What Live Vetting Catches

For two decades, remote technical assessment followed a stable pattern: an online test, a take-home exercise, and a video interview. Each step assumed that the person completing it was the person being hired and that the work reflected their own ability. Neither assumption is safe any more. Real-time AI assistants can answer interview questions invisibly, frontier models solve the algorithmic puzzles that coding screens depend on, and organized schemes use stolen identities and deepfakes to place workers inside companies. This article reviews the public evidence on how assessment has been compromised, what decades of selection research say about which methods predict performance, and how a multi-gate process anchored by a live, unscripted session with a senior practitioner addresses the gaps.

The assessment problem has two parts

Hiring fraud in the AI era takes two distinct forms, and they require different defenses.

• Capability misrepresentation. A real candidate uses AI tools to perform better than they can on their own: during an online test, in a take-home, or live in an interview.
• Identity misrepresentation. The person assessed is not the person who will do the work, or the identity presented does not exist at all.

Most assessment processes were designed against neither. Both are now documented at scale.

Capability misrepresentation: what the evidence shows

Real-time assistance is a commercial product

In 2025, a Columbia University student built Interview Coder, a tool that displays AI-generated answers in a hidden window that screen-sharing software does not capture. He reported using it to obtain an Amazon internship offer; he and his co-founder were suspended by Columbia, and the product was relaunched as Cluely, which raised a $5.3 million seed round and reported surpassing $3 million in annual recurring revenue by April 2025¹. Amazon's response was to require candidates to acknowledge that they will not use unauthorized tools¹. An acknowledgment is a policy; it is not a detection method.

Interviewers did not notice

Interviewing.io ran a controlled experiment in which candidates were instructed to use ChatGPT covertly during live technical interviews conducted by experienced interviewers. Across 32 analyzed interviews, candidates passed 73% of the time on verbatim LeetCode questions and 67% on modified LeetCode questions, compared with a platform baseline of 53%. On custom questions designed by the interviewer, the pass rate fell to 25%. Not a single interviewer identified the cheating, and 72% said they were confident in their hiring decision².

In a controlled study, AI assistance raised pass rates on standard questions well above baseline, and no interviewer detected it. Custom questions were the only format where assisted candidates struggled.

The finding has two implications. Standard question banks no longer discriminate between candidates; they discriminate between candidates who use assistance and those who do not. And the format that resisted assistance was the one requiring the candidate to engage with a problem the interviewer had constructed, which is the defining feature of a live, unscripted session.

The puzzles are solved

Algorithmic screens were valuable because the problems were hard. That premise has weakened. OpenAI reported that its o3 model achieved a gold medal at the 2024 International Olympiad in Informatics and a Codeforces rating on par with elite human competitors³. Any assessment that a frontier model can complete within the time limit measures, at least in part, access to a model.

Take-home and asynchronous assessments are the most exposed

Asynchronous formats give candidates unsupervised time, which is exactly what AI assistance needs. A take-home project can be produced largely by a coding assistant; an asynchronous video answer can be read from a generated script; a written case study can be drafted by a model and lightly edited. The output may be excellent. What it no longer shows reliably is the candidate's own judgment, which is the quality the assessment was intended to measure.

Employers have noticed. Google's CEO Sundar Pichai said the company would introduce at least one round of in-person interviews "just to make sure the fundamentals are there," and other large employers, including Cisco and McKinsey, have been reported to be reinstating face-to-face formats for some roles⁴.

Identity misrepresentation: what the evidence shows

Fake candidates are a forecast trend. Gartner predicts that by 2028, one in four candidate profiles worldwide could be fake. In a Gartner survey of 3,000 job candidates, 6% admitted to participating in interview fraud, either posing as someone else or having someone else pose as them⁵.

Deepfakes are already used in interviews. In June 2022, the FBI warned that criminals were using deepfakes and stolen personal information to apply for remote IT and software roles with access to customer data, financial data, and corporate databases. It noted that the on-camera actions and lip movement of interviewees did not always align with the audio⁶.

State-sponsored schemes operate at scale. In mid-2025, the U.S. Department of Justice announced coordinated actions against North Korean remote IT worker schemes, including searches of 29 suspected "laptop farms" across 16 states. The schemes used more than 80 compromised identities in one case and obtained employment with more than 100 U.S. companies, in some instances gaining access to export-controlled data⁷. These workers passed hiring processes at real companies, often repeatedly.

What selection science says still works

The research base on employee selection is extensive, and it offers a clear signal about where to invest. For many years, the standard reference was Schmidt and Hunter's 1998 meta-analysis. In 2022, Sackett, Zhang, Berry, and Lievens revisited those estimates, correcting for what they showed was systematic overcorrection for range restriction in earlier work⁸.

Values are operational validity estimates for predicting job performance. Sources: Schmidt & Hunter (1998) as summarized by Plum⁹; Sackett et al. (2022) as summarized by SIOP⁸. Dashes indicate values not drawn on here.

Three conclusions follow for technical and professional vetting.

1. Structure matters more than format. Structured interviews, with consistent questions tied to the job and scored against defined criteria, emerge as the strongest single predictor in the revised analysis. SIOP's summary describes the shift as moving structured interviews into the role of focal predictor⁸.
2. Job knowledge and work samples remain valuable, if their integrity holds. Their predictive value assumes the candidate produced the work. Unsupervised formats undermine that assumption.
3. General puzzle-solving tests are weaker than often assumed. The revised estimate for cognitive ability is lower than the 1998 figure, and in technical hiring, algorithmic puzzles serve as a proxy for it that AI can now complete.

The practical synthesis is that the most predictive methods are those that combine structure with job-relevant content, and that live delivery restores the integrity that unsupervised delivery has lost.

A six-gate vetting process

No single assessment addresses both identity and capability fraud while also predicting performance. A layered process does. The six gates below describe how we vet engineers, clinicians, lawyers, accountants, and linguists for our expert network. Each gate is designed to catch something the others cannot.

Why the live session is the final gate

The live session is placed after the structured assessment and the real-work review deliberately. By that point, the senior practitioner has the candidate's prior answers and work in front of them, which turns the session from a generic interview into an examination of the candidate's own output. That design reflects both research findings above: it is structured, because it follows a defined set of areas and is scored against a rubric, and it is resistant to assistance, because the questions are constructed in the moment from material specific to the candidate.

In practice, a live unscripted session tests capabilities that are difficult to fake:

• Explaining decisions in one's own work. Why this approach and not the obvious alternative? What did you try first? A person who produced the work can answer quickly and specifically.
• Responding to a changed constraint. "Suppose the dataset is ten times larger," "suppose the patient is also on anticoagulants," or "suppose the counterparty is in a different jurisdiction." Real expertise adapts; scripted answers do not.
• Reasoning about failure. Where would this break? What would you monitor? Senior practitioners recognize the answers that come only from experience.
• Calibrated uncertainty. Experts know the limits of their knowledge and say so. Generated answers tend toward uniform confidence.
• Identity continuity. The person in the session should be recognizably the person whose work and assessment preceded it, in knowledge, communication style, and detail.

The practitioner's seniority matters. The interviewing.io results show that experienced interviewers using standard questions did not detect assistance². Detection depends on the ability to construct follow-up questions in real time from deep domain knowledge, which is why the session is conducted by a senior practitioner in the same discipline rather than a generalist recruiter.

Fairness and transparency

Rigor should not come at the expense of fairness. Every gate uses written criteria so that candidates are assessed on the same dimensions. Candidates are told in advance which tools are permitted at each stage, since some roles legitimately involve AI assistance and the assessment should reflect how the work is actually done. Reasonable accommodations are available. Decisions at the sign-off gate are recorded with their rationale so they can be reviewed.

How we measure the process in our engagements

A vetting process should be judged by outcomes, not by its design. We track, for each discipline, the rate at which candidates exit at each gate, the reasons recorded at each exit, the agreement between independent reviewers on sampled sessions, and the relationship between vetting scores and subsequent engagement feedback. These measures are used to recalibrate rubrics and question areas over time. We report them to clients in the context of specific engagements rather than as general claims.

A checklist for buyers of expert talent

Organizations sourcing engineers or domain experts, whether through a network, a staffing partner, or a marketplace, can use the following questions to test a provider's vetting.

1. Identity. How is identity verified, and is it re-verified at the point of live assessment and at the start of work?
2. Credentials. Are professional licenses and qualifications checked against the issuing body, not only against documents supplied by the candidate?
3. Structure. Are assessments scored against written, job-relevant rubrics? Can the provider show an example?
4. Unsupervised components. Which stages are completed without supervision, and how is the integrity of those stages confirmed later in the process?
5. Live assessment. Is there a live session? Is it scripted or constructed from the candidate's own work? Who conducts it, and what is their seniority in the same discipline?
6. AI policy. What is the policy on AI tool use at each stage, and is it communicated to candidates in advance?
7. Accountability. Does a named, accountable person approve each expert, with a recorded rationale?
8. Feedback loop. How does engagement performance flow back into the expert's record, and what triggers re-review?
9. Exit data. Can the provider describe where candidates typically exit the process and why?
10. Remediation. What happens if an expert is found to have misrepresented identity or capability after placement?

Conclusion

AI has changed what remote assessments measure. Tools that answer questions invisibly are commercially available, the problems that once separated strong engineers from weak ones are solvable by models, and organized fraud schemes have placed workers inside more than 100 companies. The selection research points to the response: structured, job-relevant assessment, with its integrity protected. A layered process that verifies identity and credentials, assesses knowledge against rubrics, reviews real work, and then tests that work in a live, unscripted session with a senior practitioner addresses both capability and identity fraud. The live session is not the only gate, but it is the one that confirms whether everything before it was genuine.