Your AI twin may be a stereotype with your name on it

Give a machine your answers, and it will answer as you, roughly

Picture a survey that runs across four sittings. It asks about your personality, your money habits and how you weigh risk and fairness, what you would pay for a product, and it sets puzzles that have right answers. More than 2,000 Americans did this for a research panel, answering over 500 questions each. [S02]

Researchers then gave each person’s answers to an AI model and asked it to answer new questions as that person. That copy is a digital twin. In 19 follow-up studies, planned and registered in advance and covering 164 measures, the researchers put the same new questions to the twins and to the people they copied. [S01]

The twins matched their people only weakly, with an average correlation of 0.20 on a scale where 0 means no relationship and 1 means perfect agreement. They did only modestly better than the same AI given no personal answers at all. [S01]

A twin can get most answers right and still not know you

An earlier report on the same panel sounded far more flattering. There, twins built from people’s answers got about 72 percent of held-back answers right. People who took the same tests a second time matched their own earlier answers about 82 percent of the time, so the twins reached about 88 percent of a person’s own consistency. [S02]

Both numbers can hold, because they measure different things. Accuracy counts matching answers. If most people give the same answer to a question, a twin that always gives the popular answer scores well without knowing anything about you. Correlation asks something harder: when you differ from other people, does your twin move with you?

The two figures come from different tasks, so they are not a direct contradiction. The gap still points at the first distortion the researchers named: twins that answer too much like the underlying model and too little like the individual. [S01, S02]

Five recurring errors

Random errors would scatter in all directions; these leaned in five ways the researchers named. [S01]

Insufficient individuation: the twins’ answers stayed close to what the underlying model says with no personal data at all. Stereotyping: they leaned on demographics, answering more like a typical member of a group than like the particular person. Representation bias: they were more accurate for people with more education, higher incomes, and moderate politics and religious habits. [S01]

Ideological bias: they gave more pro-technology and more pro-social answers than people did. Hyper-rationality: on factual questions they showed near-perfect knowledge where real people got things wrong, and they chose the textbook answer more often than the people they stood in for. [S01]

A separate study found the same flattening one level up. When language models were asked to answer as members of demographic groups, they misportrayed those groups and erased the differences within them. [S05]

Random errors would scatter in all directions; these leaned in five ways the researchers named.

Predicting a crowd is easier than predicting a person

Other work finds these models far more capable, and the difference is instructive. In a 2026 study, language models simulated the responses to 70 survey experiments that had involved about 119,000 real participants. Their predicted effects lined up closely with the real ones, with a correlation of 0.85, and 0.90 for studies too new to be in the models’ training data. That is about as well as human forecasters did, though the models tended to overestimate how large the effects were. [S03]

That task asks what happens on average when thousands of people see one message rather than another. Averages are where a model that leans toward the typical answer does well. A twin has to be right about one person, including the ways that person is not typical.

Another approach uses interviews. In the original November 2024 preprint, a version posted before peer review, researchers built agents for 1,052 people from two-hour interviews in the people’s own words. The agents reproduced answers to a standard social survey 85 percent as accurately as the people reproduced their own answers two weeks later. Accuracy gaps narrowed across racial and ideological groups compared with agents given only demographic descriptions; gender gaps were already small and changed little. The June 2026 revision reports 83 percent for interview-only agents and 86 percent for agents combining interviews and surveys. These results measure accuracy, so they share the blind spot described above. [S04]

Ask your twin to be right about how you differ

Digital twins of people are being proposed for market research, policy testing and personal assistants that act on your behalf. Five researchers writing in Nature Physics argue that the term should mean something strict: a model that is updated with new data, tested against interventions, and able to be shown wrong. [S06]

By that standard, a twin that sounds like you has proved very little. A better test is whether it predicts where you differ from people like you, and whether anyone checks.

Until twins pass that test, the safer assumption is that a persuasive copy is a compressed one: your demographics and the model’s defaults, with your name on top.