Cognitive Biases Can Mislead Radiology LLMs
Study shows authoritative and misleading prompts reduce accuracy
Large language models (LLMs) are being increasingly explored in radiology for education, report support, guideline retrieval and clinical decision support. However, a recently published study in Radiology: Artificial Intelligence demonstrated that LLMs are susceptible to cognitively biased inputs and the authors say safeguards must be put in place to mitigate their influence.
In the study conducted at the University of Toronto, lead author Nicholas T. Dietrich, MD, MEng, and colleagues investigated whether cognitively biased prompts degrade LLM accuracy in answering radiology board-style questions.
They evaluated the performance of 10 LLMs, including five non-reasoning models and five reasoning models, answering 200 text-based and 200 multimodal (text and image) questions. Non-reasoning models are designed for broad general-purpose dialogue, whereas reasoning models are embedded with multistep problem-solving strategies.
“I previously studied adversarial AI, exploring ways in which machine learning models can be exploited,” said Dr. Dietrich, a radiology resident. “I started considering other ways these models could be deceived, and one possibility that was lying in plain sight was the actual inputs.”
While cognitive biases are well-characterized drivers of diagnostic error in human decision-making, Dr. Dietrich said their effects on LLM performance in radiology remain poorly understood.
“Unlike traditional adversarial attacks, cognitively biased prompts may arise inadvertently when a radiologist embeds an assumption into a query,” he explained. “Through my work in prompt engineering, I understood how the user brings their own views, biases and heuristics into the model in ways that can affect the outcome. For safe and reliable use in clinical workflows, LLMs must be tested for their robustness in evaluating diverse and potentially deceptive inputs.”
Researchers Test Common Forms of Cognitive Bias
To test the effects of cognitive bias, the researchers evaluated 400 radiology board-style questions drawn from a single institution-based database and reviewed by a board-certified interventional radiologist.
Each question included one correct and three incorrect answers, and was presented under a baseline prompt and three additional prompts containing:
- Authority bias (ABP), offering a confident, expert-like statement.
- Complexity bias (CBP), suggesting the more sophisticated answer was preferable.
- Anchoring bias (AnBP), introducing a salient but misleading fact.
“We tested prompts that mimic distorted reasoning within human decision-making,” Dr. Dietrich said.
The researchers also attempted to mitigate the effects of the biased prompts using two approaches. The first was a bias audit, which instructed the model to identify and disregard potentially misleading cues before selecting an answer. The second was a one-shot mitigation strategy, which gave the model one example of a biased prompt leading to an incorrect answer, then asked it to avoid the same error in the next question.
Under the baseline prompts, the models averaged 84.8% (±5.5 percentage points) accuracy on text-only questions and 59.5% (±7.7 percentage points) on multimodal questions. All three biased prompts significantly reduced LLM accuracy, with declines of up to 21.1% and 44.9%, respectively.
Authority bias had the greatest effect on text-only questions, reducing mean accuracy to 63.7%, compared to 74.7% for complexity bias and 80.4% for anchoring bias. For multimodal questions, authority and complexity bias were similarly disruptive, with mean accuracies of 14.6% and 15.1%, respectively, compared with 19.9% for AnBPs.
“This likely reflects training processes that reward outputs that follow confident phrasing, leading models to treat authoritative wording as a proxy for correctness,” Dr. Dietrich said. “In contrast, prompts like those that highlight salient but irrelevant details, may be easier for models to discount.”
The cognitively biased prompts had the greatest effect on multimodal questions, suggesting that misleading language may override visual information during reasoning.
Mitigation Strategies Improve, But Don’t Eliminate Bias
In his accompanying commentary, Soroosh Tayebi Arasteh, PhD, noted that the multimodal findings deserve special attention. “Much of the future promise of AI in radiology lies in systems that combine images and language,” Dr. Arasteh explained.
“Dr. Dietrich’s study is important because it tested both the breadth of LLM knowledge of radiology and the models’ ability to maintain performance when questions are framed in a misleading but clinically plausible way,” said Dr. Arasteh, a postdoctoral AI researcher and lecturer at University Hospital RWTH Aachen, Germany. “What makes these results especially relevant for radiologists is that the biased prompts are not exotic adversarial tricks; they are recognizably human.”
The mitigation strategies partially restored accuracy, though neither returned to baseline levels. While the two approaches performed similarly overall, each showed strengths in different ways. The prompt bias audit increased accuracy by 5.6% for text-based questions and 15.8% for multimodal questions. The one-shot mitigation yielded gains of 4.0% for text-based questions and 24.9% for multimodal questions.
Among the LLMs, the reasoning models achieved higher overall accuracy compared to non-reasoning models, which were disproportionately affected in the multimodal setting.
“These findings highlight the need for further development of defense strategies,” Dr. Dietrich said. “Given the subtle and human-like nature of biased prompts, it’s unlikely that any single defense will be universally effective.”
Dr. Arasteh said the safe deployment of LLMs will require multiple layers, including adversarial stress testing, prompt-level safeguards, uncertainty signaling and external grounding via retrieval or tools. He also noted that the study’s findings could help trainees recognize how both human and machine reasoning can be distorted by framing.
“As LLMs move closer to routine educational and clinical workflows, they should be evaluated not only for what they know but also for how easily they can be led astray,” Dr. Arasteh said.
For More Information
Access the Radiology: Artificial Intelligence study, “Cognitively Biased Prompt Effects on Large Language Model Accuracy for Radiology Board-style Examination Questions,” and the related commentary, “When Framing Shapes the Answer: Cognitive Bias and Large Language Model Reliability in Radiology.”
Read previous RSNA News stories on large language models in radiology: