A new arXiv paper posted Sept. 24 examines AI decision models, which score a fixed set of choices from text. With its main model, Jev, researchers found brief, natural-sounding additions that shifted the choice from right to wrong in 312 of 508 decisions the model initially answered correctly (61.4%), while the original question, options and correct answer stayed the same.
Jev scores a fixed set of choices from a text input. Software can use that kind of decision interface to route requests or select tools without first interpreting a free-form answer. The study tested this behavior on public benchmark tasks, not in live services.
For each test item, the researchers fixed one incorrect option in advance: the wrong choice Jev initially rated highest. A separate language-model proposer, given both the right answer and that target, added background or procedural details around the original text. It could not change the question, choices or correct answer, and another model checked whether the addition remained coherent and left the item answerable.
The search used Jev’s option scores to choose which additions to refine, and could make up to 64 accepted evaluations for each decision. Successful contexts added a median of 31 words. Similar targeted flips appeared in three other decision systems, each measured on its own set of initially correct decisions.
These are results from a targeted stress test, not an estimate of how often ordinary messages disrupt deployed AI systems. The search knew the correct answer and the wrong target in advance; the experiments used bounded public tasks and did not measure effects on live services or real tools. The authors suggest tracking where context comes from, testing whether choices stay stable under answer-preserving additions, and allowing a model to abstain when its option scores are unstable. Whether those checks prevent failures in deployment remains untested.
Sources