Generalization bias in large language model summarization of scientific research

Area of Science:

• Artificial Intelligence
• Scientific Communication
• Research Integrity

Background:

• Large language models (LLMs) offer potential for science communication by simplifying complex research.
• However, LLMs may oversimplify findings, leading to inaccurate generalizations beyond study scopes.

Purpose of the Study:

• To evaluate the accuracy of LLM-generated scientific summaries.
• To quantify the tendency of LLMs to overgeneralize research conclusions.

Main Methods:

• 10 prominent LLMs were tested, generating 4900 summaries.
• LLM summaries were compared against original scientific texts for accuracy and scope.
• LLM summaries were directly compared to human-authored summaries.

Main Results:

• Most LLMs, even when prompted for accuracy, overgeneralized scientific results.
• Specific models like DeepSeek, ChatGPT-4o, and LLaMA 3.3 70B showed high rates of overgeneralization (26-73%).
• LLM summaries were nearly five times more likely to contain broad generalizations than human summaries (OR=4.85, p<0.001).

Conclusions:

• Widely used LLMs exhibit a significant bias towards overgeneralizing scientific conclusions.
• This bias poses a substantial risk for misinterpreting research findings on a large scale.
• Mitigation strategies include adjusting LLM parameters and developing accuracy benchmarks.