Can AI personas stand in for real users?

Would AI personas have found the problem On Behalf solves?

The September incident happened because a real person tapped Allow always. If AI personas behave like real people, some of them should make the same choice. If none do, a team relying on them would never have seen the problem.

Ten people described by their circumstances

I wrote ten UK personas from their circumstances: age, where they live, work, money and how they use their phone. None says what they think of AI, so their views had to come from the model, not from me. They range from 19 to 74, cover all four UK nations and include one screen reader user.

Each persona ran in its own AI agent, with no access to my work, the case study or the other personas. Each answered two parts.

First, ten questions worded as closely as I could to six published surveys: Ofcom, the UK government’s AI tracker, Pew Research, the University of Melbourne and KPMG, Accenture and Global Payments.

Second, the keyboard task from the On Behalf prototype, in words. Each persona chose what the agent may do on its own, what it must ask and what it must never do. Then they chose when it stops, and Allow once or Allow always.

I ran all ten twice. I then ran the Allow once or Allow always choice a third time on its own, without the rest of the task, in case the detailed questions had made the personas more careful. One more agent answered as “a typical adult living in the UK”, with no persona at all.

They got the direction right on the obvious questions

Eight of ten would let an agent handle routine tasks. Accenture found 74% of people would. None would let an agent buy without checking first. Accenture found 9%, which in a group of ten is about one person.

They drifted to the middle and overstated caution

Seven of ten felt equally excited and concerned about AI, against 38% of Americans in Pew’s survey. Nine of ten were unsure whether to trust AI. In the University of Melbourne and KPMG study, 46% said they were willing to. All ten wanted more control over how AI is used in their lives, against about six in ten in Pew’s survey.

Their words for AI were milder too. The personas’ most common word was “handy”. When the UK government asked real adults for one word, the most common answer was “scary”.

None of 30 answers chose Allow always

Not one persona tapped Allow always, in either run or in the third test with only the two buttons on screen. Their reasons were close to word for word. A retired lorry driver in Hull and a pharmacy dispenser in Bradford both said they were not “handing it the keys for good”.

They also set up the agent the same way. In 15 of 20 answer sets, all six keyboard choices were identical, and the typical adult with no persona gave the same set. Run through the prototype, that set has nothing to regret, and it buzzes the phone five times, four of them before the keyboard is collected. The personas never weighed that cost, because they never had to live with it. For real people, being interrupted is likely one reason to tap Allow always.

Running it again gave the same answers

147 of 150 answers were the same the second time, and the totals matched on every multiple-choice survey question. More runs would not have added anything. The gap from the real data stayed the same each time.

This matches published research on AI stand-ins. Bisbee and colleagues found that AI survey answers vary far less than real people’s. Wang and colleagues found that models flatten the variety within groups of people. Nielsen Norman Group compared AI users with real interviews and found them agreeable and idealised, and said they “seem to care about everything”. Argyle and colleagues found close matches to real groups using personas built from real survey respondents, which mine were not, though Bisbee’s team used the same approach and still found too little variation.

Useful for rehearsing a study, and little else

That leaves the use Nielsen Norman Group suggests: rehearsing a study before real people see it. The keyboard task worked in words, so it can be used in a remote test without the prototype.

I would not report their numbers. And I would not use them to judge a risk. The behaviour On Behalf exists to handle is the one they could not produce. So the five-person test now has a recruiting rule: include people who already let apps act for them, because they are the ones who tap Allow always.

A small test, with clear limits

Ten personas from one AI model is not a sample. My wording only approximates each survey’s, and the surveys mix UK, US and global samples. A stronger test would build personas from real respondents, as Argyle and colleagues did, and compare answer by answer.

The pattern still lined up with the research, and with the incident: the personas agreed with each other far more than people do, and they never made the risky choice.

Download every answer as a spreadsheet (CSV). If you would run the test differently, tell me.

All persona answers were written by an AI model in October 2026. The real figures were checked against each organisation’s own publication.