Research

The Imposter Study

Study in progress · methodology published, results pending

Online research recruitment has a data-quality problem that predates and now compounds with generative AI: a meaningful share of respondents in some online studies are fraudulent, duplicate, or bot participants. They pass attention checks, they clear per-session bot detectors, and they are cheap to mint at scale. The Imposter Study measures one specific thing: how many of them a continuous-human check removes at intake, before they ever enter a panel.

Researchers studying online-panel data quality have reported fraudulent or duplicate participation as a serious threat, with the affected share varying widely by platform, incentive, and recruitment channel. We use that published range as context, not as our own result.

Online-panel fraud and data-quality literature. We cite the range; we do not claim it as a measured outcome of our method.

The question we are measuring

Not "can we detect a bot in a single session," which is the race every door-check loses over time. The question is: of the respondents who would have entered a study, how many lack any consented, continuous human record behind them, and does removing them at intake improve the downstream data quality a research team already measures.

Method

  1. Intake check. At recruitment, each prospective participant is checked against Pupul for the four claims: is a human, is continuous, has deviated, stands behind. No survey content is shared with us; the participant consents and can revoke.
  2. Two arms. The study team runs their normal intake alongside the continuous-human check, so every respondent is scored both ways without changing the study itself.
  3. Outcome measures. We compare the team's own existing data-quality signals across the two arms: duplicate rate, failed-attention rate, straight-lining, impossible-metadata rate, and any fraud flags the team already uses.
  4. Honest nulls. A participant with no evidenced record returns null, never false. Null is reported as null; the study team decides how to treat it. An absent record is not an accusation.
  5. Verifiability. The continuous-human claims rest on a public, append-only transparency log, so the "continuous" part of the claim is checkable rather than asserted.

What we will report

When a cohort completes, this page will publish: the sample size, the two-arm comparison on each pre-registered outcome measure, the null rate, and the limitations. We will report what the data shows, including if the effect is small. We will not publish a number before we have measured it.

Results

Pending. The first pilot cohorts are being scheduled. If you run online studies and want to be a measured cohort, the pilot is below.

Be a measured cohort

We run short paid pilots with research teams: your intake, your panel, our check, scored against your own duplicate and fraud rate. You keep the data-quality lift whether or not you continue, and your cohort's aggregate result becomes part of this study (never your participants' content).

Run a pilot Why elapsed time cannot be faked

Pupul is a proof-of-human signal, not a consumer report, and the record is owned by the person it describes. This page states an open methodology; no result is claimed until it is measured.