Explaining itself made the AI more persuasive and the reviewers less right

The rationale bought no additional compliance. Deference ran at about the same level whether the model argued its case or simply stamped a verdict, so what the written reasoning changed was the composition of the errors, not the amount of following ^[17]. That is the result with operational consequences, and the researchers say it contradicts the common assumption that model explanations augment human decision-making ^[14].

The 21-point gap between how often these reviewers went along with machine advice and how often they went along with human advice is the figure worth carrying into any screening design ^[16]. Direction matters more than volume, though. A rejection commits no budget and creates no owner, and the researchers reach for Xerox shutting down early Ethernet and PostScript work as their illustration of that cost, against Google Glass and Amazon's Fire Phone on the false-positive side ^[9]. Only one of those two failure modes leaves a project to write a post-mortem about.

The mechanism the authors offer is about accountability, not model quality. People weigh negative information more heavily than positive, a pattern the paper labels negativity bias ^[10], and as they put it, "Rejection is an active, eliminative decision that feels more consequential and accountable than preserving optionality," while also maintaining the status quo and requiring no resource commitment ^[11]. A fluent paragraph explaining why a proposal falls short hands the reviewer a ready-made justification for the decision they were already disposed toward, and the study reports evaluators leaning on surface cues such as fluency and coherence ^[12]. The authors describe narrative explanations as suppressing productive overrides, meaning the cases where someone independently verifies a persuasive model output before agreeing with it ^[15][4].

That distinction is where enterprise exposure sits. Screening under time pressure with only basic information is precisely the job LLM recommenders are being handed ^[18]. A stored rationale for each decision is an attractive artifact because it looks like a control and files like evidence. What it evidences is that a reason was available, not that anyone tested it.

The study's ground truth deserves stating plainly: correct means agreeing with four human experts on one challenge's submissions, so this measures alignment with a small panel rather than commercial outcomes ^[3]. Within that frame, the authors end on design rather than warning, writing that "effective human-AI collaboration requires designs that preserve rather than supplant independent human judgment" ^[13]. On this evidence, the explanation layer is a persuasion feature that has been filed under governance.