The promise is easy to sell. Sit for a long interview, let an AI learn how you talk and work, and from then on a version of you can answer questions in your place, so no one has to bother the real you again. A recent study built exactly that kind of stand-in and then checked it against the people it was copied from. The finding is clean and a little deflating : the copies resembled the group but not the person.
The researchers call their agents distribution-calibrated but identity-imprecise. In plain words, a whole crowd of these agents produced roughly the mix of opinions that a real crowd produced, yet each single agent was an unreliable stand-in for the one human it was grounded in. To work through what that means, Polora put the same question to several AI models built by different companies and had them examine the study together.
What the study measured
The setup was direct. 51 knowledge workers in the United States, all in document-heavy roles, each sat for an in-depth workflow interview, scheduled for 90 minutes and yielding around 43 minutes of recorded speech on average. From each transcript the researchers built one personal agent, using a pipeline that stores the interview, retrieves the relevant parts when a new question comes up, and has a language model reason and answer as that participant.
Then both the humans and their agents evaluated the same four imagined AI tools for working with documents : one that answers questions across many files, one that highlights key passages, one that turns documents into audio, and one that carries out multi-step tasks on its own. Each was scored on plain measures of usefulness, ease of use, and willingness to adopt or recommend, alongside open-ended written feedback. The people took the same test again three days later, which gave the study a benchmark for how consistent a person even is with themselves. The core check was blunt : line up each agent's answers against the answers of the very person it was built from.
A crowd matched, one person did not
At the level of the individual, the agents were weak. On exact agreement with their own human, the interview-based agents matched about 30% of the time, against roughly 45% when the same people simply answered again days later. The study frames that as around 67% of human consistency, which is 30 divided by 45, not 67% of answers correct. Feeding the agent the full interview barely helped : it was not clearly better than an agent given only a few demographic facts, or even one given nothing at all.
At the level of the group, the picture flipped. When you stopped asking who said what and looked only at the overall spread of answers, the interview-based agents came closest to the real spread. They were the only version that reproduced the low, unenthusiastic scores at all. An agent given no personal information simply collapsed toward a safe default, parking most answers near the top of each scale. So the group of agents looked like a real group, while any one agent did not look like its real person.
Why matching the crowd is not matching you
Picture ten people, five who like a concept and five who dislike it. A simulation could reproduce that five-and-five split perfectly while getting every single person backward. The histogram is right and every personal guess is wrong. That is an illustration rather than a claim that the agents literally reversed everyone, but it shows how a correct distribution can hide a pile of individual errors. Averaging lets the mistakes cancel out.
Several ordinary reasons fit what happened. An interview captures part of someone's world, not every condition that decides a reaction. Judgments about a tool you have never touched are often built on the spot, not read off a settled set of preferences. Where the record runs thin, a model fills the gap with plausible, generic reasoning that sounds like a typical worker rather than this one. Knowing the shape of a group tells you almost nothing guaranteed about a particular member of it.
Where the smoothing shows
The written answers make the gap concrete. For the audio tool, humans and agents raised the same headline concerns : accuracy, privacy, fitting into existing software. But roughly 30% of the real people rejected the idea outright, some saying they would not use it under any circumstances, while only a single agent pushed back that hard. People named specific dealbreakers, such as working on a desktop rather than a phone, or needing to read a document rather than hear it, or a compliance rule at their firm. The agents tended to reframe those hard stops as problems that some improvement could solve.
That difference matters because "I would adopt this if it improved" is not the same answer as "this does not fit my work." The agents also scored worst on voice and tone, producing tidy prose where people spoke in messy, hesitant, sometimes contradictory ways. One caution belongs here : these qualitative comparisons were graded by AI judges that only moderately agreed with each other, so they are supporting evidence, not a final verdict.
What it means for a copy of yourself
If a vendor or a research team offers you a persona made from your own interview and says it can stand in for you, this study is a reason for caution. The safer reading is that its new answers are predictions about you, not statements by you. Familiar vocabulary and references to your job show that it used your material. They do not show that its conclusions represent you.
In practice that means a persona built from your words should not be trusted to reproduce your vetoes, to explain why you would hesitate, or to judge how a tool would fit your actual workflow. It should not consent on your behalf, decide on your behalf, or be quoted as your view. Any answer that would be used as if you had said it should come back to you for confirmation.
Where it earns its keep, and where it does not
There is still a real, narrow use. Because a group of these agents resembles a real group, they can be a cheap first filter : screening many rough concepts before spending on prototypes, comparing broad directions such as more automation against more control, or generating candidate objections and questions to carry into real research. The study puts the running cost far below a human session, roughly $1.27 against about $12.50, though it is not fully clear the two figures are measured over the same unit of work.
Where it does not belong is anywhere the job is to be a specific person. It cannot speak for a named participant, establish someone's trust or adoption barriers, replace interviews when the workflow details are the point, or turn a synthetic comment into customer testimony. Even the screening use comes with a warning : matching the spread of today's answers does not prove the agents would rank tomorrow's concepts correctly, so it remains a promising idea rather than a demonstrated one.
One study, not a verdict
This is a single, domain-specific study : 51 American knowledge workers, four imagined document tools, one way of building the agents. It does not prove that a faithful personal simulation is impossible, and better methods may narrow the gap. What it does is move the burden of proof to the right place. Grounding an agent in your words is not the same as making it stand in for you, and that second claim now needs to be shown, not assumed.
The working rule that falls out of it is simple. Use a crowd of these agents when you only need the crowd, for a quick and rough aggregate signal. Use real people for anything that has to be a particular person, including your own decisions about a tool that would sit inside your work.
