What should I require from AI suppliers on transparency, bias and training data?
Answered
Before you contract, vendors must be able to explain how they build tools, the logic and assumptions, and how outputs are generated. Ask how they identified, evaluated and mitigated risks; what data the model was trained on; and whether there are known biases. If you buy a pre-trained model, you still need to ensure any personal data used to train it meets your privacy standards. Make ATRS-level transparency an expectation in your invitation to tender — suppliers selling tools that affect the public should be comfortable with that.
From the guidance
Primary (how) Data and AI Ethics Framework
Section: If you’re using a third-party supplier
Read this in Data and AI Ethics Framework (opens in new tab)
Secondary (normative) Data and AI Ethics Framework
Section: Scoping and data acquisition
Read this in Data and AI Ethics Framework (opens in new tab)
Secondary (normative) ATRS: Mandatory scope and exemptions policy
Section: 5.3 Dealing with commercially sensitive information
Read this in ATRS: Mandatory scope and exemptions policy (opens in new tab)
Related questions
- How should I decommission AI models and training data? Delivery, assurance and operations
- What does data-protection fairness require of an AI system? Lawful, ethical and responsible use
- Can attackers steal my model or training data through the API? Security and safe use of tools
- Can I use AI to process personal data? Lawful, ethical and responsible use
- What is data poisoning in an AI system? Security and safe use of tools