A Blanket AI Ban Only Works on the Candidates Who Obey It

What happens when a candidate ignores the ban?

Nothing, in almost every case. The submission arrives looking like everyone else's, it gets read on its merits, and the rule never touches the outcome. For the ban to bite, somebody has to establish authorship and then act on it, and both halves fail: the establishing is unreliable, and a rejection resting on it is one you would not want to explain.

The realistic sequence runs like this. A reviewer feels something is off about a submission, says so in the debrief, and somebody asks what the basis is. The honest answer is that the writing felt smooth. At that point the team either drops it, which means the ban did nothing, or acts on it, which means a person lost a job over a hunch nobody can check.

The one enforcement mechanism that does something, watching the candidate work, does not reach the failure it is aimed at either. A locked-down browser and a shared screen cover the machine you can see. The second device is a phone lying face-down on the desk, and no remote arrangement resolves it. So even the expensive version of the ban is a rule about one laptop, not a rule about a person.

Bad faith is not the point here. Those are the only two branches available, and neither of them is the one the rule was written to produce.

Why the checks cannot carry a rejection

Because each one was measured and came back short. Untrained readers asked to tell GPT-3 text from human writing performed at random chance, and three quick training methods lifted them only to about 55%, inconsistently across text types ¹. A study of 12 public detection tools plus Turnitin and PlagiarismCheck concluded the available tools are neither accurate nor reliable ². Neither result supports ending anyone's candidacy.

The errors are also not evenly spread, which is the part that turns an unenforceable rule into an unfair one. Seven detectors run over 91 human-written TOEFL essays by non-native English speakers produced an average false-positive rate of 61.3%, while the same tools were near-perfect on essays by US eighth-graders ³. Rewriting the essays with richer vocabulary cut that rate to 11.6%, which is to say the tools were reading fluency, not authorship.

Volume makes a small error rate expensive. Vanderbilt University switched off Turnitin's AI detector in 2023 and published the arithmetic: at the vendor's claimed 1% false-positive rate, the 75,000 papers the university submitted in 2022 would have produced around 750 wrongly labeled papers ⁴. A pipeline of a few thousand applications runs the same calculation with a person's job on the other end of it. Whether any of this can be run on a transcript is handled in do AI detectors work on interview transcripts.

Who a blanket ban actually removes

The compliant, and disproportionately the cautious. An unchecked rule is honored by people who honor rules, which is not the skill the req was opened for. The candidate who ignored it stays in the pool and now looks identical to the one who worked alone, so the ban narrowed the field without improving it.

The ban also clears out a quieter group. Candidates who use assistive tooling for reasons that have nothing to do with the assignment, including people writing in a second language and people using accessibility software, read a blanket prohibition as a risk they cannot size. Some of them withdraw. None of them tell you why.

Run the trait question honestly and it gets uncomfortable. The people most likely to comply with an unverified instruction are early-career candidates, candidates who need this job more than the next one, and candidates for whom an employer's stated rule is not something to test. The people least likely to comply are the ones with the most options. The rule is not neutral about which of those groups it disadvantages.

And it costs the signal the round was supposed to produce. A candidate who used four tools well and can say exactly where each one was wrong is the most informative person in the pipeline, and a ban is an instruction to hide that from you. The narrower stage-level version of the question, whether to open the interview specifically, is worked through in whether to let candidates use AI during the interview.

Replace the ban with a rule you can act on

Swap the prohibition for a posture per stage plus a conversation. State which stages are AI-open, say plainly that nothing is checked for authorship, and put fifteen minutes of questions about the submitted work at the front of the next call. The consequence then lands on the exercise itself, and that is the only version of this that survives being challenged.

Three moves, in order:

• Decide per stage, not per process. Some stages stop measuring anything once an assistant is open, and some do not care at all. Setting the AI rule stage by stage is the decision underneath this one, and answering it usually makes the blanket version unnecessary.
• Publish the posture before anyone applies. A rule stated in the posting is a term everybody accepted. A rule stated in the assignment email arrived after the candidate committed their evening.
• Make the follow-up the enforcement. Someone who cannot account for the choices in their own submission fails the exercise on its own terms, with no claim about authorship anywhere in the record. The design work behind that is making the AI rule hold without watching anyone.

If the decision is still a ban, publish it in the same plain words and go in knowing what it is: a request that most people will grant, backed by nothing, sorting your pipeline on a trait you did not mean to select for.

Whichever way the decision goes, write down what the rule is for. A prohibition set to protect one stage should be dropped when that stage is rebuilt, and a prohibition set because leadership wanted a public position should be recorded as that. The distinction matters in six months, when somebody asks whether the rule is still doing anything and the file has no answer.

Read the evidence >