← All media signals
Governance

The Human Count

Research bodies now ask studies to state what AI did and how many real people answered. New tests find synthetic respondents that misread taste and miss real people.

What is it?

Research bodies now ask studies to say what AI did and how many real people answered. Two single studies find synthetic respondents biased toward liking, or leaving real people uncovered. How much a disclosed synthetic answer can be trusted is the open point.

Why now?

In May 2026 the American Association for Public Opinion Research’s task force called synthetic responses the most risky of the core tasks it considers, and ESOMAR advised buyers in 2025 on validation. Whether disclosure will grow into pass marks is unsettled.

Who’s building it?

AAPOR’s task force writes the disclosures, ESOMAR advises buyers, and academic teams test the synthetic respondents.

AAPOR task forceESOMARGallupPewSynthetic-data vendors
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — vendors marketing samples made entirely of synthetic responses.
Buyers — buyers of synthetic data, advised by ESOMAR to ask suppliers how they validate.
Gatekeepers — the AAPOR task force, ESOMAR and researchers testing synthetic respondents.
What’s at stake
Whether disclosure is enough: AAPOR asks for tasks, human counts and validation, and sets no pass mark.
AAPOR’s report calls synthetic responses the riskiest core task.
One study found synthetic taste data skewed toward liking.
What’s unsettled
Whether holdout testing against data held back becomes a buyer’s standard question.
Whether whole-profile comparisons become part of how synthetic respondents are judged.
Whether a stated human count changes how readers weigh a finding.
02 · How the demand matureda progression, not a timeline
Standard
An AAPOR task force report sets out required disclosures for AI in surveys.
Guidance
ESOMAR’s 2025 guidance suggests buyers ask suppliers how they validate.
Taste
A June study of 554,940 synthetic respondents found a systematic bias toward liking things.
Coverage
An October preprint found six models left real respondents uncovered on two US surveys.
03 · Where the market goesthe forks, not a prediction
→Disclosure or threshold: studies may keep stating AI tasks and human counts, or a body may later add a pass mark.
→Buyers may add holdout testing and whole-profile comparison to the questions they put to suppliers.
→Vendors of fully synthetic ‘samples’ may adopt AAPOR’s preferred term, synthetic responses, and state the human count.
→More tests may check whole answer profiles, not answers one at a time.
The path depends on whether disclosures are matched by shared tests: held-back data, whole-profile comparison, and a human count readers can see.
04 · The Compliance-as-Feature deskadjacent markets, moving now
The Synthetic Stamp
EU rules put marks on AI text.
OpenAI says it will watermark ChatGPT text in the EU.
TechCrunch
The Labelled Avatar
New York requires ads to disclose synthetic performers.
New York law fines undisclosed synthetic performers in ads $1,000, then $5,000.
New York State Assembly
The Human Seal
Books can now carry a human-authored mark.
About 6,500 titles carry the Authors Guild’s human-authored mark
AFP via The Star
05 · The Wire
01
AAPOR asks studies to state their human respondent count.
AAPOR ↗
02
Synthetic taste data skews toward liking everything.
arXiv ↗
03
Six models left real respondents uncovered in one test.
arXiv ↗
04
ESOMAR tells buyers to ask how suppliers validate.
ESOMAR ↗
The panel on this signal
They agreed on
–AAPOR’s task force asks studies to state AI’s tasks, the human count and how results were checked.
–ESOMAR’s 2025 guidance suggests buyers ask how suppliers validate.
–Synthetic responses estimate what people would say; they are not a sample of people.
The public-opinion pollster disagreed on
–A human count is not a margin of error: pollsters infer from a known sampling frame, and synthetic responses have none.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.