Answers
AI polling accuracy
AI polling is only as good as the published miss. The calibrated 50-state benchmark: 7.07-point average miss on the best 80% of 1,521 scored questions (1,216 retained). 10.02 points across every scored question. MAE is an average miss in percentage points against published poll toplines. A narrower April test on 5 places with 10,000 respondents scored 7.47 points on 404 non-electoral questions. Electoral-margin questions are a separate table on the methodology page.
The number to quote
The calibrated 50-state benchmark: 7.07-point average miss on the best 80% of 1,521 scored questions (1,216 retained). 10.02 points across every scored question. The best 80% keeps the 1,216 lowest-error questions. The other 305 questions stay in the 10.02-point average. Every study is calibrated against published survey data before it reaches you. Each question on that file uses 200 simulated respondents. The file is state-benchmark-per-question.csv.
A narrower April test on 5 places with 10,000 respondents scored 7.47 points on 404 non-electoral questions. State and category tables, including the older April bank, are on methodology. Coverage (650,000 census-grounded respondents · all 50 states and D.C. · 16 occupation panels.) is on coverage.
Pew Research Center reported on Sept 30, 2026 that AI respondents "differed from their human counterparts by an average of 12 percentage points" across nearly 300 questions, after building a twin per panelist from demographics and political typology answers, on a question set that is not Lewsearch's 50-state file.
The Lewsearch Report grades forecasts that were frozen before the official number: 3 wins, 1 miss (Issue 06), and 3 awaiting a grade (Issues 03, 04, and 07). Issue 05 is a win on Guess #2.
Worked example: published before the hearing
On July 6, 2026, Lewsearch fielded a 600-respondent, census-based read on the Prince George's County data-center pause and posted it before the July 7 Council vote. The council passed the moratorium. The artifact is a public-opinion study, not a probability sample of recorded votes. It is a timestamped directional read. The report and JSON sit at /reports/prince-georges-data-centers.
A separate check against Polymarket is written up on /proof. Quote the 50-state miss for poll toplines.
| Read | n | Result |
|---|---|---|
| Calibrated 50-state, best 80% | 1,216 of 1,521 questions | 7.07 points |
| Calibrated 50-state, full set | 1,521 scored questions | 10.02 points |
| Separate held-out set, non-electoral | 14 scored of 22 drawn | 9.97 points |
| Electoral-margin questions | See methodology | Harder than the 50-state headline. Category tables stay on the methodology page. |
FAQ
- How accurate is AI polling?
- AI polling is only as good as the published miss. The calibrated 50-state benchmark: 7.07-point average miss on the best 80% of 1,521 scored questions (1,216 retained). 10.02 points across every scored question. MAE is an average miss in percentage points against published poll toplines. A narrower April test on 5 places with 10,000 respondents scored 7.47 points on 404 non-electoral questions. Electoral-margin questions are a separate table on the methodology page.
- Can an AI poll call an election?
- Electoral-margin questions are harder than the 50-state headline, and Lewsearch does not use them as that headline. Do not treat a synthetic panel as a substitute for a probability sample or a certified canvass. Category tables are on /methodology.
- Did Lewsearch publish anything before a real vote?
- Yes. The Prince George's County data-center moratorium read was fielded July 6, 2026, and published before the July 7 Council vote. The council passed the pause. That is a public-opinion read, not a poll of recorded votes. The report is at /reports/prince-georges-data-centers.
- How did the panel do against Polymarket?
- A separate check against Polymarket is written up on /proof. The 50-state miss above is the figure to quote for poll toplines.
Check the work
Related answers
Canonical: https://lewsearch.com/answers/ai-polling-accuracy