Pharma · · 8 min read
HCP survey fraud: how to verify physician respondents when AI can pass the screener
Acumen Editorial · Updated November 21, 2026
Pharma insights teams have always worried about professional respondents and people who bluff their way into an oncology survey. That problem has changed shape. Respondents can now be software, written to pass the exact checks the industry uses to screen them.
This matters twice for pharma. Bad respondents distort the study in front of you. And if that data later trains or calibrates a synthetic HCP persona, the error gets copied into every simulated answer after it.
How big is the HCP survey fraud problem?
Large, and worst wherever a survey link can travel freely.
NORC at the University of Chicago reviewed the research on fraudulent respondents and bots in April 2026. It reports estimated fraud rates of 15 to 30 percent across market research, and up to 45 percent on some platforms. In one health study it cites, 94.5 percent of responses recruited through social media were fraudulent.
Healthcare professionals are not a safe niche. In a University of Wisconsin–Madison study published in JMIR Formative Research on September 2, 2026, researchers surveyed US clinicians who had worked with patients with limited English proficiency. They recruited through professional newsletters and healthcare groups on Facebook and WhatsApp, and offered a $15 gift card. Of 5,886 entries, 40 were verified as legitimate.
A JMIR news feature from September 21, 2026 describes a long COVID study that drew 3,638 responses within 24 hours of launch. About 9 percent came from genuine participants.
HCP studies pay real incentives. The money that makes HCP research possible is also what makes it a target.
Can AI bots really pass survey quality checks?
Yes. Sean Westwood of Dartmouth built an autonomous synthetic respondent, described in PNAS in November 2025. Across 43,000 tests it passed 99.8 percent of attention checks and made no errors on logic puzzles. It cost about five cents per response, against about $1.50 for a typical human respondent. Every AI detection method Westwood tested failed to identify it. His summary: "We can no longer trust that survey responses are coming from real people."
NORC's review reaches the same verdict on the usual defenses. Vendor fraud flags fail against sophisticated attacks. Domain-knowledge checks and open-ended questions have become ineffective against LLM-powered bots. Evidence that AI text detectors work is lacking. Some fraudsters start a survey by hand and switch to a bot partway through.
The quiz question that once caught a fake cardiologist is now the easiest part for a model. An LLM knows the first-line therapy. What it cannot fake is a license, a practice address and a verifiable identity.
EPHMRA, the European pharma market research association, ran a session in June 2026 on detecting and deterring AI-assisted fraud. It named "increasingly sophisticated synthetic respondents" as a risk. The industry has noticed. Most study specs have not caught up.
Why does no single check work?
Because each one catches a different kind of fraud, and attackers adapt to whichever check is most common.
The Wisconsin team is clear on this. No single indicator was sufficient. Their strongest signal, a pattern of submissions arriving in regular waves, would still have missed nearly a quarter of the fraud. Their honeypot field caught zero entries.
| Check | What it catches | Where it fails |
|---|---|---|
| Attention checks and logic traps | Careless humans, simple bots | LLM agents pass them almost every time |
| Honeypot fields | Basic form-filling bots | Caught nothing in the Wisconsin HCP study |
| Vendor fraud scores, IP and VPN flags | Duplicates, offshore farms, proxies | NORC: fail against sophisticated attacks |
| Clinical knowledge questions | People bluffing a specialty | LLMs answer them correctly |
| AI text detectors | Some machine-written open ends | NORC: evidence of efficacy is lacking |
| License and identity verification | Non-clinicians, invented identities | Needs record matching and a contact step |
| Trained human review of open ends | Generic, homogeneous or inconsistent answers | Costly and slow, so it must be targeted |
The pattern is plain. Checks on what a respondent says are failing. Checks on who the respondent is still hold. As the JMIR feature puts it: "It's no longer what you say on the questionnaire but rather how it's said."
What should an HCP respondent verification checklist include?
Use this before fielding any physician or HCP study, and again before any data is used to build or calibrate a model.
- Recruit from closed, known channels; never post an open survey link on social media
- Match each respondent to a licensed practitioner record (NPI registry or state licensing board) before the survey starts
- Confirm specialty, practice setting and patient volume against independent sources, not only self-report
- Issue single-use, individual survey links tied to a verified identity
- Collect paradata (timing, submission patterns, device and IP signals) and write exclusion rules before fielding
- Flag batched or wave-pattern submissions, duplicate IPs and disposable email domains
- Have trained reviewers read open ends for generic, uniform or contradictory answers
- Recontact a sample of completes to confirm participation
- Hold incentives until verification is complete
- Report verification results by source, so you know which channels to stop using
NORC also notes that aggressive screening can exclude hard-to-reach clinicians. Verification should check identity, not filter out rural, international or less digitally active doctors.
What does fraud mean for synthetic HCP personas?
It is the bigger long-term risk. Many teams now calibrate synthetic personas and simulated panels on fielded HCP data. If a quarter of that data came from bots or impostors, the persona learns their answers.
NORC points out that LLM-generated responses tend to be more homogeneous than human ones. Contaminated data makes a persona sound more confident and more uniform than real physicians are, exactly the error that makes simulated research look better than it is.
So the verification standard for persona training data should be higher than for a one-off tracker, not lower. Every record should trace to a verified practitioner, with consent for model use, collected inside a documented process. We covered the wider vendor questions in our synthetic persona vendor checklist.
What should you ask an HCP panel or data vendor?
- How do you verify each respondent is a licensed practitioner, and against which registries?
- Is verification done before the survey or only after completes come in?
- Which recruitment channels feed this sample, and are any of them open links?
- What share of completes did you remove for quality on comparable studies, and why?
- How do you detect AI-agent respondents, beyond attention checks?
- Can you show verification records for a sample of respondents, with identities protected?
- If this data trains or calibrates a model, can you prove provenance and consent for every record?
A vendor that answers with its attention-check pass rate has answered a 2020 question.
How Acumen helps
We supply verified physician, HCP and patient data for pharma market research and AI personas. Every contributor is identity- and license-verified before they work, recruited through closed channels and matched to the specialty and setting the study needs. Trained reviewers check answers, and every record carries provenance and consent for model use.
See our pharma market research data for AI service.
FAQ
What is HCP survey fraud?
HCP survey fraud is any response to a healthcare professional study that does not come from a genuine, qualified practitioner. It includes people posing as physicians, duplicate entries from one person, offshore survey farms and automated bots, including AI agents. Fraud distorts study results and, if the data is reused, contaminates any model or synthetic persona trained on it.
How common are fraudulent respondents in healthcare research?
Common, especially with open recruitment. NORC's April 2026 review estimates fraud at 15 to 30 percent across market research, and up to 45 percent on some platforms. A University of Wisconsin–Madison HCP survey published in September 2026 verified only 40 of 5,886 entries as legitimate after recruiting through newsletters and social media groups.
Can AI bots pass survey attention checks?
Yes. Dartmouth's Sean Westwood built an autonomous AI respondent, described in PNAS in November 2025, that passed 99.8 percent of attention checks across 43,000 tests and made no errors on logic puzzles. It cost about five cents per response, and every AI detection method he tested failed to identify it.
How do you verify a physician survey respondent?
Match the respondent to a licensed practitioner record, such as the NPI registry or a state licensing board, before the survey starts. Confirm specialty and practice setting independently, send single-use links tied to that identity, and recontact a sample of completes. Then layer technical checks and trained review of open ends on top.
Do honeypot questions and fraud scores still work?
Only partly. They catch simple bots and obvious duplicates. In the Wisconsin HCP study, the honeypot field caught zero entries, and the strongest single indicator would have missed nearly a quarter of the fraud. NORC reports that vendor fraud flags fail against sophisticated attacks. Use them as one layer, never as the whole defense.
Why does survey fraud matter for synthetic HCP personas?
Personas and simulated panels are often calibrated on fielded HCP data. If that data includes bots or impostors, the persona reproduces their answers. NORC notes that LLM-generated responses tend to be more homogeneous than human ones, so contaminated data can make a persona look more consistent than real physicians are. Training data needs verified provenance.