Swapping Answer Options Exposes Hidden Weaknesses in AI Decision Models

Structured answers do not guarantee reliable decisions

System One decision models return structured judgments rather than conversational responses. They can select an option, assign a rubric score or answer a boolean question, often with probabilities attached. Software can act directly on those outputs in workflows such as ticket routing, content moderation and candidate scoring.

In an October 2026 preprint, Alejandro Mozo Quesada, Sergio Gálvez Rojas and Francisco López Valverde of Universidad de Málaga examine whether this interface reduces familiar demographic biases or merely changes how they appear. Their audit covers twelve models plus a configuration with thinking disabled, reporting 1,638,000 decisions across three adapted benchmarks.

The central finding is not a single ranking. According to the authors, stereotype alignment, presentation sensitivity and confidence measure different aspects of behavior, and none can substitute for the others.

An audit built around swaps

The researchers adapted BBQ, a question-answering benchmark covering eleven demographic categories; CrowS-Pairs, which contains paired sentences across nine bias types; and SilverSpoon, which compares moral judgments with labels from annotators at opposite ends of the socioeconomic spectrum.

For BBQ, models received three answer options, including an uncertainty response. Each item appeared twice, with the first two options exchanged. CrowS-Pairs sentences also appeared in both orders. SilverSpoon used a fixed boolean question, with a second pass serving as a replicate rather than an option swap.

Additional controls changed option labels for native typed models and repeated ten configurations under different shuffle seeds and a fixed batch size. These checks sought to separate presentation effects from execution variability.

Abstention can conceal a directional pattern

On ambiguous BBQ questions, insufficient information makes the uncertainty option correct. Across thirteen configurations and eleven categories, the authors report that 97 of 143 bias scores had 95% bootstrap intervals excluding zero. Of those, 94 pointed toward the benchmark’s targeted stereotypes.

Clef and Clef-flash showed stereotype-aligned scores in all eleven categories. Jev, the hosted commercial reference, looked much closer to zero but abstained on 99.5% of ambiguous items.

That distinction matters because the official ambiguous-context score shrinks as correct abstention rises. When the researchers scored probability mass rather than only the selected answer, Jev’s scores increased in every category: its less-probable substantive answers leaned toward stereotypes even when uncertainty won. This does not make abstention incorrect; it shows that selected-answer and distribution-level audits answer different questions.

Order can dominate the result

Reported BBQ answer-flip rates ranged from zero to 72.6% across model-category combinations. Von remained unchanged under the tested option swaps, while Laya changed answers on roughly 22–33% of items. Several letter-prompted configurations were substantially less stable.

Repeated execution produced much smaller changes overall, supporting the authors’ interpretation that option presentation, rather than ordinary rerun noise, drove the largest effects.

CrowS-Pairs exposed another trap. The adapted task asked which sentence was more stereotypical, so identifying a stereotype should not automatically be equated with endorsing it. Nevertheless, the order control revealed clear response artifacts: the thinking Qwen3.5-9B base almost always answered yes. Its probability of identifying the stereotypical sentence therefore changed from 1.000 when that sentence came first to 0.005 when it came second. Averaging those results looked nearly neutral while hiding a largely constant response.

Confidence is not a fairness certificate

On SilverSpoon, five of eight decision models aligned more closely with lower-socioeconomic-status annotator labels. This is agreement with a particular label set, not proof of greater moral competence. The two human groups disagreed on 41.7% of dilemmas, and a constant “wrong” response itself favors the lower-status labels under this comparison.

Confidence also offered no universal shortcut. Mean confidence was lower on flipped than stable items in 79 of 82 eligible model-dataset comparisons, but its relationship with stereotype alignment varied across models. Confident answers could remain systematically aligned.

What the findings support

The authors recommend reordered options, reversed boolean questions, stability reporting and examination of full distributions. They also propose order-invariant scoring as a procurement requirement.

The evidence remains bounded by English-only, US-centric benchmarks and the tested serving paths. SilverSpoon used one question phrasing, some unparsed responses were excluded, and confidence came from different mechanisms. The preprint therefore supports a stricter audit protocol—not a certification that any tested model is bias-free.