ATARC AIDA Guidebook - FINAL 4l

This document provides an overview of considerations for building trustworthy artificial intelligence (AI) systems. It discusses how including stakeholders in developing AI systems and demonstrating systems as they are built can help address concerns and build trust. The document also lists several human-centered metrics like observability, predictability, and adaptability that can inform the design of successful human-AI partnerships. Tools are described that can help evaluate these metrics and ensure user expectations match a system's capabilities. Finally, the document stresses that AI systems should undergo continuous assessment over time to evaluate performance, incorporate user feedback, and ensure systems continue functioning safely and without prejudice as circumstances change.

Uploaded by

dfglunt

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

41 views1 page

ATARC AIDA Guidebook - FINAL 4l

Uploaded by

dfglunt

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 1

Artificial Intelligence and Data Analytics (AIDA) Guidebook

As the designers and builders of the system communicate with the stakeholders, they need to
include the ongoing HCAI attributes. By addressing these considerations upfront, the
stakeholders that may be resistent to change can be assured that the team is following an
approach that builds trust. By demonstrating the AI system as it is built and including these
considerations in the demonstrations, stakeholders can be confident that the team is covering
all the bases in buidling a trustworthy AI system.
For further consideration, MITRE’s HMT Systems Engineering Guide identifies the following
HMT leverage points that can inform success (see the link below for definitions):31

• Observability
Example: Computer Vision
• Predictability
Models designed to translate visual data
• Directing Attention
based on features and contextual information
• Exploring the Solution Space
identified during training. This enables
• Adaptability models to interpret images and video and
• Directability apply those interpretations to predictive or
decision making tasks.
• Calibrated Trust

Tools are available for evaluating some of these human-centered metrics. MITRE’s Calibrated
Trust Evaluation Toolkit (https://comm.mitre.org/calibrated-trust-toolkit/), for example, helps
ensure that users’ expectations match the system’s actual capabilities. Value cards can be used
to estimate whether models would be accepted or preferred by various stakeholders (Shen et
al. 2021). There are also a variety of tools for exploring the fairness of machine learning models.

Adaptation Over Time

AI systems should be continuously assessed, evaluating consistency of the workflow and how
well humans and systems are working together. Appropriate channels will be in place for users
to easily give feedback on system performance and experience. The feedback systems have to
be easy-to-use and easily available to all of the users of the system. This increases involvement
as well as trust in the working AI systems. There may be a call for modifications as time goes on,
based on changing circumstances, technologies, or settings. It's particularly important to check
and assess adaptive or teachable systems to ensure that learning continues to function as
intended. These systems can be continuously usable, dependable, safe and trustworthy; free of
prejudice and ethically proper.

31
https://www.mitre.org/sites/default/files/publications/pr-17-4208-human-machine-teaming-systems-
engineering-guide.pdf, pp. 1-3.

Page 44