Can synthetic personas predict what the country thinks? They passed the test.

The same questions Pew Research Center asked the nation, word for word, answered by our panel of 2,500 synthetic Americans with no internet and no way to peek. The AI behind them finished learning before the first interview happened. Then we graded the panel against the real results.

Graded against real national polls
1.8 ptsgap on a question that had never been asked before
13 / 16top answer matched America's, across the full test
0.96rank match on America's 7 cost worries
16,000blind answers, zero internet access

Scored against the gold standard of American opinion research

Pew Research Center
How the test worked

Three steps, sealed in order.

The order is the proof. The AI's knowledge ends before the interviews begin, and the real answers are locked away before our panel is asked a single question.

STEP 01

Pew asks America

13,300 real Americans answer two national surveys on the economy and social media.

spring 2026
STEP 02

The answers are sealed

Pew's published results are recorded and timestamped before the test begins, so nothing can be adjusted after the fact.

before the test
STEP 03

Our panel answers blind

16,000 responses from a census-built panel. The AI's knowledge ends before the interviews, and it has no internet.

then we score it
first time ever asked

"Would you support or oppose a ban that prevents anyone under the age of 16 from using social media sites?"

Real America · Pew, n=9,75056%
Synthetic panel · blind54.2%

Share answering "support" · gap 1.8 points· Pew wrote this question three months after the AI's training ended. There was nothing to memorize.

The scoreboard

Real America next to the synthetic panel, question by question.

Dark bars are Pew's real America; orange bars are the synthetic panel. Every survey links to Pew's own report, topline and questionnaire, so any number can be checked at the source.

Economy and cost of living

a national Pew survey · 3,554 Americans interviewed · July 2026Pew reporttopline PDFquestionnairemethodology

How would you rate economic conditions today?

matched
  • Excellent2.0 · 7.9
  • Good21.2 · 28.8
  • Only fair41.4 · 36.7
  • Poor35.4 · 26.6

A year from now, the economy will be...

2 pt gap
  • Better23.2 · 22.5
  • Worse36.4 · 39.0
  • Same40.4 · 38.5

A year from now, your family's finances will be...

matched
  • Better34.3 · 31.4
  • Worse15.2 · 21.1
  • Same50.5 · 47.6

Concern about people finding jobs

matched
  • Very44.4 · 40.5
  • Somewhat32.3 · 38.8
  • Not too18.2 · 15.0
  • Not at all5.1 · 5.7

Concern about how the stock market is doing

matched
  • Very18.2 · 21.5
  • Somewhat34.3 · 39.2
  • Not too32.3 · 24.8
  • Not at all15.2 · 14.5

Social media rules for kids

a national Pew survey · 9,750 Americans interviewed · May to June 2026Pew reporttopline PDFquestionnairemethodology

Ban under-16s from social media?first time ever asked

matched
  • Support56.0 · 54.2
  • Oppose21.0 · 29.3
  • Not sure23.0 · 16.5

Set limits on minors' time on social media?

matched
  • Support78.0 · 70.8
  • Oppose10.0 · 15.2
  • Not sure12.0 · 13.9

Numbers are percentages. Pew value first, our panel's in orange. The scores in the hero are measured across the full 16-question test.

Kept honest

Checkable without trusting us.

Every question was asked word for word as Pew wrote it. The real answers were sealed before the test began, the AI's knowledge ends in February 2026, before Pew's interviews, and the panel had no internet access while it answered. Every Pew number on this page can be checked against the linked report, topline and questionnaire PDFs above.

Pew Research Center is the source of the comparison data and is not affiliated with, and has not endorsed, this analysis.

2,500synthetic Americans, built from US Census records
16,000blind answers, one per persona per question
0internet access while the panel answered
0questions changed or reworded, asked word for word