How to stop AI interview fraud: start with an interview candidates trust
TL;DR
- Fewer than 1% of interviews on Sapia.ai are flagged as written entirely by AI.
- The best defence against AI interview fraud is an interview candidates trust. Across 10 million+ candidates, satisfaction averages 9.0/10.
- Version 3 of Sapia.ai’s AI-content detector catches nearly 97% of AI-written answers from the latest LLMs, up from 92%.
- Detection reads the answer alone. There are no webcams, eye-tracking or lockdown browsers, and nobody is rejected automatically.
AI interview fraud is on every TA leader’s agenda
Generative AI makes it easy to paste an interview question into a chatbot and paste the answer straight back. Candidate fraud has shot up the agenda as a result.
The industry’s reflex has been to watch harder. Webcams. Eye-tracking. Lockdown browsers. Hidden background checks.
Every one of those treats the applicant as a suspect, in what is often their first real experience of your brand. Our data points to a different answer.
Fewer than 1% of interviews are flagged as AI-written
Fewer than 1% of interviews on Sapia.ai’s platform are flagged as written entirely by AI.
We think the reason is simple. Candidates who trust the experience answer in their own words.
How candidate trust reduces AI cheating in interviews
Sapia.ai’s structured, text-based interview is built to give people room to be themselves.
- Untimed. There’s no clock pushing anyone toward a shortcut.
- Mostly completed on a phone, whenever and wherever suits the candidate.
- Open about AI from the first screen. Every candidate sees a note on AI use and commits to answering in their own words.
- Personalised feedback for every candidate at the end.
Candidates notice. Across more than 10 million candidates, satisfaction averages 9.0/10, and 86% choose to leave a comment about their experience.
There’s a practical reason too. Sapia.ai scores the competencies candidates demonstrate through the behaviours and experiences they describe. Fluency earns nothing extra, so polishing an answer with AI gives no real advantage. What matters is that the story is the candidate’s own.
How we detect AI-generated interview answers
A low number only means something if you can trust the measurement. That’s why we keep improving our AI-generated content (AGC) detector.
Version 3 was trained on answers from current LLMs by providers including OpenAI, Google, Anthropic, xAI, DeepSeek, Moonshot AI and Alibaba, each paired with genuine candidate answers written before ChatGPT was released. Tested side by side on the same answers:
| Detector | AI-written answers detected (latest LLMs) |
|---|---|
| Previous version | 92% |
| v3 | Nearly 97% |
New models arrive every few months, so we test the detector against each new release and retrain it as the landscape shifts.
A nudge back to their own words
Detection reads the answer itself. No webcams, no eye-tracking, no lockdown browsers.
If three or more answers in an interview are flagged as AI-written, the candidate gets a real-time nudge. A genuine answer is wrongly flagged around 2% of the time, so in a five-question interview the chance of an honest candidate being nudged is roughly 1 in 13,000. False-flag rates are similar across genders and ethnic groups.
Nobody is rejected automatically. AI never makes the final hiring decision, and later stages of the process, led by people, can test what candidates describe in the first interview.
“Everyone is talking about candidate fraud, and the reflex is to watch people harder, at the expense of the candidate experience. Our data points elsewhere. When candidates trust the experience, almost all of them answer in their own words.” — Barb Hyman, CEO and Founder, Sapia.ai
Three questions to ask about your own approach
- Would your candidates describe your interview as fair?
- Could you explain how AI use is detected, to a candidate as easily as to your board?
- What happens to a candidate when a flag is wrong?
If any of those are hard to answer, see how Sapia.ai’s interview works.
FAQ
What is AI interview fraud?
AI interview fraud is when a candidate submits answers written by a generative AI tool, such as ChatGPT, instead of their own. It hides how the candidate actually thinks, communicates and behaves, which is exactly what an interview is meant to show.
How common is AI cheating in job interviews?
On Sapia.ai’s platform, fewer than 1% of interviews are flagged as written entirely by AI. Sapia.ai’s view is that an interview candidates trust is the strongest defence against AI cheating.
Can AI detectors tell if interview answers were written by ChatGPT or other LLMs?
Yes, with the right training data. Version 3 of Sapia.ai’s AGC detector catches nearly 97% of AI-written answers from the latest LLMs, tested against models from OpenAI, Google, Anthropic, xAI, DeepSeek, Moonshot AI and Alibaba.
Are candidates rejected if AI-generated content is detected?
No. If three or more answers in an interview are flagged, the candidate gets a real-time nudge to answer in their own words. Nobody is rejected automatically, and AI never makes the final hiring decision.
Does Sapia.ai use webcams or eye-tracking to stop candidate fraud?
No. Detection reads the answer itself. Sapia.ai uses no webcam monitoring, eye-tracking or lockdown browsers.
Is AI-content detection fair to honest candidates?
A genuine answer is wrongly flagged around 2% of the time. In a five-question interview, the chance of an honest candidate being nudged is roughly 1 in 13,000, and false-flag rates are similar across genders and ethnic groups.
Does using AI to polish an answer help a candidate score higher?
No. Sapia.ai scores the competencies candidates demonstrate through the behaviours and experiences they describe, so fluency earns nothing extra.
How does Sapia.ai keep its detector up to date?
New LLMs arrive every few months. Sapia.ai tests the detector against new releases as they appear and retrains it as the landscape shifts.
About Author
