AI is changing the skills test. Can hiring keep up?

Skills-based hiring has won the argument. More companies are moving away from degree requirements and towards testing what someone can actually do. That is the right call. A live coding test or a real work sample can reveal far more about a candidate’s capabilities than a line on a resume. 

But something else has changed. As skills-based hiring has gained ground, AI-assisted cheating in assessments and interviews has become a growing concern for employers. 

In one study of nearly 20,000 AI-led interviews run between mid-2025 and January 2026, over a third of candidates were flagged for getting AI help during the interview itself. A year earlier, that number was under one in ten. On proctored technical tests, cheating attempts more than doubled in twelve months, and for entry-level roles, the rate nearly tripled.

The tricks have moved well beyond having a second tab open. Candidates are using coding assistants during interviews, earpieces connected to AI tools, and, in more sophisticated cases, real-time face-swapping technology. Investigators have also uncovered rings using fake profiles and doctored video to secure remote developer roles under false identities. 

For anyone running recruitment, the concerning part is not the fraud number itself. It is what the fraud reveals.

If an AI tool can pass your test, your test is measuring something an AI can produce.

That is not something proctoring software alone can fix. It means the test was built to measure the wrong thing in the first place.

Two different problems, two different fixes

It helps to split this into two separate risks because they do not get solved the same way. 

The first is identity and honesty. Is this really the candidate, and are they doing the work themselves right now? That is a verification problem. It needs identity checks earlier in the process, live proctoring that watches for odd eye movement, screen activity, or timing patterns, and moments in the interview that cannot be scripted in advance, like an unexpected follow-up question or a small change mid-task. Research shows that most engineering leaders now say AI has made it harder to judge real technical skill, which is exactly why this layer matters more than it did two years ago, not less.

The second problem is how the test itself is built. A question that rewards a polished, generic answer is one that a general AI model can already answer well. No amount of anti-cheating software saves a test like that. It just gets better at catching people who found the gap the test itself left open.

That’s why platforms like Testlify are shifting from ‘catch cheating after the fact’ to designing assessments where AI assistance either doesn’t help or becomes part of what’s being measured.

The cost of getting a hire wrong

Skills-based hiring is now widespread. A recent survey of over 1,000 hiring decision-makers showed that 85% of employers are using skills-based hiring, signalling a shift from degree screens to capability tests. That shift is producing a side effect worth naming. Candidates know assessments carry more weight than they used to, and some are inflating what they can actually do to clear the bar, a pattern some recruiters call skillfishing. It shows up after the test, not during it: someone passes a technical screen but cannot repeat that performance on the job.

The stakes are not evenly distributed either.  According to LinkedIn’s Global Talent Trends research, 92% of hiring professionals say soft skills matter as much as or more than hard skills, and 89% of bad hires are attributed to missing soft skills rather than technical gaps. A candidate can get an AI to produce a working answer and still be the wrong hire, because the test never touched judgment, communication, or how someone handles being wrong. Fraud detection catches the person who did not do the work. It does not catch the person who did the work but would still struggle in the role. Both problems point to the same fix: assessments built around real scenarios and follow-up reasoning, not just a right answer.

What happens to the honest candidate who gets flagged?

Verification only works if it doesn't punish people who are doing nothing wrong. Proctoring systems built on gaze tracking and behavioural flags can misread a candidate who looks away to think, whose accent is poorly recognised by an audio model, or who loses a few frames because of a weak connection. None of that is cheating. Treated as a red flag anyway, it pushes out exactly the candidates that skills-based hiring was meant to bring in.

Regulators are already responding. Under the EU AI Act, emotion recognition in the workplace, including hiring and interviews, has been prohibited since 2 February 2025. At the same time, jurisdictions such as New York City now require bias audits and candidate notice before automated hiring tools can be used. The direction is clear: an algorithmic flag should trigger a human review, not an automatic rejection. That aligns with hiring leaders’ own views - in a 2025 survey, 93% of hiring managers said human involvement remains essential even as AI takes on more of the process. That principle matters even more when a false accusation is on the line.

The practical fix is to treat every flag as a starting point, not a verdict. Give candidates a chance to explain an anomaly before a decision is made, and audit the tool itself for who it flags most often. A system that only optimises for catching cheats will eventually catch honest people too, and lose their trust along with it.

What should employers do instead? 

Companies handling this well are doing two things at once, not picking one over the other.

First, they are being upfront about where AI fits in the hiring process and where it does not. Most technical and analyst roles today genuinely involve working alongside AI tools. If that is the job, testing whether someone can avoid AI entirely misses the point.  Employers should assess how well candidates direct AI, evaluate its output and catch its mistakes. That is the real skill being hired for.

But for roles that require independent thinking, whether that means handling a live customer issue, making a judgment call under pressure or diagnosing a problem no one has seen before, the assessment needs to test what AI cannot do for the candidate. That usually means live and applied, not a take-home assignment with no time limit and no one watching.

Second, they are backing that up with real verification, not just a line in the instructions asking candidates not to use AI, which nobody checks, and everyone knows. This is where assessment platforms are evolving: role-specific tests paired with live proctoring, identity checks built into the assessment itself rather than added after an offer is made, and formats that ask candidates to explain their reasoning out loud, not just submit a finished answer. Testlify brings these capabilities together in its assessment platform, aiming to make cheating harder from the start rather than simply catching it after the fact.

The resume era had one weak point. A polished resume could hide a hollow candidate. The skills assessment era has a new one. A polished performance can now hide AI behind the curtain. The answer is not to walk back skills-based hiring. It is to build tests precise enough that passing one still means something.