If an LLM shortlists your applicants, your no-AI rule punishes the people who obey it

The exposed party here is not the applicant deciding whether to run a cover letter through a chatbot. It is whoever signed the purchase order for the screening tool, because the ranking it produces is the organization's output, and the rule it violates is the organization's rule.

Two documents end up in tension, and they are usually written by different people who never compare notes. One is the submission instruction that bans AI-composed material and threatens disqualification for anything suspected of it ^[8]. The other is the scoring prompt handed to the model, which, on Lance Eliot's account of the research, tilts toward text that reads as machine-written ^[2][3]. Nobody reconciles them. The candidate who writes by hand complies with the first document and is graded down by the second ^[11].

The place where this becomes a governance question rather than an etiquette question is the shortlist. In the scenario Eliot sets out, hundreds of submissions arrive, human review would take months, and the model reduces the field to a handful before any judge sees anything ^[6]. The panel then spends an hour on a call and decides ^[7]. That hour is real deliberation, but it is deliberation over a set the panel did not assemble ^[12]. Whatever preference the screener carries has already been applied, silently, at the only stage where most of the field was eliminated. The committee's stated belief was that it was choosing among the best handcrafted entries ^[9].

Worth being honest about the evidence: the column asserts that studies show the preference without naming a study, a sample, or an effect size ^[2][10]. Magnitude matters for practical purposes, since a small bias inside a noisy rubric is not the same as a determinant. But the size of the tilt does not change the structural defect, which is that an institution cannot coherently forbid a behavior and then reward it in the same process.

The second-order mess is worse than a single bias. Eliot notes that screeners are sometimes also instructed to flag AI-generated content, so material tuned for machine receptivity can trip the detector ^[5]. Run both instructions at once and the survivors of the screen are neither the strongest human writers nor the cleanest AI users, but whatever happens to sit in the gap between one model's preference and the same model's suspicion. Nobody designed that population, and nobody can describe it to a rejected applicant who asks how the decision was made.

The narrow auditable question is what the screener was told to reward, in writing, and whether that text contradicts the instruction sent to candidates ^[1][8].