Keep AI Reconstructions Separate from Evidence

Organisations already use interpretation legitimately. A manager explains what a customer complaint may mean. An engineer proposes how a failure happened. A researcher describes a possible user response. The problem starts when the presentation removes the distance between that interpretation and the person or event being represented.

Consider an illustrative customer avatar in an investment presentation. It can help a team explore a service idea. It cannot establish that customers asked for the service, accepted its price or understood its risks. Those claims need interviews, observed behaviour or another suitable source. Calling the avatar synthetic does not supply the missing demand evidence.

The control is attribution, not a ban on explanation. Keep the researcher's inference in the researcher's voice. If a reconstruction helps discussion, name its purpose and assumptions. Do not give it the apparent authority of a real customer, employee or independent expert whose views it merely predicts.

Ask the decision owner to review the underlying material before the polished version. Separate direct records from summaries and illustrative additions. A source pack should make it easy to find the original statement, the surrounding context and any omissions that could change its meaning.

This does not require directors to inspect every raw file. It requires a responsible reviewer to be able to explain the path from source to claim. For a consequential assertion, preserve the source, identify the transformation and show why the transformed version remains suitable for the decision.

Test that path with a small challenge. Replace a quantity, remove a qualifying sentence or change who is speaking. Does the review notice that the conclusion should change? New research on numerical claim verification reports that small numerical changes can expose weaknesses in language-model checking, while targeted training can help on the tested datasets. It is supporting evidence for testing the reviewer, not a guarantee about a particular product. ²

Run the decision once without the synthetic presentation. Use the source evidence and a plain-language account of uncertainty. If the recommendation changes after the performance is added, ask what new information the performance supplied. If it added none, the change deserves examination rather than automatic acceptance.

This is an Institute recommendation, not a claim that every emotional response is irrational. Experience and human testimony can convey important consequences that a spreadsheet misses. The aim is to keep the person who experienced those consequences distinct from someone else's model of that person.

Provide a route to challenge attribution before approval. An employee representative may question whether a simulated worker speaks for colleagues. A customer researcher may reject a composite persona as evidence of willingness to pay. A technical reviewer may identify animation that hides an unresolved failure. Give those objections an owner and a recorded answer.

The Arizona ruling concerns a US criminal sentencing process. Boards elsewhere should not copy its legal conclusion into employment, procurement or investment policy. Ask local counsel which evidentiary, privacy and likeness rules apply to the actual use. The management practice can travel without pretending the law is identical.

Language also changes the review. A translated or dubbed statement can alter emphasis, uncertainty or social meaning while retaining a recognisable face. Review the original and the translation with someone competent in the relevant language. Do not substitute fluent output for that competence.

A new Japanese-language evaluation tested eleven models on three benchmarks with five types of typing error. Effects varied by error type. That is not a study of synthetic-video persuasion, but it reinforces a useful limit: performance on clean text is not enough to establish reliability in local input conditions. Evidence for other languages and institutions must be gathered, not assumed. ³

For the next material AI-assisted proposal, ask its sponsor to mark each decisive claim as a record, an interpretation or an illustration. Start with claims that affect funding, employment, customer rights or safety. Resolve ambiguous attribution before the meeting.

After the decision, retain the evidence that justified it, the uncertainties accepted and who authorised reliance on the transformed material. Preserve the synthetic version too where retention rules permit, so a later review can see what decision-makers actually encountered. Access and retention should follow the sensitivity of the underlying records.

Good synthetic media can make a difficult idea easier to discuss. Its value is explanatory. Do not let that strength quietly change the status of the evidence. A decision should remain defensible when the compelling face and voice are gone.