Loading page…
An example of how a report could combine public preferences, competing arguments, and open questions for companies.
Our example questions focus on control over model behavior, personal memory, and uncertainty. They illustrate how different preferences can be visible without declaring a winning position to be the technically safe answer.
Some harms affect people who never agreed to use the system. Individual preferences cannot settle every boundary.
How to investigate the preference: Consider an adult who wants more direct answers about a controversial subject. Then consider a user who wants permission to expose another person’s private information. Would you give the same kind of control in both situations?
Consent to a conversation is not necessarily consent to a permanent profile.
How to investigate the preference: You ask for travel suggestions for a relative with limited mobility. Next month, the assistant assumes the same constraint applies to you. Should it have asked before keeping that detail, or should a correction afterward be enough?
Too many generic caveats make it hard to recognize the warnings that actually matter.
How to investigate the preference: An answer correctly names a program but gives an unverified application deadline. Would you prefer a warning on the whole answer, a note beside the deadline, or a request to check the official announcement?
These percentages use only the fixed fictional counts in the seed dataset. They do not include your local votes. There was no sampling, recruitment, verification, or collection of real responses. A future report would need those details, alongside exact question wording, dates, missing responses, and any sponsorship.
Full transparency notesBegin with a defined audience and recruitment method. Preserve the full question wording and options, report missing answers, and examine whether different wording changes the response. Follow up with people whose priorities are poorly represented by the options.
Separate a preference from a proposed product change. A desire for more control could lead to clearer memory settings, narrower permissions or better explanations; testing those designs is a different stage from measuring opinions. A future report should show what action followed and how its effects were evaluated.
Context for interpreting preferences and designing a real study. These sources are not evidence for the fixed simulated percentages above.
Explains sampling, question design and transparent reporting. It does not validate this site’s simulated results.
Explains memory controls and deletion. Check which experience and settings apply to your account.
Evaluates support for individual factual claims rather than treating a long answer as entirely right or wrong.
A framework for identifying, measuring and managing generative AI risks across the system lifecycle.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence