Do admissions officers penalize essays they suspect are AI-written?

The paper's account of why AI-written essays were admitted less often runs through an experiment with admissions officers. The introduction describes it as a way "to understand the source of this apparent penalty." The abstract reports the finding: officers "can often recognize AI writing and rate essays they believe to be AI-generated lower than essays they believe to be human generated." The discussion adds that staff "could often recognize AI-written essays and appear to have penalized them." The word "appear" is the authors' own hedge. The excerpt supports a pattern in which ratings follow suspected authorship. It treats the penalty as a plausible reading of that pattern, not as a demonstrated cause.

The mechanism turns on belief about origin. The abstract ties the ratings to what officers "believe" an essay is, so the variable doing the work is perceived provenance rather than the text's measured quality. Two claims are separable here. One is discrimination: officers can often tell AI from human text. The other is penalty: they rate what they suspect is AI lower. The excerpt reports the first only as "often," with no figure. A penalty keyed to suspicion would also fall on human essays that are wrongly suspected, and the excerpt does not say whether that happened.

Against the neighbors, the rhetorical-sensitivity note reports that an LLM reviewer's scores move with rhetorical framing while reported content is preserved. The officer experiment has a similar shape, an evaluator whose judgment shifts with a cue about the text. The cue differs, though: here it is perceived provenance, not framing, and the excerpt does not describe a content-held-fixed comparison. The temporal signature note detects AI contribution from process data, while this paper relies on output detectors whose error rates the excerpt does not report. The sibling note states the overall penalty from application records. This experiment is offered as its explanation, and both come from the same paper, so they are not independent confirmation of each other.

The excerpt does not establish the experiment's design, the number of officers or essays involved, how often officers' beliefs were correct, or whether the lower ratings in the experiment carried over into the admissions decisions. The link between the two is proposed, not measured in the same sample. What follows is a practical risk and a test. If a penalty tracks suspicion, it can land on human writing as well as AI writing. A useful next check would ask whether the penalty persists when officers cannot see provenance, or when essays are judged without any suspicion of AI use. The excerpt raises that question and cannot answer it.