Whether technology serves people depends on decisions made inside organisations.

Image-generation models can now create realistic images of identifiable people in situations that never occurred. For one of the world’s largest technology companies, the question was which prohibitions were missing from policy, and why.

We combined policy research with scenario-based testing. First, we reviewed academic literature across law, philosophy, socio-technical studies, and computer science; current and draft regulation in more than a dozen jurisdictions; public debate; expert judgment; and frontline harm documentation.

Second, we used that research to generate a harms taxonomy and corresponding test prompts. We tested these on the image-generation model to identify where existing policy did not prevent harmful outputs. This gave the client a map of policy gaps grounded in observed model behaviour.

We gave the client a map of where its image-generation policy fell short. Each of more than twenty recommendations specified the gap it addressed, the harm it was meant to prevent, the evidence behind it, and the strongest argument against adopting it, so the policy team could decide what to change and say why.

To train AI systems well, companies often need data that has no ordinary market price. In this case, the data was personal photographs of a sensitive kind. The ethical problem was to avoid two failures at once: paying so little that the arrangement becomes exploitative, and paying so much that consent is distorted by inducement.

We built a framework for pricing sensitive training data. The work drew on analogous domains where payment is permitted but ethically constrained: research participation, medical trials, data work, biospecimen donation, and institutional compensation schemes.

We connected those cases to the academic literature on exploitation, undue inducement, vulnerability, consent, and fair pay. The resulting model answered the question what companies should pay, in different regions, when the data is sensitive, the participants may face unequal bargaining power, and both underpayment and overpayment create ethical risks.

We delivered a payment model with region-specific figures, fair-pay ranges, and safeguards against exploitation, inducement, and fraud. Rather than a general statement of ethical intent, the model gave the client a defensible basis for payment decisions.