Trying to detect AI use on coding assessments is a losing strategy, and the numbers say so. Cheating attempts roughly doubled in 2025, to around 35% of assessments, and even the vendors selling detection admit they cannot reliably identify AI assistance. The companies that assess engineers at the largest scale have stopped fighting and changed the question instead.
When we were shaping hunr's approach, we did the competitive research properly, because "just detect the AI" is the first thing everyone suggests. Here is what that research actually found, and why we concluded the detection arms race is already over.
How common is AI use on coding assessments?
Common enough that pretending otherwise is negligent. The fraud statistics that assessment platforms themselves publish show cheating attempts roughly doubling in 2025, reaching about 35%. That is not a fringe behavior you can police away. That is a third of your funnel.
And the detection side of the arms race is candid about losing. Vendors in this market have publicly admitted they cannot reliably detect ChatGPT-assisted solutions. When the people whose product is detection tell you detection does not work, believe them.
It is worth being precise about what "cheating" even means here, because the word is doing a lot of work. On an assessment that bans AI, using an agent is cheating. In the actual job the candidate is being hired for, using an agent is Tuesday. The rules being broken are rules the industry invented for the test, not rules that exist in the work. That should make us suspicious of the rules, not just the candidates.
Why can't detection win?
Three structural reasons, none of which improve with effort.
The signal decays with every model release. Detection tools are trained to spot the artifacts of today's models. Each new generation writes more naturally, varies more, and looks more like a human engineer. You are aiming at a target that moves every few months, funded by companies whose explicit goal is to make output indistinguishable from human work.
False positives are catastrophic. An AI detector that flags an honest candidate does not just cost you that candidate. It costs you the trust of every candidate who hears about it. Accusing someone of cheating on the strength of a probabilistic classifier is a brand-destroying way to be wrong, and at scale you will be wrong regularly.
Take-homes are unproctorable by nature. A repo-based work sample happens on the candidate's machine, in their editor, over hours. There is no camera angle that covers that honestly. Our research note put it bluntly: proctoring is largely inapplicable to a repo take-home; design-based resistance is the durable lever. You cannot surveil your way to integrity in an open environment. You have to design for it.
There is a fourth reason that is less structural and more empirical: agents are getting very good at defeating the countermeasures themselves. Reward-hacking is pervasive in the research literature — agents that read protected test directories, hardcode expected outputs, even patch the grader. An arms race against an opponent that adapts faster than you can ship countermeasures is not a race you enter on purpose.
What did the industry leaders do instead?
They changed the question. The winners here — Meta, Shopify, Google, Karat, CodeSignal — pivoted to a stance you can summarize as: allow AI, grade whether you understand and can direct it.
That pivot shows up across the market once you look for it. Pure deterministic test-scoring, the run-the-tests-count-the-passes model, has become commoditized; Codility, Coderbyte, and Qualified all sell it, and it no longer differentiates anyone. The premium tier layers human or AI judgment on top of the tests: Woven's human-reviewed scenarios, Byteboard's structured interviews, HackerRank's AI-assisted review for senior engineering work. That judgment layer is expensive when humans do it, which is exactly why it commands a premium: the high-signal players charge on the order of $300 to $4,000 per hire or per review for it.
Even the marketing language has flipped. Assessment vendors now advertise challenges as "tested for AI solvability" — meaning they check whether an AI can trivially solve the problem before shipping it. Two years ago the pitch was catching AI. Now the pitch is designing around it. That is what capitulation looks like, and in this case capitulation is correct.
What should a hiring assessment measure now?
If you cannot know whether an agent helped write the code, then "did an agent help" cannot be the thing you measure. What you can measure is whether the candidate understands the work and can direct it: whether they can explain the design, defend the trade-offs, reason about failure modes, and extend the system when the requirements move. Those are also, not coincidentally, the skills that matter most in a team where everyone uses agents.
This is the position hunr was built on. hunr is the agent-allowed, understanding-verified technical hiring platform: candidates ship real code against role-specific challenges using any AI agent they like, and hunr then verifies they understood what they shipped. We hold both halves of the industry pivot at once. The challenge shape makes one-shotting by an agent a low-signal, low-reward move, because the difficulty lives in messy codebases, ambiguity, and judgment rather than in bare code generation. And a reasoning defense verifies understanding directly, on the candidate's own submission, where borrowed competence has nowhere to hide.
One more finding from the research, because it reframes the whole debate: assessment completion rates collapse when the ask is unreasonable. One 15-hour assignment saw roughly 10% completion; well-scoped take-homes of two to four hours hit about 92%. The industry spent years making assessments longer and more surveilled to fight cheating, and mostly succeeded in driving honest candidates away. The fix was never more enforcement. It was a better-designed test.
The arms race is over. The interesting work is what you build once you stop fighting it.