Can a PR Survey Use AI Respondents?

Why this is a PR question

Commissioned surveys are one of the oldest tools in earned media. A company pays for a poll, finds a striking number — “68 per cent of finance leaders say…” — and builds a pitch or a press release around it. The number is the story, and the company's name rides along in every article that repeats it.

Fielding a real survey takes weeks and costs money. AI models offer an apparent shortcut: describe the respondents you want, ask the model to answer as each of them, and have results by the afternoon. These are usually called synthetic samples or silicon samples. The question for anyone who sells or buys survey-led PR is whether the numbers they produce are numbers about people.

On 30 September 2026 Pew Research Center published the most careful public test of that question we have seen, in an overview, five companion reports and a methodology note. The headline, from Pew itself: “at this time, AI models are not an adequate replacement for traditional polling on topics of broad public importance.” This guide sets out what Pew did and found, and what it means for the specific kind of number a PR survey produces.

What Pew did

Pew did not ask a model to imagine “Americans”. It used what it calls a “digital twins” approach, the most demanding version of the method. For each real member of its American Trends Panel who had answered both a 2026 survey and Pew's 2025 political typology survey, the model received that person's self-reported demographics and 71 further variables from the typology survey. A second model, GPT-5.1, wrote an “expert reflection” on each profile to draw out implications a raw data dump might miss. The answering model then took the same survey the human had taken, in order, with the same instructions and programming.

Three waves were replicated: January 2026 (politics, Venezuela, Greenland, immigration enforcement, data centres), March 2026 (international affairs and knowledge questions) and April 2026 (presidential approval, national problems, Iran, sleep). The main model was Anthropic's Claude Opus 4.6, chosen because it performed best of several tested; for one wave Pew also ran OpenAI's GPT-5.1. The January replication covered 6,700 panelists rather than the full 8,512, Pew says, “due to the cost of running Opus 4.6”.

Because each synthetic respondent shadows a real one, Pew could compare not just toplines but subgroups, person-built from the same profiles. That is a more demanding design than generating respondents from demographic quotas alone, and it gives the model far more to work with than a typical brief would. If the method struggles here, there is no obvious reason to expect a thinner version to do better.

How far off the answers were

Pew's error measure is simple: for each question, the average absolute gap in percentage points between the human and synthetic shares choosing each answer option. Across nearly 300 questions, that gap averaged 12 points. By wave it was 11.3, 14.7 and 11.2.

Averages hide the tails. The gap exceeded 15 points on about 28 per cent of questions, and Pew says it was “not uncommon to see differences exceeding 20, 30 or even 40 percentage points on individual answers”.

No subgroup was well served. Pew says no group had an average error below 12 points; its own table puts the lowest, college graduates, at 11.9, and several groups at around 15 or more.

The consensus problem

The error is not random noise. It has a direction, and the direction is the worst possible one for a PR survey: synthetic samples make opinions look more uniform than they are.

Pew's examples are vivid. Fewer than half of real Hispanic adults, 42 per cent, said they were at least somewhat likely to follow the World Cup; 97 per cent of the synthetic Hispanic respondents did. The synthetic sample said 72 per cent of Hispanic adults think it extremely important for schools to teach in Spanish, more than double the real 32 per cent. Pew calls these stereotyping. Along party lines the model did the same thing, taking views held by many Democrats or many Republicans and making them close to universal: 86 per cent of synthetic Democrats called billionaires bad for the country against 45 per cent of real ones; at least 90 per cent of synthetic Republicans viewed Israel favourably against 58 per cent of real ones.

Outside politics the pattern held. The synthetic sample said 98 per cent of Americans think China benefits more than the US from their trade, 56 points above the real share, and erased the 10 per cent who say the US benefits more, the 24 per cent who say both equally and the 19 per cent who are unsure. It said 88 per cent think India has been getting stronger; the real figure is 30 per cent.

The structural measure is the cleanest. In Pew's synthetic poll, 47 per cent of questions had at least one answer option that not a single synthetic respondent chose. Real populations almost always contain someone in every box. Synthetic ones empty the minority positions out.

Too sure, and too knowledgeable

Two further distortions matter for any survey that reports awareness or knowledge, which many PR surveys do.

Synthetic respondents rarely admit uncertainty. Pew told the model explicitly that it was realistic for its personas not to know things. Even so, across opinion questions offering an explicit “not sure”, the typical human chose it roughly four times as often as the typical synthetic respondent.

And synthetic respondents know too much. On 13 factual questions about subjects such as the Constitution and NATO, real panelists answered correctly about half the time on average, and no question was answered correctly by more than three-quarters of them. The synthetic panel was right about 80 per cent of the time, and on six questions it estimated that 98 per cent or more of the public knew the answer. When a persona did not know, it almost always said “not sure” rather than guessing wrong; real people guess wrong constantly.

A PR survey that claims “only 23 per cent of consumers understand X” is making a knowledge claim. A synthetic sample will systematically overstate that kind of number, and understate the ignorance that usually makes the story.

Frozen in time

Models have a knowledge cutoff; Pew notes that Opus 4.6's was May 2025. The synthetic sample behaved as if the world had stopped there.

Real approval of the president, measured seven times by Pew between early 2025 and May 2026, fell from 47 per cent in February 2025 to 37 per cent in January 2026 and 34 per cent in April 2026. The synthetic sample put it at 46 per cent in January and again at 46 per cent in April — close to where real approval last stood in early 2025, and unmoved by anything since. Among Republicans the miss reached 19 points.

New issues were worse. In January 2026, 25 per cent of real adults had heard a lot about data centres and about a quarter had heard nothing; the synthetic sample put both at 3 per cent and lumped 94 per cent into “a little”. At least 97 per cent of synthetic respondents said data centres are mostly bad for energy costs, the environment and nearby quality of life, against real shares of 41 to 53 per cent among those who had heard of them. On electricity prices the model went the other way, estimating that 23 per cent of Americans were very concerned when the real figure was 52 per cent.

PR surveys are usually commissioned precisely because something is new: a technology, a regulation, a market shift. That is the territory where a synthetic sample has the least to go on.

Change the model, change the public

A survey should give similar answers if fielded twice the same way. Pew ran its January survey through both GPT-5.1 and Claude Opus 4.6, with everything else identical, and got two different Americas.

Opus averaged 11.4 points of error, GPT 13.3. But the more important finding is that they were wrong in different ways. GPT said 100 per cent of Americans were dissatisfied with the direction of the country; the real figure was 69 per cent, and Opus said 70. On whether there are clear solutions to most big national problems, GPT said 57 per cent, close to the real 56, while Opus said 25. On abortion, real adults split about 60 to 39 in favour of legality; Opus reproduced that split but said 0 per cent think it should be illegal in all cases, against a real 11; GPT showed a country slightly more opposed than supportive.

Pew traced a consistent temperament to each model. GPT favoured the extreme ends of scales, “always” and “never”; Opus favoured the middle, “somewhat” and “neither”. Real people chose a middle option 45 per cent of the time on ordered questions. Pew's conclusion is that “the choice of model alone is enough to yield large, systematic differences between synthetic samples that were otherwise built in exactly the same way”.

For a commissioned survey that means the vendor's choice of model, which a client will rarely be told, can decide the headline.

What it means for a survey-led pitch

Pew's test was about public opinion on matters of broad importance, and Pew is careful to say it “does not address other uses of AI models in polling”, such as drafting questions. Within that scope, the findings line up badly with how PR surveys are used.

A survey-led pitch rests on one or two striking percentages. Pew's results say a synthetic sample pushes percentages toward the extremes of agreement, removes minority and “not sure” answers, overstates knowledge, and misses anything that changed after the model's cutoff. Every one of those errors makes a number more striking. A synthetic panel is, in effect, a machine for producing the kind of finding that gets a pitch opened and does not survive scrutiny.

The scrutiny is real. Serious outlets ask where a number came from — how verification works is covered in how fact-checkers work — and a methodology line reading “synthetic respondents” is unlikely to survive that question. A methodology line that hides it is worse: it is a fabricated statistic with your company's name on it, and statistics travel. They get repeated in other coverage and, increasingly, in AI answers that cite that coverage, which is the subject of do people trust what AI says about companies.

None of this rules out cheaper research. A small real sample reported honestly, with its size and method stated, is more useful to a journalist than a large synthetic one. So is a company's own operational data, which no one else has. The thing a synthetic sample cannot provide is the one thing a survey exists to provide: evidence about what people actually think.

Methodology and limitations

This report draws entirely on Pew Research Center's study published on 30 September 2026: the overview “Silicon Samples and Synthetic Surveys: Can AI Stand In for Human Respondents?”, five companion reports on replication accuracy, model choice, certainty and knowledge, diversity of opinion and current events, and the methodology note. All were read on 1 October 2026. Figures are Pew's; where we derived a number, such as the real share of adults saying China benefits more from trade (from Pew's statement that the synthetic figure was 56 points higher) or the Republican favourable figure on Israel (by adding Pew's “very” and “somewhat” columns), the chart notes say so. The two illustrative charts use examples Pew chose to show the pattern, and should not be read as a sample of all questions; the representative measure is the 12-point average error. Four limitations apply. Pew tested public-opinion questions of broad importance among US adults; it did not test consumer or B2B audiences, which is where most PR surveys sit, and the method could perform better or worse there. Pew tested two models at one point in time; newer models may produce smaller errors, as Pew itself notes. Pew's digital twins used rich individual profiles; a panel built with less information about each respondent was not tested, and we do not know how it would compare. And we have not tested any synthetic-panel product ourselves.

Conclusion

Pew gave AI models every advantage a synthetic survey can have — a real person's profile behind every respondent, 71 typology variables, an expert reflection, the identical questionnaire — and the results still missed real Americans by 12 points on average, by more than 15 on over a quarter of questions, and by 40 on some answers. The misses had a shape: majorities became near-unanimous, minority answers and “not sure” disappeared, knowledge was overstated, opinion was frozen at the model's cutoff, and the choice of model changed the picture of the public. Those are the properties of a striking statistic, which is what a survey-led pitch is built on. A PR survey can use AI to draft questions or test wording, but on this evidence it cannot use AI respondents and report the result as what people think. Pew, whose business is measuring opinion, says it has no plans to use them at all.

Frequently Asked Questions

Sources

1. survey research Pew Research Center, Silicon Samples and Synthetic Surveys: Can AI Stand In for Human Respondents?, 30 September 2026. Overview: 12-point average error across nearly 300 questions; 28 per cent above 15 points. Read 1 October 2026.
2. survey research Pew Research Center, AI survey samples poorly replicate human public opinion, 30 September 2026. Wave and subgroup errors; World Cup, Spanish instruction, billionaires, Israel examples. Read 1 October 2026.
3. survey research Pew Research Center, AI polling results change based on the model used, 30 September 2026. GPT-5.1 against Claude Opus 4.6. Read 1 October 2026.
4. survey research Pew Research Center, How AI survey respondents express certainty and factual knowledge, 30 September 2026. “Not sure” and knowledge-question results. Read 1 October 2026.
5. survey research Pew Research Center, AI surveys fail to capture the diversity of public opinion, 30 September 2026. China trade, India, zero-selection options, sleep. Read 1 October 2026.
6. survey research Pew Research Center, AI surveys struggle with questions about current events, 30 September 2026. Presidential approval, data centres, electricity prices, knowledge cutoff. Read 1 October 2026.
7. survey research Pew Research Center, Methodology, 30 September 2026. Digital-twin construction, 71 typology variables, expert reflections, the 6,700-panelist subsample and its cost rationale. Read 1 October 2026.