Sapia.ai Updates AI Detector for Fairer Hiring Interviews

As generative AI becomes a routine part of job searching, employers face a new recruitment challenge: distinguishing candidates’ own experiences from answers generated by large language models. Hiring technology provider Sapia.ai says fewer than 1% of interviews on its platform are flagged as entirely AI-written. The company has now released version 3 of its AI-generated content detector, which it says identifies nearly 97% of AI-written answers produced by the latest large language models. Its approach puts the emphasis on candidate transparency and human-led hiring decisions rather than automated rejection or intrusive monitoring.

Sapia.ai Updates AI Detection for Candidate Interviews

The growing availability of generative AI has complicated an important part of recruitment: assessing what candidates know, how they have handled workplace situations, and whether their interview responses genuinely reflect their own experiences.

Applicants can use AI tools to improve grammar, organize thoughts, and prepare responses. But employers may also encounter answers generated almost entirely by chatbots, raising questions about the reliability of remote assessments and the fairness of comparing candidates.

Recruitment technology provider Sapia.ai is addressing this issue with the third version of its AI-generated content (AGC) detector. The company says the updated system identifies nearly 97% of AI-written answers produced by the latest large language models, improving on the previous version’s reported 92% detection rate.

Sapia.ai also reports that fewer than 1% of interviews on its platform are flagged as written entirely by AI. The figures suggest that fully AI-generated interviews remain relatively uncommon within its own candidate population, although they do not establish the overall prevalence of AI assistance across the wider recruitment market.

A Detector Designed to Prompt, Not Punish

Sapia.ai’s approach differs from recruitment processes built around extensive candidate surveillance.

Its structured, text-based interview is untimed and generally completed on a mobile phone. Candidates receive personalized feedback after submitting their responses. Before starting, they are informed about AI use and asked to commit to answering in their own words.

The detector analyzes the text of each answer. According to the company, it does not depend on webcam monitoring, eye tracking, or lockdown browsers.

If three or more answers in an interview are flagged as AI-written, the candidate receives a real-time prompt. The purpose is to encourage authentic responses rather than automatically disqualify applicants.

Barb Hyman, CEO and founder of Sapia.ai, argues that employers risk damaging candidate experience when concerns about fraud lead to increasingly intrusive monitoring. The company’s position is that clear expectations and a candidate-friendly assessment process can encourage applicants to provide their own answers.

This is an important distinction for talent acquisition teams. A detection alert is not proof of dishonesty, and Sapia.ai says its system does not automatically reject candidates because of a flag.

What Changes in Detector Version 3?

The latest detector was trained using AI-generated answers from models associated with OpenAI, Google, Anthropic, xAI, DeepSeek, Moonshot AI, and Alibaba.

Those responses were paired with genuine candidate answers written before ChatGPT became publicly available. Sapia.ai says it tested the updated detector against the previous version using the same answers, increasing reported detection performance from 92% to nearly 97%.

The company also says it will continue testing against newly released models and retraining the detector as generative AI capabilities change.

That maintenance requirement is significant. AI writing models evolve rapidly, and systems trained on older examples may become less effective as newer models produce different writing patterns. Recruitment technology vendors therefore face an ongoing challenge in maintaining detection performance across changing models, writing styles, languages, and candidate populations.

However, the 97% figure should be interpreted carefully. It is a company-reported detection result for the tested AI-generated answers, not a guarantee that 97% of AI-assisted candidates will be identified in live recruitment. The announcement does not provide a complete independent benchmark, test-set composition, or detailed performance breakdown by language and model.

Why Candidate Experience Is Part of Hiring Integrity

Sapia.ai’s argument is that the design of an interview can influence how candidates behave.

The company reports that its platform has been used by more than 10 million candidates, with an average satisfaction score of 9.0 out of 10. It also says 86% of candidates choose to leave a comment about their experience.

These are company-reported platform metrics, but they help explain the commercial rationale behind its approach. Recruitment assessments are often among the first direct interactions applicants have with an employer. A process that feels excessively restrictive or impersonal can undermine the employer brand before a candidate reaches a human interview.

Sapia.ai’s structured format is designed to make the experience accessible across devices and give candidates time to describe their behavior and workplace experiences.

The company says its assessment evaluates competencies demonstrated through those examples rather than rewarding polished writing. In that model, using AI to improve the fluency of a response should not necessarily create an advantage. What matters is whether the underlying example reflects the candidate’s own experience and can be explored further during subsequent interviews.

The Accuracy and Fairness Questions Employers Should Ask

AI-generated text detection is not a straightforward authorship test. Detection systems classify patterns in submitted text; they do not independently establish who wrote it or whether a candidate intended to misrepresent their abilities.

Sapia.ai reports that genuine candidate answers are wrongly flagged around 2% of the time. It says that, in a five-question interview, the probability of an honest candidate receiving enough false flags to trigger a prompt is approximately one in 13,000. The company also reports similar false-flag rates across genders and ethnic groups.

Those claims are relevant to employers evaluating the technology, but buyers should request the underlying validation evidence. The probability of a multi-answer alert depends on how errors are distributed across questions, and the company’s published announcement does not provide the full statistical methodology behind the calculation.

There are broader reasons for caution. A 2026 review published in AI and Ethics highlighted the limitations of AI-text detectors, including inconsistent classifications across tools and the risk of treating text patterns as definitive evidence of authorship. The research focuses on academic writing rather than recruitment, so its findings should not be directly transferred to Sapia.ai’s specific detector. It nevertheless reinforces the importance of independent validation before using detection scores in high-stakes decisions.

For employers, this means asking vendors about false-positive rates, language coverage, demographic testing, model updates, and the process for reviewing disputed flags. Detection should be one signal in a wider assessment process, not a substitute for evaluating job-related skills.

Human Judgment Remains Central to Recruitment

Sapia.ai says its AI does not make the final hiring decision. The initial interview assesses competencies through candidates’ described behaviors and experiences, while later stages involving people can explore those examples in greater depth.

This human-led approach is particularly important when employers are assessing candidates from different educational, cultural, and language backgrounds. Written fluency, familiarity with AI tools, and access to technology should not be confused with job competence.

The issue also connects to wider research on AI use in recruitment. A 2026 study published in the International Journal of Selection and Assessment, based on two applicant samples, found that self-reported generative AI use in pre-hire assessments increased between 2024 and 2025. It also found that warning statements reduced reported use, suggesting that communicating expectations can be a useful deterrent. <a href=”https://onlinelibrary.wiley.com/doi/10.1111/ijsa.70056″>The study</a> offers useful context for why candidate guidance may matter alongside technical detection.

For HR teams, the emerging challenge is to protect assessment integrity without discouraging legitimate AI literacy or making the hiring process unnecessarily invasive.

Sapia.ai’s v3 release reflects one possible direction: use detection to support transparency, give candidates an opportunity to respond in their own words, and retain human judgment in consequential decisions.

As generative AI becomes more deeply embedded in everyday work, employers will need to distinguish between candidates who use technology to communicate more effectively and those whose responses do not genuinely represent their experience. That distinction will require more than a detector. It will require well-designed assessments, clear policies, validated tools, and recruiters who know how to interpret the evidence.

Market Landscape

AI-assisted applications are creating demand for more reliable assessment methods, but recruitment technology providers must balance assessment integrity with accessibility, privacy, and candidate trust.