Large language models can appear remarkably consistent when they recommend a meal, rank job candidates, or help users make everyday choices. Yet new research suggests that their preferences may be far more fragile than their confident answers imply. A study published in PNAS Nexus finds that the order in which options are presented can substantially influence what an artificial intelligence system selects—and, in some situations, can even reverse the model’s underlying preference.
The researchers, Haonan Yin of Iowa State University, Shai Vardi of the University of South Florida, and Vidyanand Choudhary of the University of California, Irvine, examined whether large language models display systematic “order effects.” In psychology, an order effect occurs when the position of information in a sequence changes a decision, even though the available information remains exactly the same. Such effects are familiar in human behavior: people may favor the first item they encounter, remember the most recent item more easily, or use presentation order as a shortcut when alternatives seem similar.
To investigate whether AI systems behave in comparable ways, the researchers tested nine widely used language models: GPT-4o-mini, GPT-4.1-nano, Claude 3 Haiku, Claude Sonnet 4, Llama 3 8B, Llama 4 Scout, Gemini 2.5 Flash, Gemini 3 Flash, and Qwen 3 32B. The models were evaluated in a deliberately low-stakes task that asked them to choose a paint color for a child’s bedroom. By changing the order of the same color options across repeated prompts, the researchers could determine whether a model’s choice was driven by the qualities of the colors or by where those colors appeared in the list.
The results revealed a striking quality-dependent pattern. When all of the available options were judged to be high quality, the models tended to favor the first option. When the choices were of lower quality, however, the models more often selected options appearing later in the sequence. This means that there was no single, universal “first-option bias” or “last-option bias.” Instead, the direction and strength of the positional effect changed depending on the overall quality of the alternatives.
That finding is important because position might ordinarily be dismissed as a harmless tie-breaker. If two options were genuinely indistinguishable, selecting the first one would not necessarily indicate a meaningful distortion in the model’s judgment. But the researchers found evidence that order could do more than resolve a tie. In some cases, rearranging the same options caused a model to select an item it had previously ranked lower, suggesting that the model’s apparent preference itself had been altered by presentation order.
Technically, these effects may arise from the way language models process sequences. A model does not assess a list in the same manner as a human decision-maker comparing objects in a stable internal ranking. It processes tokens through layers of neural attention, generating a response based on the entire prompt and the relationships among its words. The beginning and end of a sequence can receive different effective attention, while linguistic framing, recency, and learned patterns from training data may influence the final output. The result can be a preference that is sensitive not only to what an option says, but also to where it appears.
The researchers also tested three of the models in a resume-screening task, moving the investigation closer to a consequential real-world application. In this experiment, the names attached to resumes were varied while the underlying resume details were held comparable. The study reports that models sometimes preferred certain names regardless of which qualifications were associated with them. The authors attempted to use demographically similar names, yet the systems still produced measurable name-based differences. Claude 3 Haiku, for example, selected “Christopher Taylor” over “Andrew Harris” in 64 percent of cases.
Such behavior raises concerns for automated screening systems. A model may be instructed to focus on education, experience, and skills, but its output can still be influenced by superficial features that are unrelated to job performance. If the candidate information is presented in a different order, or if one applicant’s name appears more frequently in favorable contexts in the model’s training data, the resulting recommendation may shift. In high-stakes settings, even modest and inconsistent effects could accumulate across thousands of decisions.
The study’s broader message is that AI bias cannot be understood only by looking for familiar human categories such as race or gender. Large language models can reproduce social biases present in their training data, but they may also develop system-specific tendencies that are difficult to anticipate from human psychology alone. These “fragile preferences” may remain hidden during a single evaluation, particularly when a model is run at a low sampling temperature, a setting that makes outputs more predictable by reducing randomness.
To expose these latent tendencies, the authors propose repeatedly querying models at higher temperatures. Temperature controls how broadly a model samples among possible next tokens: lower values generally produce more deterministic answers, while higher values permit greater variation. If repeated high-temperature trials reveal that a model alternates between options, reverses its choices when list order changes, or displays unusually unstable preferences, that decision may deserve additional scrutiny. The approach could serve as a diagnostic tool before AI systems are deployed in hiring, healthcare, education, or other environments where recommendations affect people’s lives.
The researchers emphasize that their findings do not mean language models are incapable of making useful comparisons. Rather, they show that an apparently rational answer may depend on hidden features of the prompt, including the sequence in which information is presented. As AI systems move from chat interfaces into decision-support tools, developers may need to test the same case in multiple orders, compare outputs across sampling settings, and audit both item-level and name-level effects. The study suggests that reliable AI will require more than accurate answers on average: it will require evidence that those answers remain stable when irrelevant details change.
Subject of Research: Order effects and model-specific biases in large language models
Article Title: Fragile preferences: A deep dive into order effects in large language models
References: PNAS Nexus, article published 11-Aug-2026
Keywords
Large language models, artificial intelligence, AI bias, order effects, position bias, machine learning, resume screening, algorithmic fairness, model reliability, temperature sampling
