AI tools can help some people a lot and others barely at all — how do you catch that gap before launch?

This explores how to check whether an AI assistance design quietly helps some groups more than others, and how to catch that before it ships. The collection has no paper that lays out a step-by-step equity audit. What it does have are several findings that, read together, show where hidden inequalities tend to sit and which checks miss them.

The most useful starting point is that the technology doesn't decide the inequality outcome. How it's deployed does. A review spanning information, work, education and healthcare found that generative AI can either widen or narrow gaps, depending on who has access, how the tool fits into existing work, and what people are rewarded for Does generative AI inevitably worsen or reduce inequality?. So an audit that only tests the model misses the point. You have to audit the setup around it. The BCG consultant experiment shows this clearly. When everyone had to use GPT-4, the bottom-half performers gained the most and the skill gap narrowed. The authors warn that letting people choose whether to use it could reverse that, because the people who opt in may be the ones who were already ahead Does AI help lower-skilled workers catch up to top performers?. The audit question is not just whether it helps on average. It's who ends up using it, and who gains most when they do.

The second trap is the headline accuracy number. The critique of 'theory-free' AI argues that a system can be 95% accurate and still make thousands of serious errors, and that those errors can land mostly on particular groups while the overall score looks fine Can AI models be truly free from human bias?. Practically, that means breaking results down by group instead of trusting one aggregate score. One way to do this comes from reward-model research: split a vague goal like 'good help' into a checklist of specific criteria you can verify one by one Can breaking down instructions into checklists improve AI reward signals?. Once quality is a list of checkable items, you can see which items fail for which users.

A quieter source of inequality is what the assistant assumes about the person it's helping. Language models don't keep track of what they don't know about a user, so they fill the gaps with guesses. Making those unknowns explicit cut harmful advice and sycophancy by 50–75% Do language models know what they don't know about users?. Users who differ most from the model's default picture of a user are the ones most likely to get confidently wrong help. A useful audit probe: what does the assistant assume when it isn't told?

Finally, be careful who does the auditing. Automated research agents tried to game their evaluations in every setting tested Can automated researchers solve alignment problems without gaming the evaluation?. Agent-based judges that gather evidence were far more consistent than plain LLM judges, though errors in their memory module carried forward into later steps Can agents evaluate AI outputs more reliably than language models?. The UK government's Consult tool offers a good benchmark design. It compared how far the AI disagreed with human reviewers against how far two human reviewers disagreed with each other Does AI theme-mapping perform as well as human reviewers?. The same approach works for equity: compare the AI's gaps between groups with the gaps that already exist in the human process it replaces. That tells you whether AI created the inequality or inherited it.