What Prolific’s “100% Human” Research Guarantee Covers

Prolific’s “100% human” guarantee has a specific reimbursement rule. Its help article dated August 4, 2026 says submissions identified as non-human through its Bot authenticity checks qualify for double the submission costs back as Prolific credits. The article also describes protections including ID verification, more than 110 pre-screening checks and real-time bot detection. That gives researchers a concrete remedy tied to a named check, with credits rather than an unconditional cash refund.

I think this is a useful guarantee with a narrower job than its headline might suggest: it covers a defined authenticity failure, while eligibility and the usefulness of an answer still need separate checks.

If you’re a founder buying responses for a customer study, that boundary belongs in your planning. You can receive an answer from a real person who has never bought your category. Or someone who qualifies but rushes through a question. Neither case necessarily falls under a guarantee about non-human submissions.

What the detection figures tell you

In its February 4 announcement about Authenticity checks, Prolific separately reports 98.7% precision for detecting AI-generated free text and 100% accuracy for AI-agent detection in its testing.

I like that these claims name what was being detected. Generated text and an agent completing a study are different targets. A human can enter generated text, so detecting one does not automatically establish the other.

The metrics differ too. Precision describes how often a positive detection was correct. The reported 98.7% precision does not mean the detector caught 98.7% of all generated text. To understand how much misuse went undetected, you would need evidence about missed cases as well.

The reported 100% agent-detection accuracy describes Prolific’s testing. It does not establish that every response in every live study will be authentic. Real deployments can include behaviour and conditions a test did not cover.

I’d give Prolific credit for publishing specific performance claims and attaching a remedy to detected failures. I would push back on reading either test result as proof that response review is now unnecessary. The announcement’s metrics and the help article’s reimbursement rule need to be read separately; neither should silently expand the other.

A real person can still be the wrong participant

Suppose you sell refillable cleaning products and want to understand why first-time buyers don’t reorder. You recruit people who buy household cleaning products.

Those people may all be human. They may answer carefully. But if they haven’t bought your refill product, their answers cannot tell you why your first-time buyers stopped. You have recruited a plausible audience for a different question.

That example is hypothetical, but it’s the kind of mistake I’d worry about before debating a detector’s last decimal place. An authenticity guarantee cannot repair a mismatch between the decision you need to make and the people you asked.

Even with the right participants, your survey can make weak answers likely. Ask, “How much did the inconvenient refill process affect your decision?” and you have supplied both a complaint and a possible explanation. A careful human might follow your framing. We covered ways to avoid that in How to Write Survey Questions That Don’t Lead People.

I’d rather ask whether they reordered, then invite them to describe what happened when they next needed cleaning supplies. Their account might support your theory. It might give you a reason to abandon it.

A checklist for evaluating a response guarantee

These are the questions I would use when comparing providers. This is editorial guidance, not a statement of additional Prolific terms.

What triggers reimbursement?

Write down the exact event that makes a submission eligible. Does a provider’s detector have to flag it? Can your own evidence trigger a review? Are particular checks required to be enabled before the study starts?

For the Prolific guarantee described above, the stated trigger is a submission identified as non-human through Bot authenticity checks. I would not assume that an irrelevant answer, failed screener or suspected use of generated text qualifies under that same rule.

Ask about any reporting deadline and what evidence you need to retain. If the published terms don’t answer those questions, get clarification before relying on the remedy.

What do you receive?

Record the form of reimbursement and the costs it covers. Prolific specifies double the submission costs in Prolific credits. That should not become “double your study budget back” in an internal purchasing note.

Credits may be useful if you plan to run another study with the provider. They still leave you with the work of reviewing affected responses and deciding whether to rerun part of the research. Ask whether any credit restrictions or expiry terms apply rather than assuming either way.

What does detection cover?

Separate checks for bot participation from checks for generated free text. Ask which response types are assessed and when checks run. Where a provider publishes a performance figure, keep the target and metric attached to it.

Then look for the unanswered questions. What evidence is available about missed cases? How were the tests conducted? How closely do their conditions resemble your study?

You may not get every detail, particularly where disclosure could help people evade detection. You can still record what is known and what remains uncertain. An unanswered question should stay unanswered in your assessment.

How will questionable responses be reviewed?

Before launch, document what makes someone eligible and what would prompt a closer look at their answers. Decide who will review flagged submissions and where they will record the reason for keeping or excluding each one.

Use flags to start a review. A short answer can be perfectly adequate. A participant can misunderstand a question because you wrote it badly. I don’t want a process that treats polished prose as evidence of authenticity or an unexpected opinion as evidence of fraud.

Keep the original response alongside the review decision. Where the platform allows it, use its review or dispute process. If a finding depends on a handful of questionable submissions, compare the interpretation with and without them before acting.

Your next study

For a small customer study, I would choose a provider partly on whether it explains these boundaries clearly. A defined reimbursement rule is useful. So is knowing which judgments remain yours.

At Probe, participants we recruit come from our recruitment partner’s vetted panel, and you see one all-in price per study before launch. You can also send a study to your own audience for free up to 10 questions and 25 responses. Neither route removes the need to define who qualifies or review whether the answers support your decision.

Before you launch, write two short notes beside your study plan: the provider’s exact reimbursement trigger, and your own rule for reviewing questionable responses. If you can’t yet write the second, settle it before you see an answer you desperately want to believe.