Trustworthy AI

Trustworthy AI (TAI) is a framework to mitigate the various forms of risk arising from AI. Created by the US Department of Health and Human Services in 2021, the framework offers detailed guidance to ensure that health-related AI is implemented in a way that is ethical, effective and secure.  

Trustworthy AI has six major principles: Fair/Impartial, Robust/Reliable, Transparent/Explainable, Responsible/Accountable, Privacy and Safe/Secure.  

• Fair / Impartial
  AI applications should include checks from internal and external stakeholders to help ensure equitable application across all participants. Data will be assessed to ensure that they meet established fairness requirements, are representative of target populations, and have been corrected for data biases. 
   

  Example: An AI-based tool used to help public health experts flag geographic areas for targeted outreach should be trained on data that is representative of the target population so it performs well for all sub-groups.
   

• Robust / Reliable
  AI systems should have the ability to produce accurate and reliable outputs consistent with the original design and to learn from humans and other systems. Data will be assessed for accuracy, data drift, and data contamination. 
   

  Example: If new research comes out about the relationship between biomarkers and brain deterioration, an AI model that predicts brain deterioration should be retrained to reflect the updated research.
   

• Transparent / Explainable
  All relevant individuals should be able to understand how AI systems make decisions and how data is being used; algorithms should be open to inspection. Relevant stakeholders will need to be consulted to ensure there is effective communication with all groups impacted by the AI system.
   

  Example: If an organization uses AI to match patients to clinical trials, the organization needs to be able to explain the solution’s decision-making process to patients, providers and researchers.
   

• Responsible / Accountable
  Policies should outline governance and who is held responsible for all aspects of the AI solution (e.g. initiation, development, outputs, decommissioning.) There should be people responsible for continuously monitoring against drift, bias, or failure, and people should be able to retrace how an AI system arrived at a certain recommendation or solution.
   

  Example: If stakeholders observe that an AI pharmacy benefits processing tool used to aid pharmacists is inappropriately accessing system data, they should be able to quickly identify the human custodian to resolve the issue.
   

• Privacy
  Privacy for all Individuals, groups, or entities should be respected. Peoples’ data should not be used beyond an AI tool’s intended and stated use. In addition, data used by an AI tool must be approved by a data owner or steward. 
   

  Example: A research team is studying the use of mobile technology to track symptoms of Parkinson’s disease. If a team uses the mobile data to build an AI model that provides tailored recommendations to users, they first need to obtain consent from patients.
   

• Safe / Secure
  AI systems should be protected from risks that may directly or indirectly cause physical and/or digital harm to any individual, group or entity. Security risks should be assessed for level of risk, potential impacts and what controls are in place to mitigate vulnerabilities.
   

  Example: An NLP-based AI solution that interprets handwritten medical records needs to have data encryption, user authentication and other applicable security controls to prevent hackers from stealing records.

As the Trustworthy AI Guidelines make clear, every application of AI involves tradeoffs across these dimensions. The key to an evaluation of trustworthiness is identifying risks and developing strategies to mitigate them.

Find a one-page summary of these principles here.