We commit to introduce processes proportionate to risk that mitigate unintended discrimination across groups in our production AI systems, enabling Fairness by Calibration.
Human and model bias is present throughout the end-to-end AI system lifecycle; bias evaluation supports understanding who benefits, who carries risk, and where a system’s errors concentrate. In systems built on general-purpose models the bias has two layers: 1) what the model inherited from training data that its deployer cannot inspect, and 2) what the application introduces through its own data, prompts, retrieval sources and thresholds.
Furthermore, fine-tuning amplifies both layers, as it rewrites the model’s behaviour and makes the result the fine-tuner’s responsibility. This is why it is important to ensure processes and frameworks are in place, proportionate to the risk involved, to understand the extent and scale of the bias present. This enables mitigation of the group-skewed risks identified throughout the system.
01 — WHERE IT FAILS
Where it fails
Bias is ubiquitous in AI systems, however negative impact concentrates where nobody is measuring, and a system built on a general-purpose model inherits skew its team never chose.
- Inherited model bias propagates into every downstream decision the system takes.
- Application data, prompts, retrieval sources and thresholds amplify skew the base model did not have.
- Fine-tuning shifts behaviour in ways that go unnoticed without re-evaluation.
02 — PRACTICAL CONTROLS
Practical controls
Teams must combine quantitative tests with domain knowledge, and rely on processes rather than tools alone.
- Run disaggregated behavioural evaluations across the groups the system affects.
- Review provider model documentation for known biases and evaluation gaps.
- Re-evaluate after fine-tuning, model updates and changes in deployment conditions.