Can AI Replace Survey Respondents? What Pew's Synthetic Sample Study Found

“Synthetic respondents” have become a popular pitch over the past two years: instead of surveying real people, ask a language model to answer as a 45-year-old nurse in Ohio, a thousand times, and use the results. It is fast and cheap, and several research tools now offer it as a feature.

On 30 September 2026, Pew Research Center published a study that tested the idea against its own high-quality survey data. The headline is in the title: Can AI Stand In for Human Survey-Takers? Not Really.

What Pew Actually Tested

Pew took three waves of its American Trends Panel, a probability-based panel of US adults, and generated AI “respondents” matched to the characteristics of the real people who answered. It then compared the synthetic answers with the human ones across nearly 300 questions.

As of early October 2026, the reported results include:

• An average absolute error of about 12 percentage points between synthetic and human answers.
• Larger errors for some groups, with reported averages above 15 points for some demographic subgroups.
• Synthetic respondents that were far more “knowledgeable” than real people, answering factual questions correctly much more often than the human panel did.

That last finding is worth dwelling on. A model asked to role-play an average person still answers like a model: well-informed, consistent, and confident. Real people are none of those things all the time, and that variation is exactly what surveys exist to measure.

Why This Matters Beyond Polling

You may never run a national opinion poll. But the same pattern applies to smaller uses of AI personas that have become common at work:

• Asking a chatbot “how would our customers react to this pricing?”
• Generating synthetic interview transcripts to “test” a product idea.
• Using AI-written reviews or survey answers to fill out a thin dataset.

In each case, the output looks like data. It is actually the model’s average expectation, smoothed and confident. It can be useful for drafting survey questions or brainstorming hypotheses. It is not evidence about what real people think.

Where AI Personas Can Still Help

Pew’s finding does not make language models useless for research. They are reasonable tools for:

• Drafting and testing survey wording, catching ambiguous or leading questions before real people see them.
• Summarizing open-ended answers from real respondents, with human checking.
• Generating hypotheses to test with actual data.

The line is between using AI to prepare or process research and using AI as the source of the findings. If a report or slide cites AI-generated answers as if they were customer views, it should say so clearly, and readers should treat it accordingly. The broader concern about AI-generated filler at work is the same one discussed in workslop and your professional reputation.

Writing About Research Honestly

If you write reports, proposals, or marketing copy that cite research, this is a good moment to tighten your wording:

• Say where data came from: “a survey of 412 customers in September” rather than “research shows”.
• Label synthetic or simulated results explicitly.
• Do not let a rewrite tool upgrade “a few customers told us” into “customers overwhelmingly prefer”.

That last point is a real risk. Language models tend to make claims sound more certain than the source supports, which is one reason a rewrite should always be checked against the original. How to keep a paper trail of your AI decisions covers a lightweight way to record what came from where.

A Wrivio Context for research summaries could say:

Rewrite this research summary in clear, neutral language. State the source, sample size, and date of every finding exactly as written. Keep hedges such as “some”, “a few”, and “in our sample” unchanged. Do not strengthen claims, add percentages, or generalize beyond the sample described.

Press Ctrl+Shift+Space, paste the draft, and read the diff specifically for certainty creep.

Common Questions

What did Pew find about AI survey respondents?

In a study published on 30 September 2026, Pew found that AI-generated respondents differed from real survey answers by about 12 percentage points on average, with larger gaps for some groups, and concluded they are not an adequate replacement for traditional polling.

Why are synthetic survey respondents inaccurate?

Models tend to answer like a well-informed average rather than reflecting the real variation, uncertainty, and gaps in knowledge among actual people. Pew found synthetic respondents answered factual questions correctly far more often than humans.

Can I use AI to help with surveys at all?

Yes, for drafting and testing question wording, summarizing open-ended responses from real people, and generating hypotheses. The findings themselves should come from real respondents.

Should I disclose AI-generated research data in a report?

Yes. Label any simulated or AI-generated results clearly so readers do not mistake them for real customer or public opinion data.

Download Wrivio for Windows to tighten research summaries without letting the wording claim more than your data supports.