A Primer on Artificial

Intelligence (AI)
Overview

What is AI?

AI branches and techniques

The emerging landscape

Skills, data, and infrastructure supporting AI

AI uses and opportunities in the Government of Canada

Risks and ethical considerations

Legal and policy environment

Annexes: guidance, available learning, and AI assistants

 AI, data, and algorithms
Artificial intelligence (AI) AI Process Flow
An AI system is a machine-based system that infers how to
generate outputs such as predictions, content,
recommendations, or decisions from the input it receives Data input Processing Outputs
AI is also a category of technologies; a common explainer
• Speech • Content
Interpreting and
is “technology that performs tasks that would ordinarily • Text • Predictions
processing data
require biological brainpower to accomplish, such as • Images • Actions
through algorithms,
making sense of spoken language, learning behaviours, or • Numerical • Problem-solving
models, and rules
solving problems”1 values

Data
Data refers to structured and unstructured values such as
numbers, text, images, and videos. AI systems get their Revisions Assessment
value from processing massive amounts of data – and are
generally required to process that data in the first place
To the model or • Analysis
data • Feedback
Algorithm
An algorithm is a set of rules or instructions a machine (and
especially a computer) follows to achieve a particular goal
What is AI? AI branches and techniques
AI is best thought of as a set of interconnected fields and subfields. Rather than a single AI technology, a range of different
techniques and approaches are used to solve different problems

Reinforcement techniques that allow computers to improve outputs over time by testing multiple processing approaches within
Machine Learning (ML) the model, and assessing outputs against success benchmarks, then adjusting

Robotic Process Uses automation technologies, scripts, and rules to fulfill administrative tasks such as inputting data, digitizing forms, and
Automation connecting data sources. May or may not use machine learning

Methods to acquire and make sense of digital images, usually divided into activities recognition, images recognition, and
Computer Vision
machine vision

A class of algorithms loosely modeled after the neuronal structure of the brain that improves its performance without being
Neural Networks
explicitly instructed on how to do so

Natural Language Tools that interpret text (or transcribed speech) for analysis or to allow conversational interaction with software (e.g., Chatbots,
Processing (NLP) GenAI). May or may not use machine learning
Expanding into the mainstream
Generative AI, large language models, AI assistants, and bots
have quickly emerged and are becoming increasingly mainstream

A category of AI that accepts natural language and other media

Generative AI (GenAI) prompts to generate new content (text, images, audio, or other
forms of data) that is statistically probable in response to a prompt

Large language model LLMs power generative AI. They use machine learning algorithms
(LLM) to process vast amounts of data and generate human-like textual
responses based on that data

Software that uses AI to increase productivity to streamline and

AI Assistant automate workflows, generate content, connect software, manage
calendars, support decision-making, and more

A software application that performs automated tasks on the

Bot internet and within systems based on human instructions
provided through programming
 Unclassified | Non classifié

Timeline and
Emerging emerging landscape
lndscap

1970 2000 2010 2020

Automated
Chat Bots
transcription Intelligent
Noise reduction and Document and virtual
enhancement content analysis assistants
Speech Facial
recognition recognition
Language Generative pre-trained
translation transformer (GPT)

Language Processing & Understanding • Generative: capable of producing coherent and contextually relevant outputs
• Pre-trained: on extensive datasets to analyze the relative probability of
Communication & Interaction
content following other content
Multimedia Analysis • Transformer: mechanism that helps to understand words in context
 Unclassified | Non classifié

Timeline and
Emerging emerging landscape: recent
lndscap

1970 2000 2010 2020

Radical acceleration in the last decade: Why?

• Increase in available data
• Access to software including open source
• Increase in computing power / decrease in cost
• Shared or proprietary models, algorithms, and techniques
• Commoditization and mainstreaming of AI tools
 Unclassified | Non classifié

Timeline and
Emerging emerging landscape: now
lndscap

1970 2000 2010 2020

Emerging now:
• Generative AI that includes references to sources (called
Retrieval-augmented generation (RAG)) to increase the
reliability of AI systems
• Generative AI incorporating GC- and program-specific
data behind GC firewalls
• Multimodal generation – text, audio, images, and videos
 Unclassified | Non classifié

Timeline and
Emerging emerging landscape: next
lndscap

1970 2000 2010 2020

What’s next:
• Continued investment and experimentation in this space will
generate new tools and uses, though AI is susceptive to “hype
cycle” predictions and the most productive, sustainable uses will be
revealed over time
• Mature AI systems still represent narrow AI (AI systems good at
specific tasks), not artificial general intelligence (AI systems that
can self-improve at a wide range of tasks)
Skills, data, and infrastructure supporting AI
For every AI use, there’s a lot of work
and tooling below the surface. Ultimately, APPLICATION
“Using AI to do [X]”
every AI project depends on data

MODELS DATA
Training data, choice of AI approaches, Data that’s ideally cleaned, normalized, unbiased, free
statistical methods, algorithms of personal or copyright data, and often massive and
costly to analyze. Cleaning and preparing data for
analysis and processing can be 50%+ of the work and
cost involved
SKILLS
Data science, mathematics, statistics, research,
data collection, and programming INFRASTRUCTURE AND TOOLS
Substantial and specific computing power and/or access to
Cloud-based tools, open source or proprietary models, and
specialized software
BUSINESS QUESTIONS
Ultimately, the foundation of AI work is a
well-defined business problem to solve
Opportunities and current uses for AI in the GC

Better public Service and Risk-based regulation Open policy

decisions operations and compliance development

• Analysis of complex data and • Automation of routine activities • Surveillance, monitoring, and • Analyzing public input
information tracking • Sustainable and inclusive growth
• Rapid provision of
information and analysis • Safety and security• Discourse analysis
• Big data analysis • Targeting testing and
• Accessibility and inspections • Increased business and
translation support stakeholder intelligence
• Regulatory
• Service demand analysis
forecasting
Risks of AI in the GC

Keeping up with a changing technology environment in operations

Efficiency lag while the GC develops, procures, or adopts policy-compliant solutions and approaches

Securing increasingly substantial, valuable, and linked data holdings

Policy considerations regarding AI use in all sectors

Protecting public trust when using AI for government business

Informational advantage of external parties leveraging AI in dealings with government

12
Ethical considerations
As AI becomes more advanced and use becomes widespread, there is a greater risk that it may - even unintentionally - be misused, perpetuate
inequality, or exacerbate existing societal problems. This list includes only some of the many ethical considerations. Others may include impacts on
job markets, environmental impacts, and questions about humans’ relationship with technology

Bias and fairness Transparency and Privacy, security, Data provenance Manipulation
accountability and governance and copyright and deception
Bias in AI means unfair Transparency is required AI systems process massive Data provenance refers to This category involves the
decisions or showing skewed around how AI systems amounts of data, and AI tools the origins, ownership, ethical considerations related
outputs. The GC has a operate and, if they support are often Cloud-based or collection, and reliability of to the use of AI in generating
responsibility to make sure that decision-making, how data based on externally created source data. Organizations and disseminating misleading
AI tools treat everyone fairly was analyzed to produce resources, code, or models. using data may need to track of false information.
and without discrimination. outputs. This includes and document the sources,
openness, clarity, Any data processing of transformations, and External entities may use or
Bias can be a product of the traceability, and personal or sensitive data usage of data throughout propagate disinformation,
algorithm/model or the explainability of the AI needs to be governed and data lifecycles. misinformation, or deepfakes
training or input data. system. protected. using AI.
Many datasets powering
In short, historical data with a Actors – individuals or Generative AI in particular Generative AI may create
context of systemic racism and organizations – leveraging AI have massive, opaque data content that includes false
discrimination is likely to result systems may not feel, or sources that likely include information.
in a biased AI output. take, responsibility and personal information, or
accountability for actions, direct or derivative
outputs, or decisions made by copyrighted works
the system.
AI and representation
Researchers and advocates have identified a number of potential and proven risks and harms of AI. These are likely to
disproportionately impact marginalized communities

Bias and fairness

Absence or under-representation in training data: e.g., in simulations, self-driving cars were
found not to stop for people in wheelchairs, which were absent in the training data.*1
Bias in AI means unfair
decisions or showing skewed
Generative AI underrepresents minority communities in generated images of many professions;
outputs. The GC has a
responsibility to make sure attempts to correct this through the processing algorithms have so far instead created
that AI tools treat everyone stereotypical racially coded images.2
fairly and without
discrimination. Social media or content-streaming applications are likely to generate recommendations based on
“what people like you” like, siloing communities with different life experiences.
Bias can be a product of the
algorithm/model or the
training or input data.
Over-representation in training data: E.g., advocates have demonstrated that predictive
In short, historical data with a policing systems were systematically over-policing marginalized communities, over-represented
context of systemic racism in the training data.3
and discrimination is likely to
result in a biased AI output.

*Jutta Treviranus, Director of the Inclusive Design Lab at OCAD, notes that AI may be a “double-edged sword” for people with disabilities: there’s a danger of decisions or systems based on data that excludes them,
or systems generating outputs that serve the majority because they’re designed for efficiency. On the other hand, automation technologies (e.g., self-driving cars) could also create options and supports for people.
The policy
The legal and
andpolicy environment
legal environment
For Government of Canada internal use: For industry and society:

• Directive on Automated Decision-Making • Artificial Intelligence and Data Act (in development)
• Guide on the Use of Generative AI • Voluntary Code of Conduct on the Responsible
• OCIO has just launched an initiative to shape Development and Management of Advanced
an AI Strategy for the federal public service, Generative AI Systems
scheduled for completion in Fall 2025

The Artificial Intelligence Federal institutions are Can consult Chief

The Directive requires
and Data Act was tabled expected to align with Information and
an Algorithmic Impact
in 2022 as part of Bill C- the principles of fair, Security Officers and
Assessment where GC
27. As of January 2024, is accountable, secure, use systems that have
use of AI supports, or
in committee for study transparent, educated, gone through privacy
renders, administrative
and relevant (FASTER) and security screening /
decisions about
use of generative AI the AI Supply List
individuals
Annex A: FASTER principles for Generative AI
Ensure that content from these tools does not include or amplify biases and that it complies
Fair with human rights, accessibility, and procedural and substantive fairness obligations.

Take responsibility for the content generated by these tools. This includes making sure it is
Accountable factual, legal, ethical, and compliant with the terms of use.

Ensure that the infrastructure and tools are appropriate for the security classification of the
Secure information and that privacy and personal information are protected

Identify content that has been produced using generative AI. Notify users that they are interacting with an AI
Transparent tool. Document decisions and be able to provide explanations if tools are used to support decision-making.

Learn about the strengths, limitations and responsible use of the tools. Learn how to create
Educated effective prompts and to identify potential weaknesses in the outputs.

Make sure the use of generative AI tools supports user and organizational needs and contributes to
Relevant improved outcomes for Canadians. Identify appropriate tools for the task; AI tools aren’t the best choice in
every situation.
Guide on the use of Generative AI - Canada.ca

16
Annex
The B: AI Meeting
Emerging Assistants
Landscape of AI

An emerging trend is AI tools in videoconferences

AI assistants that integrate

within videoconferencing AI assistants that
software and provide attend meetings as
transcripts, summaries, participants
action items, and more

17
Annex B: Use of AI Meeting Assistants

Innocuous Automated Language Noise reduction Accessibility

transcription translation and enhancement features

Increasingly Intelligent virtual

Smart scheduling
Meeting summaries Document and
expected assistants and action items content analysis

Gesture Facial
Engagement tracking Sentiment analysis
recognition recognition

Contentious
+ more emerging uses as the market for AI assistants
is seeing rapid expansion and experimentation

18
Annex B: Real-Time Business Intelligence in Meetings
AI assistants are already being embedded in major enterprise software applications (e.g.,
Word, Excel, Teams, PowerPoint)

In a meeting context, that means integrated with Customer Relationship Management

(CRM) systems (e.g., Dynamics, Salesforce)

Parties meeting with government officials will increasingly have rich, real-time, access to:

Facts and figures Sentiment analysis

Policy positions Suggested conversation prompts

Notes from every previous meeting with an official or an organization

19
Annex B: Privacy, Ethics, IM, and Security in Meetings

If a meeting is being transcribed and

summarized, it’s being recorded

Bots and integrated software may not trigger the For recordings and
recording notification transcripts recorded by GC
officials, records will fall
Transcripts and recordings are then held by: under IM and ATIP legal
• Other organizations and policy frameworks
• Often cloud-based AI assistant providers

Some AI assistant software could have weak data

protection – or be explicitly designed to collect
data

20
Annex B: Managing for Privacy in Meetings

Ensure that all participants in a meeting are identified

and known

Consider whether assistants – human or AI – should be

involved in the meeting discussion

In general, we cannot assume that there’s no AI and

no recording - a participant could have a standalone
device recording, or a home assistant like Alexa

Ultimately, this is a question of trust in the participants

and the sensitivity of the discussion

21
Annex C: CSPS learning resources
Discover Artificial Intelligence
Courses Using Generative AI in the Government of Canada
Ethical Considerations in AI

Decoding AI Assistants in Online Meetings

Using Large Language Models (like ChatGPT) in the Federal Public Service
Microlearning Articles
Demystifying Artificial Intelligence
OpenAI's ChatGPT Explained

Events Artificial Intelligence Series: ongoing

22