Trustworthy AI: the HAT test
The HAT test characterises trustworthy AI as Human-centric, Accountable, Transparent, operating in an expected, legal and fair manner. Explainability and transparency are the keys to trust; PETs support privacy-enhanced AI. The mirror image is Distrustful AI.
This course characterises trustworthy AI with three traits. Distrustful AI is the mirror image → black-box decisions, unfair outcomes, no explainability, a diminished human experience.
Human-centric · Accountable · Transparent → trustworthy AI operates in an expected, legal and fair manner.
- 🧑 Human-centric → AI must amplify human agency → a positive, not negative, impact on the human condition. Help, not hinder.
- 👔 Accountable → organisations are ultimately responsible for the AI they deliver, irrespective of the number of contributors.
- 🪟 Transparent → understandable to the intended audience → technical for engineers, legal for counsel, plain for users.
Building it inside an organisation - Structure: a top-down approach supported by leadership, embedding responsible AI principles into operating systems, with governance structures of diverse roles. Operations: risk management frameworks covering privacy, accountability, auditability; systems must be robust, secure, transparent, ensuring fairness and human oversight; adhere to technical standards.
Transparent & explainable in practice - provide meaningful information to explain the development, training, operation and deployment of the system; make stakeholders aware when they are interacting with AI; and enable those adversely affected to challenge the output.
The capacity to describe an AI system, its expected impact and potential biases → requires understanding how the system operates and the data it was trained on. Together, transparency and explainability are the keys to trust.
Collect, store and use personal information in ways that protect individuals' rights. PETs are digital solutions letting information be used while protecting confidentiality → they help prevent intentional misuse as well as accidental or negligent misuse from hacks, bugs or misunderstandings of policies.