Asking People What They Want From AI Is Only the Beginning

A person gives an hour to explain how an AI tool affects their work. Months later, the product team announces a change. Unless the organization connects that decision to the feedback—or explains why it went another way—the interview can feel like a listening exercise with no listener.

Anthropic’s September 29 announcement describes a study about people’s experiences and preferences concerning AI. Participants can opt to publish a complete interview with their country while leaving out account identity. The effort builds on earlier research, but a self-selected interview pool is not a representative mandate and does not bind governance decisions. Source

My position: consultation should come with a public decision ledger. It should show which themes changed a product or policy, which did not, and why. The ledger should preserve disagreement instead of converting a complex set of views into a single “people want” statement.

Participation is a beginning, not an outcome

Interviews can reveal experiences that a benchmark or usage dashboard will miss. A person may explain why a feature feels coercive, how an AI tool changes a workplace relationship, or what an interface assumes about language and ability. That qualitative detail can help a team discover a problem worth investigating.

But who participates matters. People who volunteer for an interview may be more interested in AI, have more time, or feel strongly about a particular issue. A study should describe recruitment and limits. It should not present a self-selected set of conversations as though every community had been represented equally.

Consultation stage	Accountable practice	Reader’s check
Invite	Explain purpose, selection and participation terms	Who could take part, and who may be missing?
Listen	Preserve context and consent choices	What information will be public, and what remains private?
Decide	Connect themes to a decision and its owner	What changed, and who made the call?
Report back	Record accepted, rejected and deferred ideas	Is disagreement visible, with reasons?

This is a proposal for accountable practice, not a description of commitments made by the study. It offers a way to evaluate whether consultation has a path beyond collection.

Make the decision trail visible

A decision ledger could list a feedback theme, the product or policy question it raised, the evidence considered, the decision and the reason. It should include the status: accepted, rejected, deferred or still under review. A reader should be able to see whether an idea influenced a roadmap, prompted more research or did not fit the organization’s priorities.

Imagine a hypothetical set of interviews in which participants describe difficulty distinguishing AI-generated suggestions from actions already completed. A product team might decide to add a confirmation step before an external action. The ledger could connect that theme to the change and explain how the team assessed the added friction. If it chose not to make the change, it could describe the competing concern and the evidence behind that decision.

The example is fictional. The point is that public feedback does not need to dictate outcomes to matter. It needs to be taken seriously, assessed against explicit criteria and answered in a form participants can inspect.

The ledger should also preserve dissent. If many participants support a feature and a smaller group raises a distinct risk, the summary should not erase that minority view. A decision-maker can prioritize one course while recording the costs and unresolved concerns attached to it.

Consent has to travel with the interview

Publishing a full interview can provide context that a short quote loses. It can also expose personal information even when an account name is omitted. Country, job details, a distinctive story or a combination of facts may make someone recognizable to people who know them.

Participants should understand what “complete interview” means, how redaction works, whether they can review the final transcript and whether consent can be withdrawn before publication. The choice to share should be separate from the choice to participate. A person should not have to publish an interview to contribute to research.

Consent is not a one-time checkbox if the use changes. If an interview is later repurposed for a different audience or analysis, the organization should explain that use. Responsible consultation should protect the participant’s ability to choose what becomes public.

The scalability objection is real

Decision ledgers take staff time. Organizations may receive more feedback than they can individually answer, and product decisions often depend on engineering, legal, operational and business constraints that cannot all be published. A detailed response to every suggestion could slow work without changing the outcome.

The practical answer is to report themes and decisions, not promise a personal reply to every participant. A team can group similar concerns, publish how it prioritized them and name the office responsible for follow-up. It can protect confidential deliberations while still stating the rationale at a useful level.

That process also helps teams internally. A structured record makes it harder for feedback to vanish between research, policy and product groups. It gives future teams a way to see why a decision was made and what evidence might justify revisiting it.

Measure the loop, not the number of interviews

Organizations should publish whether participants came from different contexts, what the study cannot represent, and how the research affected decisions. The count of interviews may indicate effort, but it does not show whether the organization listened or responded.

Useful measures might include the share of major themes with a documented disposition, the time between research and response, and the number of decisions revisited after new evidence. These should not become simplistic targets that reward paperwork. A short, honest record can be more valuable than an elaborate process that produces no discernible answer.

Public consultation will not make AI governance democratic by itself, and interviews do not substitute for regulation, worker participation or independent research. They can give people a meaningful channel into decisions if the organization is willing to show the route from experience to action.

Participation becomes concrete in design choices such as the correction controls discussed in sign-language AI input. At the institutional level, our disclosure essay argues for recurring evidence that outside readers can verify rather than a one-off statement of intent.

Asking what people want is the opening. Accountability begins when the organization shows what it heard, what it chose and what it is still willing to reconsider.