the report includes the corresponding references. The categories highlight the main
themes identified during the report’s development and are not exhaustive of all potential AI applications
within central banks. Appendix 2 provides a summary of these use cases.

Use case 1: Economic analysis, forecasting and policy analysis

The ability of AI to analyse large data sets and capture nonlinear relationships makes it highly effective for
economic analysis, forecasting and policy analysis.
• GDP nowcasting: central banks and organisations use machine learning algorithms to nowcast
GDP growth (Richardson et al (2021) and Dauphin et al (2022)).
• Inflation forecasting: machine learning has long been part of some organisations’ forecasting
procedures (Benford (2024), Buckmann et al (2023), Joseph et al (2022), Chakraborty and Joseph
(2017) and Burgess et al (2013)).
• Textual analysis: natural language processing (NLP) and large language models (LLMs) have
helped to enhance economic analysis and forecasting through textual analysis from newspapers,
reports and social media articles (Chen et al (2023), Denes et al (2022), Gascon and Werner (2022)
and Kalamara et al (2020)). Some central banks use NLP to analyse survey responses on economic
indicators and perform sentiment analysis on official publications.
• Policy analysis and implementation: some central banks use AI to develop models that
enhance policy processes. These models help to analyse and assess policy implications,
investment impacts, and impacts on the financial sector and the macroeconomy.

Use case 2: Payment systems

• Payment system enhancements: AI can detect abnormal transactions, strengthening the
functioning of payment systems (BIS (2024)). For example, using synthetic transaction data,
Phase 1 of the BIS Innovation Hub’s Project Aurora showed that machine learning algorithms
could detect money laundering networks more effectively than traditional methods
(BISIH (2023)). Some central banks employ artificial neural networks to identify anomalies in
payment patterns (Rubio et al (2020)).
• Research on payment systems: AI can be used to research standards and innovations that can
enhance the resilience and robustness of payment systems.

Use case 3: Regulation

• Regulatory complexity: some central banks use NLP to calculate complexity measures of
prudential and banking regulations (Amadxarif et al (2021)).
• Impact analysis: some central banks use AI to analyse the impact of bills and regulations through
web scraping, classification algorithms and similarity analysis. This provides timely alerts about
legal provisions that may require the attention of the central bank.

Governance of AI adoption in central banks 5

Use case 4: Supervision and oversight
• Supervision activities: some central banks have used machine learning, NLP and generative
artificial intelligence (GAI) to extract insights from supervisory agents’ information or to provide
insights for supervisors (Benford (2024), McCaul (2024) and ECB (2023)).
• Monitoring financial institutions: AI tools help assess and monitor financial institutions,
covering aspects like credit risk, exposure and portfolio risks (eg Devys and von Kalckreuth
(2022)).

Use case 5: Banknote production and distribution

• Banknote and coin production: central banks have used deep learning techniques to enhance
the efficiency of printing operations and banknote quality control (eg Kerdsri and Treeratpituk
(2022)).
• Forecasting banknote demand and distribution needs: AI is used to forecast cash flow and
plan production and distribution schedules for banknotes and coins.

Use case 6: Anomaly detection

AI can analyse vast data sets to detect unusual patterns quickly, flag anomalies for investigation and
support decision-making. Central banks can use AI for:
• Data quality enhancements: in 2022, the BIS published a compendium on the application of
data science and machine learning in central banks, including AI use cases for anomaly detection
to improve data quality (Araujo et al (2022)). Some other organisations and central banks have
employed machine learning to detect anomalies or incorrect reporting (Accornero and
Boscariol (2022), Cagala et al (2022), Faria da Costa et al (2022), Haghighi et al (2022) and Jiménez
and Serrano (2022)). Some other central banks use AI to detect anomalies in price data and
account balances – identifying outliers and unusual changes.
• Cyber security enhancements: the BIS Innovation Hub Nordic Centre is working on Project
Raven, which proposes a solution for central banks and regulatory authorities to assess cyber
security maturity in their financial sectors using AI (BISIH (2024)).

Use case 7: Risk assessment

AI excels at identifying patterns, analysing granular data and capturing nonlinear relationships for risk
assessment (Nistor (2023)). Key examples include:
• Early warning systems: AI can analyse diverse data sources to detect early signs of potential
risks or financial crises (Bluwstein et al (2020)).
• Stress testing: AI can generate synthetic data for scenario analysis and enhance stress testing
with complex simulations (Petropoulos et al (2019)).
• Risk event analysis: NLP is used to categorise risk events, identify trends and extract actionable
insights.

Use case 8: Customer and corporate services

• Research assistance: LLMs assist with summarising documents, providing coding support,
preparing slides, drafting reports and analysing frequently asked questions from the public
(Benford (2024)).

6 Governance of AI adoption in central banks

 • Public services: some central banks are using AI to implement chatbots that answer public
questions.
• Corporate services: some central banks use AI-powered chatbots to assist employees with
questions about processes, HR benefits and policies (Benford (2024)).
• Information technology: central banks use AI to enhance code development processes and
database management.
• Communications: some central banks are exploring the use of AI for sentiment analysis on
publications to ensure appropriate wording. AI also aids in producing official speeches with a
consistent institutional style. Additionally, central banks use AI to support communication
processes through document translation, grammar refinement and presentation transcription.

2 Risks and their impact

While there are significant benefits to the use of AI, its use also entails risks that must be identified,
analysed and mitigated. The following subsections outline the main risks associated with the adoption of
AI, provide examples of how to classify AI risks and propose methods for mitigating these risks.

2.1 Risks associated with the adoption of AI

The use of AI can introduce new risks and vulnerabilities and amplify existing risks. Like other emerging
technologies that access and process large volumes of data, AI carries a risk for users. The level of risk will
be determined by the access granted to information and data used by AI models and tools.
Below is a proposed classification of AI risks for central banks. Appendix 3 includes a summary
table of risks associated with the adoption of AI.

Proposed classification of AI risks for central banks Table 1

Type of risk Description

Absence of a clear strategy for the use of AI can lead to the application of AI in ways that negatively affect
Strategic the reputation of the organisation. For example, using GAI could raise questions around bias.
AI adoption could introduce additional complexity in the delivery of internal and external services,
exacerbating or amplifying a range of operational risks including:
Legal uncertainty. AI outputs may not meet legal obligations such as copyright protection.
Compliance. Lack of explainability or the “black box” nature of some AI models may exacerbate compliance
risks due to the difficulty of explaining outcomes or providing evidence on the reasoning behind a decision.
Processes. Failures or deficiencies in the design, implementation, or ongoing management of processes or
controls when incorporating AI tools.
Operational People. Inadequate digital skills among staff or the lack of a solid understanding of AI models and safe usage
of AI tools, which could place central banks at a strategic disadvantage or make them more vulnerable to
current threats. A lack of AI knowledge, specialised training and security awareness among central bank staff
can lead to greater reliance on AI providers. This dependence on third-party expertise may pose additional
operational risks, such as the inappropriate use or mishandling of data by those third parties.
Data governance. AI models are data intensive and leverage new data sources or introduce new ways to
use the existing ones, thereby increasing data governance challenges. In particular, the volume, velocity,
variety and quality of data may require intensified efforts for effective data governance.

Governance of AI adoption in central banks 7

 Resiliency. Increased use of AI can amplify traditional threats to business continuity such as geopolitical
conflicts or wars, civil conflicts, terrorism and other illicit activities including organised crime. This could
make it more difficult to prevent or respond to business disruptions. Awareness and prevention around AI-
related threats will be critical to foster resiliency.
Information security and privacy. GAI outputs may contain confidential data, leading to security breaches,
unauthorised access or privacy violations.3 It could produce incorrect results (integrity) or suffer disruptions
(availability). The use of sensitive information or critical assets may present a risk of data leakage, data
corruption, intellectual property rights violations and data privacy violations. Users can send sensitive and
proprietary information as part of prompts. These data could be stored by AI tools to continue training the
model and may be inadvertently leaked to other users.
Cyber security
New entry points. New AI systems and their integration with other central bank systems add entry points,
eg through third or nth parties, for cyber attacks expanding the attack surface. Novel attacks include prompt
Information injection, training data poisoning, and model denial of service and model theft.
security, privacy Data poisoning. The use of AI models could corrupt the original data, affecting its integrity, reliability and
and cyber security completeness.
Malicious activities. AI executes tasks as requested by its users, typically without questioning the context,
intentions or taking into account ethical considerations. Malicious actors or criminal groups can therefore
exploit AI to conduct malicious activities, both within and beyond cyberspace, including scams, fraud, deep
fake videos and disseminating misinformation. GAI tools may augment attackers’ capabilities, allowing them
to carry out more precise and faster attacks such as hacking, malware and phishing. LLMs are already able
to find new vulnerabilities in systems and write code to successfully exploit them. AI can facilitate new types
of scams and frauds, with attackers and fraudsters deploying more sophisticated techniques to target
individuals and organisations, for example, deepfakes (visual and audio content that appear real but were
generated with AI) using the image of central bank Governors have been reported lately.
ICT system risk. ICT risks are often originated from the ICT systems themselves, including sources such as
software bugs, hardware failures and inadequate system design. Challenges in ensuring compatibility
between new technologies and established ICT infrastructure can exacerbate these risks, potentially leading
to systems downtime, operational disruptions and reduced reliability, impacting the overall efficiency and
effectiveness of financial operations.
Information and ICT infrastructure and maintenance. Infrastructure failure, compromised information attributes or deficient
communication backups, which can impact business continuity, operations and resilience.
technology Functionality and permissions. AI systems with excessive functionality or permissions may make decisions
(ICT) and undertake actions with unintended consequences. For example, an AI tool for reading documents may
inadvertently be given permissions to delete documents as well.
Lack of governance. During the implementation and maintenance of AI systems throughout their life cycle,
a lack of governance might lead to inappropriate usage and introduce new vulnerabilities. AI systems might
be unavailable due to non-malicious causes (eg incorrect patch or human error). Incompatibility issues
between AI systems and legacy systems may use up significant resources.
Dependence on external AI models or tools developed by third-party providers have several implications for
central banks:
Third-party privacy and information security risks. If a vendor/provider has access to personal or sensitive
information, there are additional risks that must be considered, such as intentional or unintentional misuse
of data. Also, third-party vendors that store or transmit data through unregulated systems and networks, or
process and host data in certain countries, can expose the organisation to compliance risk.
Third party Third-party concentration risk. Dependency on a single or a few suppliers can create significant risks. Any
issues on the side of the supplier could affect the operations of organisations that depend on it. Additionally,
the lack of alternatives limits negotiating power and exposes the organisation to security and compliance
risks if the supplier fails to meet required standards.
Third-party cyber security risk. One of the most relevant risks is the impact on central bank processes caused
by a failure or cyber attack on these providers, which limits the response capacity of the central bank in the
event of an interruption or failure.

3
The General Data Protection Regulation (GDPR) is applicable to any company that processes the data of European Union (EU)
citizens or the personal data of subjects who are in the EU or the European Economic Area.

8 Governance of AI adoption in central banks

 Inflexibility. Dependency on third-party providers means that central banks cannot modify or adjust the
technologies according to the evolution of their business needs.
Interpretation and reliability
Imprecise results. The accuracy of AI models is highly dependent on the quality of data. Poor data quality 4
and poor prompting lead to poor analysis or predictions, biases and errors. In many cases, the data sources
used to train AI models to generate knowledge or make decisions may be of low quality or cover general
topics. Therefore, certain AI models may struggle to respond to a question or execute activities on very
specific, rare or exceptional topics. There are other causes of inaccuracy in the results related to the models
and data, such as overfitting,5 data sparsity and biases.
Reasoning errors and “hallucinations”. AI tools, especially GAI, may be susceptible to reasoning errors and
“hallucinations”, ie perceiving patterns or objects that do not exist, creating meaningless or inaccurate
results. Likewise, handling large quantities of data (big data) by AI could result in apophenia, ie “seeing
patterns where they do not actually exist, simply because enormous amounts of data can offer connections
that radiate in all directions” (Boyd and Crawford (2012)). Similarly, overfitting can lead to establishing
spurious relationships.
Model risk Repeatability issues. AI tools, especially GAI may not produce consistent responses over time, even when
given similar inputs.
Overconfidence in generated output. AI models and tools can generate what appears to be authoritative
output, potentially weakening due diligence efforts to confirm validity. This increases the exposure to error
risk in decision-making.
Transparency and accountability
Lack of transparency. Absence of public and detailed knowledge about the programming of AI models can
lead to governance and accountability issues for those who design or operate them. This includes a lack of
understanding or inappropriate controls of databases and information used to “train” the algorithms of AI
applications.
Self-modification of algorithms. GAI and ML algorithms can be programmed to automatically incorporate
different or additional data beyond their initial training data. These algorithms can initiate changes to
themselves, leading to the identification of new patterns or correlations with other data. However, this
“evolutionary nature” can also produce outcomes not intended by their designers.
Energy consumption and carbon emissions. The use of AI models, particularly GAI and LLMs, requires large
energy consuming technologies and demand significant computational resources, consuming considerable
energy and impacting the carbon footprint. Several factors affecting AI models’ carbon emissions. These
factors include the volume of training data, the cost of running the model, the location of the training server
Environmental, and its energy grid, the complexity of the training procedure and the hardware used.
ethical and social Limitations in understanding ethical and social context. When using AI, individuals interact with computer
systems whose “understanding of reality” is not comparable to that of a human. Therefore, AI tools may
sometimes provide inappropriate responses from an ethical or social point of view. Biases in the data used
to train a model can also result in unethical, discriminating or stigmatising outcomes.6
Reputational risks may be a consequence of the realisation of any of the above risks. AI-related controversies
Reputational can attract intense media scrutiny, amplifying negative impact on reputation.

Source: authors’ elaboration.

4
Poor data quality includes missing historical, unlabelled, biased, incomplete, noisy, untimely, inaccurate or inadaptable data.
5
Overfitting occurs when too many variables are included in the algorithm and it fits its training data too closely or even exactly,
resulting in a model that cannot make accurate predictions from any other data.
6
For instance, according to the literature on the ethical impact of data science, privacy is a concept that evolves alongside
technological and social changes. In this regard, people currently perceive privacy differently in public spaces, such as parks
and streets, compared with socially and legally designated private spaces, such as homes. Training technologies like ML, AI and
GAI with data and information from multiple perspectives or “dimensions” can reduce this risk. See Mulligan et al (2016).

Governance of AI adoption in central banks 9

2.2 Broader considerations and impacts

Differentiating between risks and other impacts is important when adopting AI models and tools. Risks
are identifiable challenges that can be anticipated and managed. In contrast, other considerations refer to
unexpected outcomes that may emerge due to the complex and dynamic nature of AI technologies.
Therefore, a comprehensive approach to AI adoption in central banking must encompass a thorough
assessment of both risks, as well as detect and incorporate other considerations that are a natural part of
using AI.
Examples of such other considerations include:
• Unplanned uses – GAI can help identify new issues in central banks’ processes that need to be
explored.
• Job impact – not all employees benefit from AI in their roles, nor is AI necessarily a solution to
addressing all business challenges.
• New activities – staff responsible for information, processes and technological tools may need
to interact given the new activities introduced by AI.
• Complexity challenges – the complexity of training or deploying AI models can be
underestimated, as can the effort needed to integrate new tools with legacy systems, especially
when the information categories of the systems differ. This can lead to delays rather than
potential productivity increases. Additionally, some processes will require specially trained or
personalised models, which may require significant resources for maintenance. Acquiring and
retaining more specialised skills can often be more difficult than continuing existing processes
that rely on analysts.
• Project management – AI tools are currently in constant evolution, which has made it difficult
to precisely programme their implementation and the pressure to rapidly adopt AI tools may
lead to rushed implementations, increasing the probability of errors or vulnerabilities.
• Impact on technological infrastructure – the widespread use of AI tools within central banks
could affect their technological infrastructure performance. AI tools, particularly GAI, require a
vast amount of energy and a lot of computing and storage resources for training and responding
to user queries.
• Dependency on AI – if AI models produce accurate and useful outcomes, they could quickly
become key elements of central banks’ processes, generating a new dependency on this type of
tool and introducing new costs.

10 Governance of AI adoption in central banks

 Box B

Examples of malicious use and failures

Below are some examples illustrating the risks of various artificial intelligence (AI) applications:

GPT – Derner and Batistič (2023) identified six main risks when using this technology: (i) generation of fraudulent
services; (ii) collection of harmful information; (iii) disclosure of private data; (iv) generation of malicious text; (v)
generation of malicious code; and (vi) production of offensive content. Computer systems are often designed without
explicit ethical standards and lack parameters to make value judgments.

Machine learning (ML) – as ML systems base decisions on algorithms, probabilities and data, the future actions
suggested by ML tools may lead to errors with impacts that are challenging to quantify (Babic et al (2021)). Due to
their evolutionary nature, ML systems can create discrepancies between the data they are initially trained on and the
data they subsequently process. The complexity of ML systems makes it challenging to identify the error or the logical
reason behind an unforeseen result when it occurs.

Generative artificial intelligence (GAI) – according to interviews with cyber security experts, GAI may enhance
organisations’ cyber defence capabilities worldwide but can also be used by attackers to develop more potent cyber
weapons that will be harder to defend against (Moody’s (2023)). Cyber security departments may lack sufficient data
to feed their GAI-based cyber defence systems as most organisations are reluctant to share technical details of cyber
attacks with other companies.

According to OECD (2023), risks or incidents resulting from the use of GAI grew exponentially between December
2022 and September 2024, reaching an unprecedented peak of 730 incidents in June 2024, as illustrated in Graph B.1.

Hyperlink BIS
Increase
Hyperlink BIS
in the number of registered risks or incidents resulting from the use of GAI Graph B.1

Hyperlink BIS Count

Hyperlink BIS
800
700
600
500
Number of incidents

400
300
200
100
0
mar-2021
dec- 2019

dec- 2020

dec- 2021

dec- 2022

dec- 2023
mar-2019
jun-2019

mar-2020
jun-2020

jun-2021

mar-2022
jun-2022

mar-2023
jun-2023

mar-2024
jun-2024
sep-2019

sep-2020

sep-2021

sep-2022

sep-2023

sep-2024

Source: adapted from OECD (2023).

Considering the above, as well as the increasingly widespread use of AI worldwide, some countries are evaluating the
need to regulate its usage. For example, in the European Union, the Artificial Intelligence Act came into force on 1
August 2024 and will be fully applicable from 1 August 2026. Appendix 4 of this report provides an overview of current
AI frameworks (European Parliament (2024a)).

Governance of AI adoption in central banks 11

3 Risk management for AI

3.1 Risk management strategy

When defining an AI risk management strategy, the most important consideration is that the
implementation of AI models and tools is supported by a comprehensive risk management model. This
strategy should: define the AI risk profile; identify, evaluate and select AI projects; leverage and
adapt governance management models that are already in place; and protect information through
the full life cycle.
1. Define the AI risk profile. Determine the central bank’s appetite for AI-related risks, taking into
account the intended strategic goals and outcomes for AI use. This can then guide risk
management decisions on AI development, management and use. This also helps to comply with
regulations and align usage with the central bank’s resilience objectives (CGRM (2023)). By its
nature the output from GAI is not precise so it may be necessary to accept certain risks in favour
of innovation with AI. Central banks should carefully consider whether the output of AI or the
means to manage the risks around AI are aligned with its risk appetite for innovation.
At the initial stages of AI adoption, it is often safer to use AI only in internal processes that are
not critical for central banking functions until the risks associated with such technology are
properly assessed and controlled.
2. Establish a process to identify, evaluate and select feasible AI projects. Establish a
multidisciplinary team responsible for selecting feasible AI projects that align with the strategic
and innovation objectives of the central bank. The team’s main objective is to evaluate the
suitability of adopting certain AI models and tools, and verify whether the proposed models and
tools align with the central bank’s risk appetite. This process should take place before other risk
management activities.
Such a process starts with the identification and analysis of AI use cases, initiatives and projects,
their classification by risk level depending on the sensitivity of the information used to train them
and the information accessed and generated by the model (Graph 1). The second step is to
determine the appropriate controls, striking a balance between risks and controls that is
determined by the risk profile of the institution. One factor to consider here is risk introduced by
the use of third-party information or solutions. Once it has been decided that a use case, initiative
or project is feasible to implement, the new controls should be designed, implemented and
tested, and the technological infrastructure adapted if needed. Finally, the solution should be
integrated into the risk mitigation mechanisms that are already in place and must be monitored
during its operation.
These approaches allow for more efficient allocation of resources and greater precision in terms
of managing vulnerabilities. Box C explains an approach that central banks could follow to start
addressing AI risks according to AI models’ intended use.

12 Governance of AI adoption in central banks

 AI project evaluation process diagram Graph 1

Source: authors’ elaboration.

Box C

AI risks based on use cases, initiatives and projects

Classifying artificial intelligence (AI) risks based on use cases, initiatives and projects allows more flexibility to
focus on the level of risk of each use case (Amazon AWS (2023)). Central banks may employ third-party AI services or
applications for non-sensitive daily use. Such arrangements warrant specific considerations for each use case, initiative
or project in particular.

Risks associated with the daily use of AI tools

Buying or using open AI models for daily use, often using non-sensitive information, includes the following scenarios:

• Consumer application. The organisation consumes a public third-party AI service, either at no cost or paid.
The organisation does not own or know the training data or the model. Usage is through application
programming interfaces (APIs) or direct interaction with the application, according to the terms of service
of the provider.

Enterprise application. The organisation uses a third-party enterprise application that has AI features
embedded within, and a business relationship is established between the organisation and the vendor.

The risks associated with the daily use of AI tools, such as consumer or enterprise applications, are generally
lower compared with the risks tied to newly developed AI projects. Daily use scenarios involve established third-party
services in which the organisation interacts with public or enterprise applications under defined terms of service. These
scenarios, while not free from risk, benefit from the inherent stability and predictability of using pre-existing AI models
and services. Organisations must remain vigilant about compliance and data privacy, but the risks are typically well
managed through robust vendor relationships and standardised security protocols.

3. Leverage and adapt governance management models already in place. Using and adapting
existing governance management models is important for both completeness and consistency.
For example, the three lines of defence model, which separates management and control
responsibilities into three distinct layers, should be adapted and used to clarify roles and
responsibilities in managing AI risks.
▪ First line roles: these are directly aligned with delivering products and/or services to internal
and external stakeholders, including support functions. They own and manage related risks
with AI outputs.

Governance of AI adoption in central banks 13

 ▪ Second line roles: these are responsible for the regulation and oversight of AI standards,
supporting the first line in identifying relevant risks and challenging their risk assessments
and control effectiveness.
▪ Third line roles: these provide independent and objective assurance and advice on the
adequacy and effectiveness of governance and risk management.
The use of the three lines of defence model has several advantages in managing AI risks. The
clear structure of roles and responsibilities facilitates the integration of AI risk management,
though it may be necessary to consider specific aspects of each line’s function for risks unique to
AI. This model enhances the ability to quickly identify and mitigate AI risks and strengthens
resilience against potential incidents. In addition to the three lines of defence model, central
banks have developed robust risk management frameworks and processes to identify, monitor
and mitigate many kinds of risks. Risks associated with AI should be included in these existing
mechanisms, provided they are updated to capture risks that are unique or more pronounced
with AI use (see Box A).
The table below contains some AI considerations for the three lines of defence (3LoD) model.

AI considerations for the 3LoD model Table 2

Function Considerations
– Identify AI use cases, initiatives and projects that are aligned with the central bank’s strategy or add
innovation value. Identify use cases that are valid and reliable with clear benefits.
– Identify and assess risks. Conduct risk assessments (identification, analysis and evaluation) that
consider AI-specific threats and vulnerabilities, and determine whether use cases match the risk
appetite.
– AI-specific technical controls. Implement new controls, based on strengths, weaknesses, opportunities
First line and threats (SWOT) analysis, such as AI model robustness testing, training data and outcome validation,
and bias detection techniques.
– Monitoring mechanisms. Set up continuous monitoring mechanisms to detect performance and
security issues. Monitor data drift in AI models where changes to the statistical distribution of the
underlying data can potentially cause a decline in model performance.
– Deliver specialised training. Ensure technical and risk management skills are regularly refreshed.
– Alignment of AI use with risk appetite and risk profile. Support the first line to determine if the use
case, initiative or project is aligned with the central bank’s risk appetite and profile.
– Risk methodology and prioritisation. Develop the institutional risk management methodology,
coordinate the execution of risk assessments made by the first line and prioritise risks based on the AI
risk profile.
– AI-specific policies and guidelines. Update existing policies or develop new policies and guidelines to
Second line address AI-specific risks and promote the adoption of AI while ensuring ethical and secure use. Update
data governance policies if necessary.
– Compliance and ethics. Supervise compliance with regulations and adherence to ethical principles.
Analyse legal framework according to the finance sector and central bank contexts.
– Specialised training and awareness. Design and provide AI-specific risk training materials and instil
awareness.
– Technical and ethical audits. Perform audits that evaluate the security controls as well as the fairness,
transparency and ethics of AI models.
Third line
– Continuous review. Provide ongoing reviews and recommendations for improving AI controls and
policies, adapting to fast-evolving technology and associated risks.
Source: authors’ elaboration.

14 Governance of AI adoption in central banks

 When applying a risk management framework to the adoption of AI models and tools, it is crucial
to define specific objectives based on principles. For example, a key principle is that central banks should
implement a robust ICT risk management programme in alignment with their operational risk
management framework.7 Existing frameworks, based on standards like ISO 310008, can be very useful
regarding promoting strong data governance and operational resiliency (BCBS (2021b)). Key elements
include:
▪ governance (discussed in more detail in Section 5);
▪ operational-digital (integrated) risk management;
▪ business continuity plan;
▪ mapping interconnections, interdependencies and dependency management (as part of
third-party risk management);
▪ incident management;
▪ ICT and information security risk management; and
▪ cyber security.
4. Protect information through the full life cycle. Due to the large volume of data that AI models
require to operate and the opaqueness associated with AI models, one of the main challenges
associated with AI risk management is protecting information, in terms of confidentiality,
integrity, availability and privacy. A resilience perspective ensures that organisations not only
prevent cyber attacks and protect ICT infrastructure, but also have the capacity to quickly recover
and maintain essential operations during and after cyber incidents. As a result, protecting
information throughout the AI life cycle is of the utmost importance.

3.2 Information security, privacy and cyber security risks

Information security and cyber security programs that have been developed based on standards such as
ISO/IEC 27001:2022 – Information security management9 or the National Institute of Standards and
Technology (NIST) Cybersecurity framework10 will continue to be adequate to manage the risks posed by
AI models. These frameworks are flexible and scalable, so they can be adapted to different industries and
technologies. Nevertheless, they need to be modified to capture the specific AI-related risks described in
Section 2. A key aspect of AI models is that systems or tools that support these models should be valid,
safe and reliable. In this way, cyber security is crucial for constructing reliable AI systems; by directly
contributing to the confidentiality of sensitive information, mitigation of integrity and availability risks, as
well as ensuring adequate access management.
Beyond preventing attacks and mitigating vulnerabilities, several standards focus on data
governance and ensure that the results from AI are explainable and interpretable:
• NIST Artificial intelligence risk management framework (NIST AI RMF);11

7
See principle 10 of BCBS (2021a).
8
ISO (2018).
9
ISO and IEC (2022a).
10
NIST (2024).
11
NIST (2023).

Governance of AI adoption in central banks 15

• ISO/IEC 23894:2023 – Information technology – artificial intelligence – guidance on risk
management;12 and
• ISO/IEC 38507:2022 – Information technology – governance of IT – governance implications of the
use of artificial intelligence by organisations.13
These publications promote an integrated AI risk management framework into general risk
models already in place. NIST AI RMF is a specialised technical approach to managing AI risks and
complements the widely used general risk management model NIST Risk management framework (NIST
RMF)14. These security frameworks might be considered part of the corporate mitigants, within the
information security corporate definitions and risk framework to ensure the adequate treatment of
information and data, as well as information technology (IT) assets, when adopting AI models. For example,
if an AI model needs access to a database, it is necessary to:
− know the classification and sensitivity level of the data stored in the database in terms of
confidentiality, integrity and availability;
− establish the control baseline for that particular classification and sensitivity;
− identify who or which systems are allowed to access those data and manage access; and
− monitor threats and report incidents related to information security and technological assets
involved in the AI model.
If central banks decide to use or rely on third-party AI services, tools, components or algorithms,
a third-party risk management model must be in place. This model begins by identifying external
participants, assessing risks, determining the central bank risk profile and level of acceptance of external
risks and their potential impact on critical processes. Central banks should perform due diligence to select
third parties and negotiate contracts that define the rights and responsibilities of all parties, considering
specific AI model risks.
When evaluating third-party data sets used by AI systems, it is important to have a clear
understanding of data quality, training data sources, data ownership and traceability. Regarding AI model
attributes, it is necessary to clarify the type of model, learning method, biases that may be present,
autonomy level and how much human oversight employed. In this way, a central bank should align the
identified third-party AI risks with the current corporate risk evaluation, before taking decisions on legal
agreements, to ensure that the risk appetite is adequately represented by the conditions and duties of
third-party service providers.

3.3 Specific actions to mitigate GAI risks

The complexity of GAI models requires special attention on trying to understand the outputs of those
models, including potential biases, limitations and robustness, as well as considering the greater human
supervision they will require. Some specific controls and practices examples to address GAI risks are listed
below.

12
ISO and IEC (2023a).
13
ISO and IEC (2022b).
14
NIST (2018).

16 Governance of AI adoption in central banks

 Examples of controls to address GAI risks Table 3

Risk category Approach Specific controls or practices

Define organisational functions related to GAI adoption

that include creating a risk management culture for all
Define key functions phases of the GAI systems life cycle, assessing and
monitoring GAI risks, and responding to and treating GAI
risks.
Establish a dedicated committee to identify and assess use
cases, complementing existing governance structures.
The key is to identify and carefully select cases in which AI
generates a productivity impact and those where it does
not.
Enhanced governance
structures It is necessary to maintain an inventory of use cases to
avoid project duplication and encourage innovative uses.
Establish formal approval processes for use of GAI tools.
Clear allocation of roles and responsibilities between
Strategic model owner/developer/user and model validators.
Establish specific GAI requirements in existing policies or if
necessary in an AI-dedicated policy that: define permitted
and prohibited practices, and encourage its usage while
balancing trustworthy AI principles or characteristics. 15
Examples of prohibited practices are: (i) publishing AI
generated documents without proper review due to
urgency; (ii) using GAI tools in critical processes before
Enhanced policies and these tools are well calibrated; (iii) using sensitive
guidelines information to train GAI models; and (iv) letting GAI tools
modify their algorithms by themselves.
If an AI-specific policy or guideline is created, it must be
aligned to existing policies and governance mechanisms
without duplicating work.
Protect data used to train GAI models according to existing
data policies.
Review applicable end-user licence agreement with
Legal certainty
support from the legal team.
Align GAI usage to regulatory requirements and include it
in the monitoring and reporting processes. Regulatory
requirements must be considered when sensitive
information assets or personally identifiable information
(PII) are used as inputs of AI models. Particular attention is
Compliance
Operational needed when AI solutions are deployed on public clouds,
or third parties without proper contracts in which clear
legal duties and responsibilities are established in case of
any negative impact on the attributes of sensitive
information.
Business continuity Review and, if necessary, update business continuity plans.
Involve people with experience in various areas to define
People
GAI usage guidelines to foster acceptable use of tools.

15
NIST (2023) defines the following key characteristics of trustworthy AI systems: valid and reliable, safe, secure and resilient,
explainable and interpretable, privacy enhanced, fair (with harmful bias managed), and accountable and transparent.

Governance of AI adoption in central banks 17

 Information security and privacy
Determine what information can be used in AI models
based on their classification.
Data classification
Classify AI models according to the most sensitive data
used for training.
Cyber security
Identify GAI solutions that staff are using without the
knowledge of the IT and cyber security units (shadow AI)
and propose how these use cases should be addressed in
the organisation. This identification could be done through
network traffic analysis, considering the main AI public
platforms.
Limiting use to AI tools which are installed locally to avoid
Limitation on the use of AI
sharing data with external parties.
tools
Information security, Blocking access to public online models except for non-risk
privacy and cyber activities.
security Use separate networks or sandboxes for AI tools.
Limit access to relevant data and/or limit the data that can
be used for creating inputs (prompts) to non-sensitive
data.
Filter outputs at the application level to remove sensitive
Data protection
or inappropriate information.
Perform security checks to open source or third-party
models according to cyber controls.
Identify and mitigate specific GAI and LLM systems
vulnerabilities OWASP (2024).
System security Consider specific GAI threats in application threat models,
ie prompt injection, insecure plugin design, supply chain
vulnerabilities and training data poisoning.
Make sure AI systems follow both AI and security
standards.
Analyse current and future resource needs, including
processing power, data storage and bandwidth.
Information and Ensure that the infrastructure can handle AI workloads.
communication Capacity planning Deploy or utilise elastic architectures, based on containers
technology (ICT) or virtual machines.
Continuously monitor and optimise resource usage for
maximum efficiency and performance.
Understand the provider’s terms of service and privacy
policy, and review them periodically as they can change
without notice.
Third party Terms and conditions Understand the service’s data flow, particularly if it stores
information, uses it to improve the model or shares it with
fourth parties.
Understand ownership of prompts and responses.
Understand the data used to train the model and the
Reputational Explainability quality of the information. Only use explicable AI
outcomes.

18 Governance of AI adoption in central banks

 Interpretation and reliability
Even though AI models were trained with large amounts of
data, their results must be considered as drafts that must
be reviewed and discussed. These reviews must be carried
out by humans and in many cases are new process
Human supervision activities that should be done by experts. Notably, output
can seem accurate even when it is not. This requires close
attention.
Establish procedures for output validation that consider
the review of a human expert on the subject.
Training people to ensure the risks and limitations of AI
models are known.
Employee education
Learn how to create adequate inputs (prompts) to increase
AI models
the probability of obtaining the expected results.
Transparency and accountability
Disclose AI usage. Appropriately denote, label or identify
Transparent usage
work and processes that use GAI tools.
Requirement for more validation and authorisation even
Enhanced validation
for lower risk use cases.
Environmental
Efficient energy use Optimise the life cycle of AI systems.
Ethical and social
Elaborate internal rules or manuals that set out the
Internal rules and manuals
responsible use of AI tools.

Source: authors’ elaboration.

4 Governance

The inherent complexity and uncertainty associated with AI systems makes their adoption challenging.
Effective governance16 mechanisms can help balance the risks and rewards of AI adoption in a consistent
risk-based manner. This is important because AI may affect business processes and decision-making across
the institution. AI governance is important not only for complying with national and international
strategies, laws or regulations (including multilateral agreements between countries), but also for ensuring
the alignment of any AI initiative with the organisation’s strategy. Effective governance mechanisms should
support efficiency and innovation in the organisation while effectively identifying and balancing associated
risks. That said, AI governance does not have to start from scratch but can be built on policies, procedures
and management tools that are already in place.

16
AI governance refers to the set of policies, rules, frameworks, principles and/or standards that guide organisations in their
adoption or development of secure, responsible and ethical use of AI.

Governance of AI adoption in central banks 19

4.1 Current industry frameworks
Several institutions worldwide17 have developed AI frameworks for organisations to exercise AI governance
(Appendix 4). While guidelines can be a good starting point for setting up governance practices, these
need to be commensurate with each organisation’s operational environment and risk appetite. In the case
of central banks, these include the generally low risk appetite and the importance of fulfilling its mandate
and maintaining transparency.
For AI to be trustworthy, the majority of industry guidance considers balancing each of the
principles listed below:
• secure – AI systems should be robust, secure, effective and resilient;
• data privacy compliant – AI users should ensure data privacy internal governance is maintained;
• explainable and transparent – AI users should ensure that decisions taken with the support of
AI are understandable and clear;
• reliable – AI systems should consistently perform as expected;
• ethical – AI users should ensure that the results of AI systems help promote social and
environmental well-being, ensure diversity, and are non-discriminatory and fair;
• responsible – AI users must ensure that AI systems assist the human decision-making process,
allowing people to oversee the system and override its decisions; and
• accountable – users should always know and follow the internal governance of AI usage.

Common AI principles of different frameworks Graph 2

Source: authors’ elaboration.

17
Eg the International Organization for Standardization (ISO), the Organisation for Economic Co-operation and Development
(OECD) and the National Institute of Standards and Technology (NIST) of the US Department of Commerce.

20 Governance of AI adoption in central banks

4.1.1 Adaptive governance frameworks
To deal with the fast pace of change in the field, AI governance needs to be dynamic and adaptative. A
governance framework is dynamic if it is able to keep up with developments in the field and is flexible and
responsive. A dynamic and adaptive framework can be based on the following principles:
1. Control – the organisation defines guard rails to comply with policies, requirements and
regulations.
2. Agility – delegating roles and teams with authority to make distributed and/or mandated
decisions.
3. Autonomy – people in the organisation can make autonomous governance decisions based, for
instance, on real-time data and in line with the enterprise’s objectives and goals.
Graph 3 sketches an adaptive AI governance framework, in which AI projects are evaluated on
their ethical, data security, transparency, accountability, non-discrimination, regulatory and compliance
implications and can only be implemented once these points have been analysed and addressed.

Adaptive AI governance framework Graph 3

Sources: Grasso (2021); Gartner (2022).

4.2 Proposed actions for AI governance at central banks

The CGRM has identified the following actions as useful for governance of the adoption of AI in central
banks:
1. Establish an interdisciplinary AI committee
Prior to establishing specific AI governance, central banks should establish a dedicated AI
committee, which will serve as an oversight body to help guide the implementation of
governance requirements. This oversight body should be sufficiently interdisciplinary, given the

Governance of AI adoption in central banks 21

 far-reaching nature of AI risk, and at a sufficiently senior level to ensure organisational buy-in.
While this could mean establishing a specific committee to oversee AI use, it could also be an
additional responsibility of an existing committee.
2. Define principles for responsible AI use
The committee overseeing AI use should establish a set of principles that outlines the
organisation’s philosophy and approach to AI use. These principles should be sufficiently clear to
capture the central bank’s overall posture towards AI use and consistently guide governance
decisions. The information from standard-setting bodies (such as ISO) or other organisations
(such as OECD) can help with defining these principles.
3. Establish an AI framework and update existing guidance
Set up a robust AI governance framework in line with the central bank’s strategy, objectives,
values and established AI principles. This framework should be used when assessing AI initiatives
proposed by different parts of the bank and when managing projects.
4. Maintain an AI tools inventory
An inventory ensures that all AI systems within an organisation are known and can be properly
managed, monitored and aligned with the central bank’s AI governance framework.
5. Map AI tools and stakeholders
Once the principles and inventory are defined, there is a need to establish an understanding of
which processes might benefit from AI usage and the corresponding stakeholders (current users,
potential users, data providers and third parties etc). This mapping would constitute the use cases
that are relevant for the central bank. Incorporating stakeholders’ participation, both internal and
external, is crucial to ensure that AI systems meet regulatory and ethical requirements.
6. Perform a detailed assessment of risks and controls
The assessment of risks and controls should not only involve ICT technical aspects but also other
relevant functions such as cyber security, legal, information security operational risk and external
third parties. It should identify controls that need to be in place before testing starts or before an
AI tool goes live.
7. Perform regular monitoring
The second line of defence should develop compliance monitoring and reporting to the AI
committee to ensure that systems, operational practices, information treatment and ethical
implications are in line with defined guidelines. Effective compliance monitoring for AI
governance includes:
a. The deployment of tools that continuously monitor AI systems for data drift, performance,
fairness, bias and adherence to defined parameters. They should also ensure that data used
for AI training and operations are handled in compliance with privacy laws and security
standards.
b. Implementing processes to ensure the quality and integrity of data used by AI systems.
c. Developing a policy to respond to incidents involving AI systems, including their
investigation, mitigation and reporting. Conducting root cause analysis for any compliance
breaches or performance issues to prevent future occurrences.
d. Considering human oversight for compliance monitoring as another component to validate
the accomplishment of the defined guidelines.

22 Governance of AI adoption in central banks

8. Report anomalies and incidents
Once AI tools have been implemented, there should be a process for reporting incidents involving
AI to the overseeing AI committee to capture residual risks that require further reduction or
controls that require further improvement.
9. Develop and improve workforce skills
The proper functioning of central banks largely depends on the specialisation and expertise of
their staff. AI applications reduce intensive manual processing, allowing staff to dedicate a larger
share of their time to more productive and innovative activities. To build AI knowledge, central
banks should define basic and specialist training and awareness programmes on AI usage,
governance and compliance.
10. Perform ongoing reviews and adaptations to the framework
Given the fast pace of change in AI, governance frameworks need to be reviewed regularly to
ensure validity and appropriateness on an ongoing basis. Central banks need to regularly review
the outcomes of AI systems to ensure they align with their organisational goals and risk
management measures. They should also consider the outcomes of risk and control assessments
performed, as well as incident reports involving any AI tools.
Together, these actions could form an AI governance system for central banks. These actions
should all be aligned to the central bank’s objectives and do not need to be created separately from
existing forms of technology governance.

General proposal for governance and risk management associated with the use
of AI models by central banks Graph 4

Source: authors’ elaboration.

Governance of AI adoption in central banks 23

5 Conclusions

Central banks are increasingly adopting AI to improve data quality, enhance operations and support
decision making. AI provides powerful tools for addressing complex challenges in areas such as data
analysis, risk assessment, forecasting, operations, and customer and corporate services.
But the use of AI tools can also pose new risks and have the potential to exacerbate existing risks
in central banks. Such risks are best managed holistically at the design, implementation and operations
phases. Information security risks deserve special mention, given the criticality and sensitivity of the
information handled by central banks. In general, central banks could start adopting AI in non-critical
processes where risks can be better controlled.
A safe AI adoption could cover the following domains: (i) governance; (ii) legal and compliance;
(iii) information security and privacy; (iv) cyber security; (v) third-party risk management; (vi) business
continuity; and (viii) other operational risks associated with the level of digitalisation and exposure of the
organisation.
Instead of reinventing the wheel, such a framework can build on existing risk management
frameworks and governance schemes. The task force did not identify any entirely new approaches to
managing AI risks, but recommends adapting existing mechanisms to the specific risks posed by AI tools.
Central banks can use existing risk management mechanisms such as the three lines of defence model
with specific adjustments.
We recommend that the AI risk profile is discussed and defined by upper management before
adopting AI models as this would help to allocate resources and set priorities. Benefits need to be weighed
against risks posed to information and core functions and be aligned with the risk appetite and security
capabilities of the organisation.
Given the transformative potential of AI technology, both in terms of its business impact and
possible externalities to society at large, developing a governance framework for AI adoption is of high
importance for central banks interested in the use of AI. This involves revising policies that cover various
procedures for the organisation’s governance and operations, such as systems and risk management,
compliance and data maintenance, and transparency and communication with internal and external
stakeholders.
Good practices laid out in international standards can serve as a starting point. These practices
include:
• systems and risk management – updating systems to integrate AI while ensuring robust risk
management practices;
• compliance and data maintenance – ensuring AI systems comply with existing regulations and
maintain high standards of data integrity and privacy; and
• transparency and communication – enhancing transparency in AI decision-making processes
and effectively communicating these processes to internal and external audiences.
By leveraging these international standards and best practices, central banks can effectively
incorporate AI into their governance frameworks – ensuring the secure, responsible and ethical use of AI
technology.

24 Governance of AI adoption in central banks

References

Accornero, M and G Boscariol (2022): “Machine learning for anomaly detection in datasets with categorical
variables and skewed distributions” in “Machine learning in central banking”, IFC Bulletin, no 57,
November, presentation given at an Irving Fisher Committee on Central Bank Statistics and Bank of Italy
workshop, 19–22 October 2021, www.bis.org/ifc/publ/ifcb57_05.pdf.
Amadxarif, Z, J Brookes, N Garbarino, R Patel and E Walczak (2021): “The language of rules: textual
complexity in banking reforms” Bank of England Staff Working Paper, no 834, November,
www.bankofengland.co.uk/-/media/boe/files/working-paper/2019/the-language-of-rules-textual-
complexity-in-banking-reforms.pdf.
Amazon AWS (2023): “Securing generative AI: an introduction to the generative AI security scoping matrix”,
AWS Security Blog, 19 October, aws.amazon.com/es/blogs/security/securing-generative-ai-an-
introduction-to-the-generative-ai-security-scoping-matrix/.
——— (2024): “What’s the difference between deep learning and neural networks?”,
aws.amazon.com/compare/the-difference-between-deep-learning-and-neural-networks/?nc2=h_mo-
lang.
Araujo, D, G Bruno, J Marcucci, R Schmidt and B Tissot (2022): “Machine learning applications in central
banking: an overview” in “Machine learning in central banking”, IFC Bulletin, no 57, November,
www.bis.org/ifc/publ/ifcb57_01_rh.pdf.
Babic, B, I Cohen, T Evgeniou and S Gerke (2021): “When machine learning goes off the rails”, Harvard
Business Review, January–February 2021, hbr.org/2021/01/when-machine-learning-goes-off-the-rails.
Baker-Brunnbauer, J (2023): Trustworthy artificial intelligence implementation: introduction to the TAII
framework, Springer, link.springer.com/book/10.1007/978-3-031-18275-4.
Bank for International Settlements (BIS) (2024): “Artificial intelligence and the economy: implications for
central banks”, Annual Economic Report 2024, June, Chapter III, www.bis.org/publ/arpdf/ar2024e3.pdf.
Bank for International Settlements Innovation Hub (BISIH) (2023): Project Aurora: the power of data,
technology and collaboration to combat money laundering across institutions and borders, May,
www.bis.org/publ/othp66.htm.
——— (2024): Project Raven: using AI to assess financial system’s cyber security and resilience, April,
www.bis.org/about/bisih/topics/cyber_security/raven.htm.
Basel Committee on Banking Supervision (BCBS) (2021a): Revisions to the principles for the sound
management of operational risk, March, www.bis.org/bcbs/publ/d515.htm.
——— (2021b): Principles for operational resilience, March, www.bis.org/bcbs/publ/d516.htm.
Benford, J (2024): “Trusted AI: ethical, safe and effective application of artificial intelligence at the Bank of
England”, speech given at the Central Bank AI Conference, Bank of England, September,
www.bankofengland.co.uk/speech/2024/september/james-benford-speech-at-the-central-bank-ai-
inaugural-conference.
Bluwstein, K, M Buckmann, A Joseph, M Kang, S Kapadia and Ö Simsek (2020): “Credit growth, the yield
curve and financial crisis prediction: evidence from a machine learning approach”, Bank of England Staff
Working Paper, no 848, January, www.bankofengland.co.uk/-/media/boe/files/working-
paper/2020/credit-growth-the-yield-curve-and-financial-crisis-prediction-evidence-from-a-machine-
learning.pdf.

Governance of AI adoption in central banks 25

Boyd, D and K Crawford (2012): “Critical questions for big data”, Information, Communication & Society,
vol 15, no 5, pp 662–79, doi.org/10.1080/1369118X.2012.678878.
Buckmann, M, G Potjagailo and P Schnattinger (2023): “Dissecting UK service inflation via a neural network
Phillips curve,” Bank Underground, 10 July, bankunderground.co.uk/2023/07/10/dissecting-uk-service-
inflation-via-a-neural-network-phillips-curve/.
Burgess, S, E Fernandez-Corugedo, C Groth, R Harrison, F Monti, K Theodoridis and M Waldron (2013):
“The Bank of England’s forecasting platform: COMPASS, MAPS, EASE and the suite of models”, Bank of
England Staff Working Paper, no 471, May, www.bankofengland.co.uk/-/media/boe/files/working-
paper/2013/the-boes-forecasting-platform-compass-maps-ease-and-the-suite-of-models.pdf.
Cagala, T, J Hees, D Herurkar, M Meier, N-T Nguyen, T Sattarov, K Troutman and P Weber (2022):
“Unsupervised outlier detection in official statistics” in “Machine learning in central banking”, IFC Bulletin,
no 57, November, presentation given at an Irving Fisher Committee on Central Bank Statistics and Bank of
Italy workshop, 19–22 October 2021, www.bis.org/ifc/publ/ifcb57_09.pdf.
Chakraborty, C and A Joseph (2017): “Machine learning at central banks”, Bank of England Staff Working
Paper, no 674, www.bankofengland.co.uk/-/media/boe/files/working-paper/2017/machine-learning-at-
central-banks.pdf.
Chen, M, M DeHaven, I Kitschelt, S Lee and M Sicilian (2023): “Identifying financial crises using machine
learning on textual data”, Board of Governors of the Federal Reserve System, International Finance
Discussion Papers, no 1374, doi.org/10.17016/IFDP.2023.1374.
Chiarello, F, V Giordano, I Spada, S Barandoni and G Fantoni (2024): “Future applications of generative
large language models: a data-driven case study on ChatGPT”, Technovation, vol 133,
doi.org/10.1016/j.technovation.2024.103002.
Consultancy.eu (2023): “European AI Act: Implications for the financial services industry”, 19 October,
www.consultancy.eu/news/9392/european-ai-act-implications-for-the-financial-services-industry.
Consultative Group on Risk Management (CGRM) (2023): Central bank digital currency (CBDC) information
security and operational risks to central banks, November, www.bis.org/publ/othp81.pdf.
Dauphin, J-F, K Dybczak, M Maneely, M Sanjani, N Suphaphiphat, Y Wang and H Zhang (2022):
“Nowcasting GDP: a scalable approach using DFM, machine learning and novel data, applied to European
economies”, IMF Working Paper, no 52, March, www.imf.org/en/Publications/WP/Issues/2022/03/11/No
wcasting-GDP-A-Scalable-Approach-Using-DFM-Machine-Learning-and-Novel-Data-Applied-to-513703.
Denes, J, A Lestrade and L Richardet (2022): “Using Twitter data to measure inflation perception” in
“Machine learning in central banking”, IFC Bulletin, no 57, November, presentation given at an Irving Fisher
Committee on Central Bank Statistics and Bank of Italy workshop, 19–22 October 2021,
www.bis.org/ifc/publ/ifcb57_13.pdf.
Derner, E and K Batistič (2023): “Beyond the safeguards: exploring the security risks of ChatGPT”, arXiv
preprint arXiv:2305.08005, arxiv.org/pdf/2305.08005
Devys, E and U von Kalckreuth (2022): “The use of AI for company data gathering – finding and monitoring
fintechs in Germany and France” in “Machine learning in central banking”, IFC Bulletin, no 57, November,
presentation given at an Irving Fisher Committee on Central Bank Statistics and Bank of Italy workshop,
19–22 October 2021, www.bis.org/ifc/publ/ifcb57_28.pdf.
European Central Bank (ECB) (2023): “Suptech: thriving in the digital age”, Supervision Newsletter, 15
November, www.bankingsupervision.europa.eu/press/supervisory-
newsletters/newsletter/2023/html/ssm.nl231115_2.en.html.

26 Governance of AI adoption in central banks

European Parliament (2024a): “EU AI Act: first regulation on artificial intelligence”, 18
June, www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-
artificial-intelligence.
European Parliament (2024b): “Regulation (EU) 2024/1689 of the European Parliament and of the Council
of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No
300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and
Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act)”, Official Journal of
the European Union, 7 July, eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689.
Faria da Costa, A, F Fonseca and S Maurício (2022): “Novel methodologies for data quality management
anomaly detection in the Portuguese central credit register” in “Machine learning in central banking”,
IFC Bulletin, no 57, November, presentation given at an Irving Fisher Committee on Central Bank Statistics
and Bank of Italy workshop, 19–22 October 2021, www.bis.org/ifc/publ/ifcb57_29.pdf.
Future of Life Institute (2024): “The AI Act Explorer”, EU Artificial Intelligence Act,
artificialintelligenceact.eu/ai-act-explorer/.
Gartner (2022): “Choose adaptive data governance over one-size-fits-all for greater flexibility”, 11 April,
www.gartner.com/en/articles/choose-adaptive-data-governance-over-one-size-fits-all-for-greater-
flexibility.
Gascon, C and D Werner (2022): “Does the Beige Book reflect US employment and inflation trends?”,
Economic Synopses, Federal Reserve Bank of St. Louis, doi.org/10.20955/es.2022.13.
Grasso, C (2021): “Governance: from research trend to enterprise standard”, Dataiku blog, 2 August,
blog.dataiku.com/governance-from-research-trend-to-enterprise-standard.
Haghighi, M, C Jones and J Younker (2022): “Machine learning for anomaly detection in financial regulatory
data” in “Machine learning in central banking”, IFC Bulletin, no 57, November, presentation given at an
Irving Fisher Committee on Central Bank Statistics and Bank of Italy workshop, 19–22 October 2021,
www.bis.org/ifc/publ/ifcb57_24.pdf.
Hirsch, T, K Merced, S Narayanan, Z E Imel and D C Atkins (2017): “Designing contestability: Interaction
design, machine learning, and mental health”, Proceedings of the 2017 Conference on Designing
Interactive Systems, pp 95–99.
Info-communications Media Development Authority (IMDA) and Personal Data Protection Commission
(PDPC) (2020): Model artificial intelligence governance framework, second edition, www.pdpc.gov.sg/-
/media/files/pdpc/pdf-files/resource-for-organisation/ai/sgmodelaigovframework2.pdf.
Institute of Electrical and Electronics Engineers (IEEE) (2023) “General Principles of Ethically Aligned
Design”, The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems, standards.ieee.org/wp-
content/uploads/import/documents/other/ead1e_general_principles.pdf.
International Organization for Standardization (ISO) (2018): “Risk management – guidelines”, ISO
31000:2018, www.iso.org/standard/65694.html.
International Organization for Standardization (ISO) and International Electrotechnical Commission (IEC)
(2017): “Information technology – cloud computing – interoperability and portability”, ISO/IEC 19941:2017,
www.iso.org/standard/66639.html.
——— (2022a): “Information security, cybersecurity and privacy protection – information security
management systems – requirements”, ISO/IEC 27001:2022, www.iso.org/standard/27001.
——— (2022b): “Information technology – governance of IT – governance implications of the use of
artificial intelligence by organizations”, ISO/IEC 38507:2022, www.iso.org/standard/56641.html.

Governance of AI adoption in central banks 27

——— (2022c): “Trustworthiness – vocabulary”, ISO/IEC TS 5723:2022, www.iso.org/standard/81608.html.
——— (2023a): “Information technology – artificial intelligence – guidance on risk management”, ISO/IEC
23894:2023, www.iso.org/standard/77304.html.
——— (2023b): “Information technology – artificial intelligence – management system”, ISO/IEC
42001:2023, www.iso.org/standard/81230.html.
Jiménez, P and T Serrano (2022): “An artificial intelligence application for accounting data cleansing” in
“Machine learning in central banking”, IFC Bulletin, no 57, November, presentation given at an Irving Fisher
Committee on Central Bank Statistics and Bank of Italy workshop, 19–22 October 2021,
www.bis.org/ifc/publ/ifcb57_23.pdf.
Joseph, A, G Potjagailo, E Kalamara, C Chakraborty and G Kapetanios (2022): “Forecasting UK inflation
bottom up”, Bank of England Staff Working Paper, no 915, March, www.bankofengland.co.uk/-
/media/boe/files/working-paper/2021/forecasting-uk-inflation-bottom-up.pdf.
Kalamara, E, A Turrell, C Redl, G Kapetanios and S Kapadia (2020): “Making text count: economic
forecasting using newspaper text”, Bank of England Staff Working Paper, no 865, May,
www.bankofengland.co.uk/working-paper/2020/making-text-count-economic-forecasting-using-
newspaper-text.
Kerdsri, J and P Treeratpituk (2022): “Using deep learning technique to automate banknote defect
classification” in “Machine learning in central banking”, IFC Bulletin, no 57, November, presentation given
at an Irving Fisher Committee on Central Bank Statistics and Bank of Italy workshop, 19–22 October 2021,
www.bis.org/ifc/publ/ifcb57_34.pdf.
Khurana, D, A Koli, K Khatter and S Singh (2023): “Natural language processing: state of the art, current
trends and challenges”, Multimedia Tools and Applications, vol 82, pp 3713–44, doi.org/10.1007/s11042-
022-13428-4.
McCaul, E (2024): “From data to decisions: AI and supervision,” February,
www.bankingsupervision.europa.eu/press/interviews/date/2024/html/ssm.in240226~c6f7fc9251.en.html.
Moody’s Investor Service (Moody’s) (2023): “Generative AI may increase cyber risk rather than lessen it”,
Cybersecurity – Cross region, 18 September.
Mulligan, D, C Koopman and N Doty (2016): “Privacy is an essentially contested concept: a multi-
dimensional analytic for mapping privacy”, Philosophical Transactions of the Royal Society A, vol 374,
no 2083, December, dx.doi.org/10.1098/rsta.2016.0118.
National Institute of Standards and Technology (NIST) (2018): Risk management framework for information
systems and organizations, nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-37r2.pdf.
National Institute of Standards and Technology (NIST) (2023): Artificial intelligence risk management
framework (AI RMF 1.0), nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf.
National Institute of Standards and Technology (NIST) (2024): The NIST Cybersecurity Framework (CSF)
2.0, nvlpubs.nist.gov/nistpubs/CSWP/NIST.CSWP.29.pdf.
Nistor, M (2023): “Artificial intelligence applications in finance: considering the benefits and risks”, The
Teller Window, Federal Reserve Bank of New York, December,
tellerwindow.newyorkfed.org/2023/12/14/ai-applications-in-finance-considering-the-benefits-and-risks/.
Njoroge, L (2024): “Role of artificial intelligence (AI) in central banking: implications for COMESA member
central banks”, COMESA Monetary Institute Special Report, March, www.comesa.int/wp-
content/uploads/2020/09/Special-Report-AI-and-Big-Data-Implication-for-Central-Banking-in-COMESA-
region.pdf.

28 Governance of AI adoption in central banks

Oh, P (2024): “Navigating the AI Maze: How ISO Standards Guide Responsible Development and
Deployment”, Medium, 21 January, medium.com/@patrick-oh-sglion65/navigating-the-ai-maze-
09ff5909ae78.
Organisation for Economic Co-operation and Development (OECD) (2022): “OECD framework for the
classification of AI systems”, OECD Digital Economy Papers, February, doi.org/10.1787/cb6d9eca-en.
——— (2023): G7 Hiroshima process on generative artificial intelligence (AI): towards a G7 common
understanding on generative AI, doi.org/10.1787/bf3c0c60-en.
——— (2024a): “Recommendation of the Council on Artificial Intelligence”, OECD Legal Instruments,
legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449.
——— (2024b): “What is AI? Can you make a clear distinction between AI and non-AI systems?”, The AI
Wonk, March, oecd.ai/en/wonk/definition.
Open Web Application Security Project (OWASP) (2024): “OWASP Top 10 for Large Language Model
Applications”, The OWASP foundation, owasp.org/www-project-top-10-for-large-language-model-
applications/
Petropoulos, A, V Siakoulis, K Panousis, L Papadoulas and S Chatzis (2019): “A deep learning approach for
dynamic balance sheet stress testing”, in D Magazzeni, S Kumar et al (eds), ICAIF ‘22: proceedings of the
third ACM international conference on AI in finance, Association for Computing Machinery, pp 53–61,
doi.org/10.1145/3533271.3561656.
Pinsent and Masons (2024): “A guide to high-risk AI systems under the EU AI Act”, Out-law guide, 13
February, www.pinsentmasons.com/out-law/guides/guide-to-high-risk-ai-systems-under-the-eu-ai-act.
Richardson, A, T van Florenstein Mulder and T Vehbi (2021): “Nowcasting GDP using machine learning
algorithms: a real-time assessment”, International Journal of Forecasting, vol 37, no 2, pp 941–48,
doi.org/10.1016/j.ijforecast.2020.10.005.
Rubio, J, P Barucca, G Gage, J Arroyo and R Morales-Resendiz (2020): “Classifying payment patterns with
artificial neural networks: an autoencoder approach”, Latin American Journal of Central Banking, vol 1, nos
1–4, doi.org/10.1016/j.latcb.2020.100013.
Sharma, N, R Sharma and N Jindal (2021): “Machine learning and deep learning applications – a vision”,
Global Transitions Proceedings, vol 2, no 1, June, pp 24–28, doi.org/10.1016/j.gltp.2021.01.004.
Université de Montréal (2018): “Montréal declaration for a responsible development of artificial
intelligence”, Montréal Declaration Responsible AI, declarationmontreal-iaresponsable.com/wp-
content/uploads/2023/04/UdeM_Decl-IA-Resp_LA-Declaration-ENG_WEB_09-07-19.pdf.
Yenduri, G, M Ramalingam, G Chemmalar Selvi et al (2024): “GPT (generative pre-trained transformer) – a
comprehensive review on enabling technologies, potential applications, emerging challenges and future
directions”, IEEE Access, April, ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10500411.

Governance of AI adoption in central banks 29

Appendix 1: Artificial intelligence definitions

Artificial intelligence risk management framework (AI RMF 1.0) (NIST (2023))
“An AI system is an engineered or machine-based system that can, for a given set of objectives, generate
outputs such as predictions, recommendations, or decisions influencing real or virtual environments. AI
systems are designed to operate with varying levels of autonomy.”
Recommendation of the Council on Artificial Intelligence (OECD (2024a))
“An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it
receives, how to generate outputs such as predictions, content, recommendations, or decisions that can
influence physical or virtual environments. Different AI systems vary in their levels of autonomy and
adaptiveness after deployment.”
Trustworthy artificial intelligence implementation: introduction to the TAII framework (Baker-
Brunnbauer (2023))
“AI systems are software (and possibly also hardware) systems designed by humans that, given a complex
goal, act in the physical or digital dimension by perceiving their environment through data acquisition,
interpreting the collected structured or unstructured data, reasoning on the knowledge, or processing the
information, derived from this data and deciding the best actions to take to achieve the given goal.”
Model artificial intelligence governance framework (IMDA and PDPC (2020))
“AI refers to a set of technologies that seek to simulate human traits such as knowledge, reasoning,
problem solving, perception, learning and planning, and, depending on the AI model, produce an output
or decision (such as a prediction, recommendation, and/or classification).”
Role of artificial intelligence (AI) in central banking: implications for COMESA member central
banks (Njoroge (2024))
“Artificial intelligence (AI) uses computing to create intelligence artificially and is described as the ability
of machines to imitate human intelligence. It entails a collection of tools that learn with given data and
understand patterns and interactions between series and values.”

30 Governance of AI adoption in central banks

Appendix 2: Summary of AI use cases described in Section 1.1

The use cases described in Section 1.1 were identified from both public sources and from responses to a
questionnaire from members of the Consultative Group on Risk Management (CGRM) task force. Graph
2.1 below displays the distribution of the number of use cases.

Hyperlink BIS
AI use cases among CGRM task force members1
Hyperlink BIS
Graph 2.1

Hyperlink BIS Number of responses

Hyperlink BIS n ana s s as n an ana s s a n s s s

a n s n an s
ankn n an s b n n a n
sk ass ss n s an a s s

a ana s s ana s s
s na s
s a ass s an

n as n
as n n n

n a n as n
n an ns
s a an an
n s
aa a
n an ns

a
na s s
b n a n a b s a
s s n s s n an ns a nn ss s n

1
This graph includes more use case categories than those described in Section 1.1.

Source: authors’ elaborations.

Governance of AI adoption in central banks 31

Appendix 3: Risks associated with AI

Risks associated with AI Table 3.1

Applies to
Applies
Type of risk Description new
to AI
technologies

Lack of strategy and governance in AI, which may impede

Strategic
central banks’ ability to achieve their objectives.  
Loss resulting from inadequate or failed internal processes,
people, systems or from external events. Includes legal
uncertainties, compliance risks due to the black box nature
Operational of AI models, deficiencies in processes, risks associated with
personnel’s digital capabilities, third-party dependency and  
capacity issues related to the data intensive nature of AI
models.
Safeguard sensitive assets and personally identifiable
information in systems, particularly in public cloud
infrastructure, while ensuring regulatory compliance.
Information
The misuse or mishandling of sensitive data or personal
security,
information by AI models, exposure of confidential
privacy and
information due to inadequate cyber security controls, and  
cyber security
vulnerabilities such as model extraction and data poisoning,
all of which that can result in significant legal and
reputational consequences.
Potential failures or misconfigurations in ICT infrastructure
that could impact the performance, availability or integration
Information
of systems, compromising sensitive data.
and
This includes risks arising from software bugs, hardware
communication
failures, inadequate system design and challenges in  
technology
compatibility with new technologies, as well as insufficient
(ICT)
safeguards for critical processes during disruptions, which
can affect business continuity and operational resilience.
Incidents originating from dependence on external AI
models or tools developed by a third party outside the
Third party  
organisation including privacy breaches, operational
disruptions, compliance failures and cyber security threats.

Reputational damage or adverse publicity because of errors,

data leaks or lack of transparency, particularly due to
Reputational  
inadequate control of complex AI models. Such incidents can
lead to public scrutiny and diminished trust in central banks.
Risks associated with AI models include issues of
interpretation and reliability from imprecise results and black
box models, as well as challenges in transparency and
AI models accountability due to algorithm self-modification.
Environmental concerns arise from the high carbon footprint 
of energy intensive AI models, while ethical issues stem from
potential biases and inappropriate outcomes.
Source: authors’ elaboration.

32 Governance of AI adoption in central banks

 Appendix 4: Review of current AI frameworks

The European Union’s Artificial Intelligence Act18

The Artificial Intelligence Act (AI Act) 202419 mandates that providers follow six ethical principles. This
approach promotes innovation while ensuring responsible and safe AI development.

Values-based principles Table 4.1

AI system creators should design these systems to promote sustainable

1. Social and environmental well- and inclusive growth, social progress and environmental well-being.
being Providers must consider the potential societal and environmental impacts
of AI systems to ensure they contribute positively to these areas.

Developers and providers of AI systems should create these systems to

2. Diversity, non-discrimination and avoid discrimination and bias, while promoting diversity. Providers must
fairness rigorously examine data sources for bias and implement appropriate
measures to mitigate any potential biases.

AI systems must be transparent. Providers should offer clear information

3. Transparency regarding the system’s capabilities, limitations and the data sources used
for training.

Providers and developers should design AI systems to be reliable,

4. Technical robustness and safety predictable and safe for use. AI providers must ensure their systems
comply with established quality management standards.

AI systems should assist humans in decision-making, allowing humans the

5. Human agency and oversight ability to override system-generated decisions.

AI system providers and developers should design AI systems with a focus

6. Privacy and data governance on data privacy and protection. The data sets used for training should be
subject to stringent governance.

Source: adapted from European Parliament (2024b).

The AI Act classifies models according to its risks, categorising them into four different levels
(Graph 4.1):20
The unacceptable risk category establishes that systems considered a clear threat to the safety,
livelihoods and rights of people will be banned. This includes social scoring by governments, devices using
voice assistance that encourage dangerous behaviour and manipulative AI.
The second category, high risk, refers to any AI system which: (i) constitutes a certain type of
product (eg medical devices, industrial machinery, toys, aircraft or cars); (ii) is a safety component of a
certain type of product (eg components for rail infrastructure, lifts or appliances burning gaseous fuels);
or (iii) AI systems that are high risk by their nature eg biometrics, critical infrastructure, exploiting
vulnerabilities, education, employment, access to essential services both public and private, law
enforcement, immigration, the administration of justice and democratic processes.

18
Future of Life Institute (2024) and European Parliament (2024a).
19
European Parliament (2024b).
20
Pinsent and Masons (2024) and Future of Life Institute (2024).

Governance of AI adoption in central banks 33

 High-risk AI providers have specific obligations including establishing a risk management system
to operate throughout the AI system’s life cycle. Such providers must also conduct data governance,
ensure adequate training is carried out, and validate and test that data sets are free of errors. Another
obligation is to draw up technical documentation to demonstrate and assess compliance, and to design
high-risk AI systems to allow deployers to implement human oversight and achieve appropriate levels of
accuracy, robustness and cyber security. Finally, they must consider establishing a quality management
system to ensure compliance.
For AI systems considered as limited risk, the AI Act defines transparency obligations for
providers and deployers of AI systems. This includes ensuring that such systems are designed to inform
people that they are interacting with an AI system. Another obligation is to ensure that the outputs of AI
systems generating synthetic content (eg audio, image, video or text), are marked and detectable as
artificially generated. Finally, there is an obligation to ensure that deployers of an AI system that generates
or manipulates image, audio or video content constituting a deep fake shall disclose that the content has
been artificially generated or manipulated. In contrast, AI systems catalogued as low and minimal risk do
not have any specific obligations.
Moreover, the AI Act, specifies other obligations regarding the AI models’ technical
documentation, including training and testing processes and the results of evaluations. There is also an
obligation to make information and documentation available to providers of AI systems who intend to
integrate the general purpose AI model in their AI systems. Additional obligations include having a policy
in place regarding copyright law and making a detailed summary about the content used for training the
general purpose AI model publicly available.

European Union AI Act risk model Graph 4.1

Source: adapted from Consultancy.eu (2023).

The Institute of Electrical and Electronics Engineers (IEEE) Global Initiative on Ethics of
Autonomous and Intelligent Systems 21
This framework consists of eight general principles applicable to the creation and operation of all types of
autonomous and intelligent systems (AIS), regardless of whether they are physical robots or software
systems in real, virtual or mixed-reality environments.

21
IEEE (2023).

34 Governance of AI adoption in central banks

Values-based principles Table 4.2

Systems should be created and operated to respect, promote and protect

1. Human rights internationally recognised human rights.

Systems should adopt increased human well-being as a primary success

2. Well-being criterion for development.

Creators should empower individuals with the ability to access and

3. Data agency securely share their data, to maintain people’s control over their identity.

Developers and operators should provide evidence of the effectiveness

4. Effectiveness and suitability for purpose of the AI system.

5. Transparency The basis of a particular AI system decision should always be discoverable.

Systems should be developed and operated to provide an unambiguous

6. Accountability rationale for all decisions made.

Systems should guard against all potential misuses and risks during
7. Awareness of misuse operation.

Systems should specify competences and operators should have the

8. Competence knowledge and skills required for secure and effective operation.

Source: adapted from IEEE (2023).

The Montreal Declaration for Responsible Development of Artificial Intelligence

The Montreal Declaration for Responsible Development of Artificial Intelligence is based on 10 principles.
It covers areas similar to the two frameworks above and additionally emphasises democratic
participation, respect for autonomy and prudence during development.

Values-based principles Table 4.3

AI systems must permit the growth of well-being for individuals

1. Well-being such as living conditions, health, and the exercise of their mental
and physical capacities.

AI systems must be developed and used while respecting

2. Respect for autonomy people’s autonomy, and with the goal of increasing people’s
control over their lives and their surroundings.

Privacy and intimacy must be protected from artificial

3. Protection of privacy and intimacy intelligence system (AIS) intrusion, and data acquisition and
archiving systems.

4. Solidarity The development of AIS must be compatible with maintaining

the bonds of solidarity among people and generations.

AIS must meet intelligibility, justifiability and accessibility

5. Democratic participation criteria, and must be subject to democratic scrutiny, debate and
control.

The development and use of AIS must contribute to the creation

6. Equity of a just and equitable society.

Governance of AI adoption in central banks 35

 The development and use of AIS must be compatible with
7. Diversity inclusion maintaining social and cultural diversity and must not restrict
the scope of lifestyle choices or personal experiences.

Every person involved in AI development must exercise caution

8. Caution by anticipating, as far as possible, the adverse consequences of
AIS use and by taking appropriate measures to avoid them.

The development and use of AIS must not contribute to

9. Responsibility lessening the responsibility of human beings when decisions
must be made.

The development and use of AIS must be carried out to ensure

10. Sustainable development strong environmental sustainability of the planet.

Source: Université de Montréal (2018).

These principles have resulted in eight recommendations that provide guidelines for achieving
digital transition within the ethical framework of the declaration. These include implementing audits and
certifications, independent controlling organisations, ethics education for developing stakeholders,
empowerment of the user and ecological sustainability.

NIST AI risk management framework (AI RMF)

The US Department of Commerce’s National Institute of Standards and Technology (NIST) has published
this framework to help organisations to manage the risks of AI. The framework proposes two lines of
action.
The first line of action is related to establishing specific processes to identify, assess and address
risks associated with the development and use of a specific AI, aiming to enhance its safe use and reduce
the likelihood of incidents related to its use.
The management process referred to in the first line of action is composed of four functions:
govern, map, measure and manage (Graph 4.2).

Process functions to manage risks associated with AI Graph 4.2

Source: adapted from NIST (2023).

36 Governance of AI adoption in central banks

 The govern function is integral to the other three, providing the necessary framework to
implement, follow up and control the entire risk management process associated with AI, while aligning it
with the organisation’s principles and policies. The map function provides context to identify risks
associated with AI in terms of various factors related to its use, for instance stakeholders, values and
interests. This function is the basis for the other two functions. The measure function addresses, analyses
and monitors the AI risks and their impacts previously identified in the map function, and reports to the
govern function by using different qualitative and quantitative tools and methodologies. Finally, the
manage function’s aim is to prioritise risks and to take action based on their estimated impact. This
function includes developing plans to respond, recover and communicate about incidents or events in the
AI development life cycle and use. Regarding the second line of action, NIST refers to implementing
general processes to validate that AI systems are equipped with features that ensure their safe use and
reliability. It also suggests that organisations should ensure that AI is used in accordance with principles
that facilitate its governance, such as the ones described in the following table.

Characteristics of trustworthy AI Table 4.4

The AI subject to evaluation must not pose risks to human life, health, property or
1. Safe the environment.22

The AI under evaluation and the ecosystems in which it operates must be resilient
and able to tolerate adverse and unexpected events in its environment, maintain its
2. Secure and resilient functions and structure despite internal and external challenges and, if necessary,
degrade safely.

AI must provide comprehensive information about its functionality and reliability,

including its results, allowing humans to understand its behaviour, an attribute that
3. Explainable and interpretable is also referred to as contestability in specialised literature.23 Similarly, AI must be
implemented with transparency and accountability mechanisms regarding its
operation.

AI must safeguard human autonomy, identity and dignity, as well as protect

information from intrusion. It should limit its observation to strictly necessary
4. Privacy enhanced parties, and consider the rights of those who operate, monitor or use it to consent
to the disclosure or control of their identities.

AI should minimise bias, and thus the damage caused to individuals, groups,
communities, organisations and society as a whole. It is therefore important to
5. Fair – with harmful bias consider multiple human and social values, such as transparency and fairness in this
management process. Methodologies have been developed to examine the role of human values
in technological contexts, in line with the theoretical currents known as “values in
design” and “value-sensitive design”.

The AI under evaluation must provide objective evidence of compliance with

6. Valid and reliable specific use or application requirements. Additionally, it must function as intended,
without failures, during a given time interval and under given conditions.

Accountability presupposes transparency and both are necessary for trustworthy

AI. Transparency provides and promotes access to information, increasing
confidence in the AI system. Having organisational best practices and governance
7. Accountable and transparent
to address risks can contribute to more accountable AI systems.

Source: NIST (2023).

22
ISO and IEC (2022c).
23
Hirsch et al (2017)

Governance of AI adoption in central banks 37

OECD framework for the classification of AI systems24
This framework is a reference for “policy makers, regulators, legislators and others to characterise AI systems
for specific projects and contexts. The framework links AI system characteristics with the OECD AI Principles
…, the first set of AI standards that governments pledged to incorporate into policy making and promote the
innovative and trustworthy use of AI.”
This framework classifies AI systems and applications along five dimensions.

Key high-level dimensions of the OECD framework Graph 4.3

Source: authors’ elaboration.

The dimension people and planet considers different criteria such as users of the system,
impacted stakeholders, human rights and democratic values, well-being, society and the environment
including sustainability and stewardship.
Economic context refers to industrial sectors. It describes the business function and business
model, the level of criticality regarding the extent to which a disruption impacts systems functions or
essential services, and is also related to scalability and maturity (breadth of deployment).
The third criterion, data and input, refers to the detection and collection of data either by
humans, automated sensors or both, the provenance of the data and input, as well as their nature
(dynamic, static, updated or real time). Further, it covers proprietary, public or personal rights and the
identifiability of personal data. ISO/IEC 19441 (2017)25 distinguishes various categories – or “states” – of
data identifiability.
The AI model dimension refers to model characteristics like AI model type, model building either
from machine or human knowledge, model evolution, machine learning, central or federated machine
learning and model inference about transparency and explainability.
Finally, the task and outputs criterion describes the tasks that the system performs, the level of
autonomy and the role that humans play. It also considers core application areas such as human language,
computer vision, automation/optimisation or robotics.

24
OECD (2022).
25
ISO and IEC (2017)

38 Governance of AI adoption in central banks

 Each of the framework’s dimensions have distinct properties and attributes that are relevant to
assessing policy considerations associated with a particular AI system. The 10 OECD AI Principles, adopted
in 201926, help to structure the analysis of policy considerations associated with each dimension and sub-
dimension.

Values-based principles Table 4.5

Such as non-discrimination and equality, freedom, dignity, autonomy of

individuals, privacy and data protection, diversity, fairness, social justice
1.People and planet
and internationally recognised labour rights. This also includes addressing
misinformation and disinformation amplified by AI, while respecting
freedom of expression and other rights and freedoms protected by
2. Human rights, privacy and fairness applicable international law. Addressing risks arising from uses outside
intended purpose.
Commit to transparency and responsible disclosure regarding AI systems.
Provide meaningful information appropriate to the context and consistent
with state of the art technology to foster a general understanding of AI
3.Transparency and explainability
systems, including their capabilities and limitations; and to make
stakeholders aware of their interactions with AI systems, including in the
workplace.
AI systems should be robust, secure and safe throughout their entire life
cycle so that, in conditions of normal use, foreseeable use or misuse, or
4. Robustness, security and safety
other adverse conditions, they function appropriately and do not pose
unreasonable safety and/or security risks.
AI actors should be accountable for the proper functioning of AI systems
and for respect of the above principles, based on their roles, the context,
and consistent with state of the art technology. Ensure traceability, apply
5. Accountability a systematic risk management approach to each phase of the AI
system life cycle including cooperation between different AI actors,
suppliers of AI knowledge and AI resources, AI system users and other
stakeholders.

Recommendations for AI policies

OECD recommends increasing public and private investment to drive
innovation and economic growth. Encourages stronger collaboration
6. Investment in research and between industry, universities and research institutions to enhance
development knowledge. Also suggests prioritising strategic fields like health, green
technologies and digital innovation for sustainable development.
Recommends creating an environment that ensures the responsible use of
data and technologies. Highlights the need for investment in technology,
7. Data, compute and technologies the need for robust cyber security measures, and stresses the importance
of education and training programmes, among others.
Recommends creating flexible and adaptive regulatory frameworks that
8. Enabling policy and regulatory can adapt rapidly to disruptive technologies, advocates for intellectual
environment property rights and recommends policies that promote fair competition
and prevent monopolistic practices.

26
OECD (2022).

Governance of AI adoption in central banks 39

 Emphasises the need for continuous upskilling programmes to prepare the
workforce for the use of AI technologies, also recommends supporting
9. Jobs, automation and skills workers transitioning between jobs or sectors particularly for those
affected by automation. Encourage the creation of new job opportunities
in emerging sectors, such as health and green energy.
Encourages international cooperation on data standards and practices to
10. International cooperation
facilitate cross-border data flows and fosters global innovation.
Source: OECD (2022).

ISO/IEC 42001 – Information technology – artificial intelligence – management system27

This international standard provides the requirements and guidelines for establishing, implementing,
maintaining and continually improving an AI management system within the context of an organisation.
The document is intended to be helpful to any organisation regardless of the nature of their activities.
This standard consists of seven key components.

ISO/IEC 42001 key components Table 4.6

Organisations should be able to understand the internal and external

factors that can impact their AI systems, including regulatory and
1. Context of the organisation contractual obligations. This involves identifying the needs and
expectations of all stakeholders.

The board must establish a privacy policy, assign roles and responsibilities,
2. Leadership and commitment and ensure the integration of AI systems requirements into the
organisation’s current processes.

This involves identifying risks and opportunities related to the processing

of personal data, setting objectives for the AI systems, and planning actions
3. Planning to address these risks and opportunities. This also includes compliance with
applicable laws and regulations.

This is related to procuring the necessary resources for the AI systems,

including personnel with appropriate skills and competencies. This also
4. Support involves ensuring awareness and communication about privacy policies
and procedures, as well as maintaining documentation.

The standard outlines the processes and controls necessary to meet the
requirements of the AI systems. This includes data protection by design
5. Operation and by default, conducting data protection impact assessments, and
managing data breaches and incidents.

Organisations must monitor, measure, analyse and evaluate the

performance of their AI systems. This implies conducting internal audits,
6. Performance evaluation management reviews and measuring the effectiveness of privacy controls
and procedures.

The standard highlights the need for continual improvement of the AI

systems. Some activities are related to identifying discords and taking
7. Improvement corrective actions, as well as proactively seeking ways to enhance privacy
management practices.

Source: adapted from ISO and IEC (2023b).

27
ISO and IEC (2023b).

40 Governance of AI adoption in central banks

 ISO/IEC 42001 offers several benefits for organisations like helping them with regulatory
compliance, and providing a structured approach to risk management to identify and mitigate risks related
to AI systems and personal data. It also offers enhanced customer trust while providing operational
efficiency and helping organisations to establish clear protocols for incident management.

ISO 42001 artificial intelligence management system Graph 4.4

Sources: Oh (2024); ISO and IEC (2023b); OECD (2022).

ISO/IEC 23894 – Information technology – artificial intelligence – guidance on risk management28

In close connection with the approach proposed by ISO/IEC 31000 – Risk management, ISO/IEC 23894
integrates specific considerations related to the adoption of AI systems by the organisation into the risk
management framework. With a focus on identification, assessment and mitigation of risks emerging from
AI adoption, this standard provides recommendations to the organisation to ensure that AI systems are
properly controlled and safe. Transparency of the risk management process, promotion of ethical AI use
and continuous monitoring of AI risks are among the practices suggested by this standard, whose main
stakeholders may include technical professionals such as IT managers, AI developers and risk managers.
In summary, the table below highlights the procedures required to adapt current risk
management frameworks to cover the adoption of AI systems.

28
ISO and IEC (2023a).

Governance of AI adoption in central banks 41

 ISO/IEC 23894 key components Table 4.7

Considering the benefits of an integrated risk management framework, this standard

1. Principles of AI risk and describes several principles that shall guide AI risk and impact assessments in
impact assessment organisations, such as dynamism, use of best available information, inclusivity,
continuous improvement, and human and cultural factors.

The standard provides guidance on adapting the current risk management structure to
cover AI systems, like ensuring risk integration to the business processes, maintaining
2. Risk and AI systems
the commitment of leadership towards the responsible adoption of AI, properly
management framework
mapping external and internal contexts, defining roles, accountability and resource
allocation.

The standard highlights potential impacts on individuals from AI implementation

3. Ethical impact
involving AI bias, personal data usage and even impacts on fundamental rights and
assessment
physical security.

Populations may be affected in terms of their social and cultural values by AI

4. Social impact assessment development and usage, so it is important that organisations consider this kind of
impact in their risk assessments.

Given the nature and amount of data required by AI systems, organisations should
5. Data governance and
assess and mitigate risks related to data governance and privacy, particularly with the
privacy risk assessment
use of large data sets for training AI systems.

Source: adapted from ISO and IEC (2023a).

ISO/IEC 38507 – Information technology – governance of IT – governance implications of the use

of artificial intelligence by organisations29
On the other hand, ISO/IEC 38507 discusses the consequences for an already existing governance structure
as the institution considers the adoption of AI in its processes. The huge differentiation between traditional
and AI systems and the recognition of their vast implications on the emerging relations between human
and AI systems are the backdrop for the practices discussed in this standard, which provides guidance to
the governing body for efficient, effective and acceptable adoption of AI within the organisation.
The document highlights the importance for the organisation to keep current governance pillars,
like human supervision and accountability over automated decision-making processes, while being
compliant with current, and potentially new, internal and external obligations (like ethically and/or legally
defined usage of data, which covers privacy and data protection – even for training models). In order to
ensure adherence to best practices in this new reality, the organisation shall consider building a clear and
detailed map of AI systems (AI ecosystem) operating across the institution, which will provide the
governing body with necessary information regarding data usage, model transparency and outcome
explicability.
Finally, keeping human supervision and accountability over AI-aided decisions in all executive
levels may require revision of internal commands that define the governance. Review of internal policies,
updates to the risk management framework and impacts on the organisation’s culture are expected in
order to align safe AI usage to the organisation’s strategic objectives.
This detailed standard has its main elements described in the following table.

29
ISO and IEC (2022b).

42 Governance of AI adoption in central banks

ISO/IEC 38507 key elements Table 4.8

Given the possibility of a great impact on the organisation’s activities, this standard
1. Keeping governance provides guidance on assessing the adequacy of the current governance structure in
the light of AI adoption.

Roles and responsibilities of the governing body (eg board of directors) in overseeing
AI initiatives are highlighted in this standard. It emphasises the need to develop/review
2. Responsibilities of the
governance structures, decision-making authority and accountability frameworks in the
governing body
face of greater technological dependency and the need for greater transparency and
explainability brought by AI systems.

The standard provides guidelines on how the governing body should define the
organisation’s AI strategy, ensuring that AI investments align with long-term goals,
3. AI strategy and provide value and are in line with its risk appetite. It also suggests attention to new
investment needs related to the technology, like improvement compliance supervision and control,
assessment of the impact of usage of AI across the organisation, and special care with
legal requirements and the consequences of deploying AI.

4. Policies review and Governing body shall ensure that proper policies, responsibility chain and human
decision-making supervision are in place for the controlled use of AI, since automated decision-making
governance delivered by AI systems does not change the responsibility of the governing body.

The standard recommends robust data governance for responsible and effective AI
5. Data governance usage. Data collection, treatment and storage processes shall be enhanced, in order to
ensure quality in data processing and output. Bias analysis shall also be performed.

As an AI system does not understand context (moral values and common sense etc) like
6. Culture and values humans, it is important that the governing body is explicit about the organisation’s
culture and values and is able to monitor, and when necessary correct, AI’s behaviour.

The governing body should seek assurances that management configures and
maintains any AI system used by the organisation to meet compliance obligations and
avoid compliance violations, such as pricing mechanisms that violate antitrust legal
7. Compliance requirements or the use of data for training that violates civil rights or is discriminatory.
Compliance management shall also be enhanced to cover new needs, like the
sophistication of AI systems (new controls may be implemented) and human
reassessment of decisions made by AI systems.

As AI adoption provides risks and opportunities, it is paramount that the current risk
management process review examines whether the risks involved are fully understood
8. Risk management and managed, especially in decision-making processes, data usage, culture and values
development, and compliance. If so, the governing body must be aware of the
acceptability of those risks with regard to its stated risk appetite.

The standard reminds the reader that the organisation’s objectives and assets shall be
carefully considered in an AI adoption scenario, like accountability, duty of care, physical
9. Objectives
safety, security and privacy, transparency and data itself (whose protection and integrity
may be considered an organisational objective).

Governance of AI adoption in central banks 43

 The governing body shall be aware of risks that depend on the nature and domain
where an AI system is deployed, as well as on the maturity of technologies used by the
organisation. Other risks, already mapped on current IT processes, may emerge with
10. Sources of risks new and unknown intensity, like risks related to data sources, poorly specified systems,
value chain, undesired bias, lack of explicability, lack of experience in AI and cyber
threats. Impacts on human professional autonomy, contractual and environmental
issues may also be expected.

Source: ISO and IEC (2022b).

To sum up, these three ISO/IEC standards can be seen as complementary, each addressing a
different layer of responsibility and scope within the life cycle of AI adoption, from operational
management and risk assessment to high-level governance and strategic oversight.

44 Governance of AI adoption in central banks

Appendix 5: Members of the Artificial Intelligence Task Force

Co-chairs
Bank of Mexico Alejandro De Los Santos

Bank for International Settlements, Angela O’Connor

Basel

Members of the task force

Central Bank of Brazil Germano Machado

Bank of Canada Stefan Smith

Maryam Haghighi

Central Bank of Chile María Jesus Orellana A

Central Bank of Colombia María Carolina Trespalacios

Federal Reserve Bank of New York Li He

Bank of Mexico Alberto Mendoza

Aldo Hernández

Julieta Carmona

Gloria Guadarrama

Alfonso Murillo

Bank for International Settlements, Steffen Grosse

Basel Graham Cameron

Secretariat

Bank for International Settlements, Christian Upper

Representative Office for the José Aurazo
Americas
Torsten Ehlers

Governance of AI adoption in central banks 45