Caught in the Web of Words: Do LLMs Fall for Spin in Medical Literature?

1. Problem Statement

Medical research often suffers from "spin," a reporting strategy where authors overstate the benefits of experimental treatments or downplay harms, particularly when primary outcomes are statistically non-significant. This bias is prevalent in article abstracts, which clinicians frequently rely on for decision-making. While human experts are known to be susceptible to this spin, the impact of spin on Large Language Models (LLMs)—which are increasingly used to synthesize, summarize, and interpret medical literature—remains unexplored.

The core research questions are:

1. Can LLMs detect the presence of spin in medical abstracts?
2. Does the presence of spin alter LLMs' interpretation of trial results (e.g., perceived treatment benefit)?
3. Do LLMs propagate or amplify spin when generating simplified, plain-language summaries for lay audiences?

2. Methodology

Data

The study utilized a high-quality, manually curated dataset from Boutron et al. (2014) consisting of 60 cancer-related Randomized Controlled Trial (RCT) abstracts.

• Structure: 30 original abstracts containing spin (where results were statistically non-significant but conclusions were overly optimistic) paired with 30 "neutral" versions manually rewritten to remove spin.
• Controls: Treatment names were masked (e.g., "Treatment A" vs. "Comparator B"), and author/journal details were removed to prevent bias.
• Extension: To test generalizability, the authors created a semi-synthetic dataset of 150 abstracts from orthopedics, emergency medicine, and obesity, using GPT-4o to generate "unspun" versions of real-world spun abstracts.

Models Evaluated

The study evaluated 22 LLMs across three categories:

• Generalist Closed-Source: GPT-4o, GPT-4o Mini, GPT-3.5, Claude 3.5 Sonnet/Haiku, Gemini 1.5 Flash/8B.
• Generalist Open-Weight: Llama 2/3 (various sizes), Mistral, OLMo2.
• Biomedical Open-Weight: AlpaCare, BioMedGPT, BioMistral, Med42, OpenBioLLM.

Experimental Tasks

1. Spin Detection: A binary classification task (Yes/No) to determine if an abstract contains spin.
2. Result Interpretation: LLMs were asked to rate 5 specific metrics on a 0–10 scale for both spun and unspun abstracts:
   • Treatment benefit
   • Study rigor
   • Study importance
   • Interest in reading the full text
   • Interest in running another trial
   • Metric: The Mean Difference in scores between spun and unspun versions was calculated. A positive difference indicates the model was influenced by spin to view the treatment more favorably.
3. Simplification (Propagation): LLMs generated plain-language summaries (targeting a 5th-grade reading level) of the abstracts. These summaries were then evaluated by other LLMs to see if the bias persisted in the simplified text.
4. Mitigation Strategies: The authors tested two prompting strategies to reduce susceptibility:
   • Explicit Labeling: Providing the model with a label indicating whether the abstract contains spin (using ground truth or model-inferred labels).
   • Joint Prompting (Chain-of-Thought): Asking the model to first detect spin and then interpret results based on that detection within a single prompt.

3. Key Results

A. Spin Detection Capabilities

• LLMs demonstrated moderate to strong ability to detect spin, with an average accuracy of 0.67 across all models (significantly above the random baseline of 0.50).
• Claude 3.5 Sonnet achieved the highest accuracy (0.97), followed by GPT-4o Mini (0.85).
• Crucial Finding: High detection accuracy did not correlate with immunity to spin. Even models with >0.80 detection accuracy were heavily influenced by spin when interpreting results.

B. Susceptibility to Spin (Interpretation)

• LLMs are more susceptible than humans: When comparing the mean difference in scores between spun and unspun abstracts, LLMs showed significantly larger biases than human experts (clinicians and researchers).
  • Example: For the "Treatment Benefit" question, the mean difference for human experts was 0.71, whereas the average for all LLMs was 3.38.
• Universal Susceptibility: All 22 models, regardless of size (7B vs. 70B) or domain specialization (Generalist vs. Biomedical), were more susceptible to spin than human experts.
• Biomedical models did not outperform generalists: Specialized medical LLMs (e.g., Med42, OpenBioLLM) did not consistently outperform generalist models in resisting spin.

C. Propagation of Spin

• LLMs propagate spin into plain-language summaries. When spun abstracts were simplified, the resulting summaries retained the optimistic bias.
• When these simplified summaries were re-evaluated, the mean difference scores remained high (e.g., 4.32 for OpenBioLLM 70B), indicating that the bias was not only preserved but potentially amplified for lay readers.

D. Mitigation Effectiveness

• Explicit Labeling: Providing the model with a "spin" label reduced the mean difference in treatment benefit scores from 3.38 (baseline) to 1.94 (using reference labels) and 2.59 (using model-inferred labels).
• Joint Prompting: The most effective strategy. Asking the model to detect spin before interpreting results reduced the mean difference for treatment benefit to 1.14. This approach significantly outperformed the baseline and explicit labeling, suggesting that forcing a reasoning step (Chain-of-Thought) helps the model critically evaluate the evidence despite the linguistic framing.

4. Key Contributions

1. Empirical Evidence of Vulnerability: The first comprehensive study demonstrating that LLMs are more susceptible to spin in medical literature than human experts, even when they can technically detect the spin.
2. Propagation Risk: Evidence that LLMs do not just passively absorb spin but actively propagate it into simplified summaries, posing a risk to patient understanding and health literacy.
3. Mitigation Framework: Demonstration that prompt engineering (specifically joint detection/interpretation) can significantly mitigate the impact of spin, offering a practical solution for developers and users.
4. Dataset Expansion: Creation of a semi-synthetic dataset to generalize findings beyond oncology to other medical fields.

5. Significance and Implications

• Clinical Decision Making: As LLMs are increasingly deployed to summarize medical evidence for clinicians and patients, their susceptibility to spin could lead to the adoption of ineffective treatments or the dismissal of valid ones based on biased summaries.
• AI Safety in Healthcare: The findings highlight a critical gap in the "reasoning" capabilities of LLMs. While they excel at pattern recognition (detecting spin), they struggle with discourse reasoning (ignoring the spin when evaluating the underlying data).
• Implementation Guidelines: The study suggests that relying on LLMs for evidence synthesis without specific safeguards (like joint prompting or explicit bias warnings) is dangerous. Future medical AI systems must incorporate mechanisms to explicitly counteract reporting bias.
• Future Research: The authors call for larger, diverse datasets and human-in-the-loop evaluations to further refine mitigation strategies and ensure LLMs serve as reliable, unbiased tools in evidence-based medicine.