Bioalignment: Measuring and Improving LLM Disposition Toward Biological Systems for AI Safety

Most frontier LLMs show systematic bias toward synthetic solutions: Testing 10 state-of-the-art language models revealed that the majority exhibit preferences for non-biological technological approaches across materials, energy, manufacturing, and algorithms—a potential safety concern if these models influence real-world decision-making in these domains.

AI biases can be measured and tracked: The researchers developed a Kelly criterion-inspired evaluation framework using 50 carefully designed prompts to quantify "bioalignment," demonstrating that we can systematically measure whether AI systems favor certain solution types, which is essential for responsible AI oversight.

Fine-tuning can shift AI preferences without sacrificing performance: Training two smaller models on biological problem-solving content (~22 million tokens from scientific articles) significantly increased their preference for biological solutions while maintaining their general capabilities, suggesting targeted interventions can reshape AI decision-making.

Small models can be effectively retrained: Even modest-sized open-weight models (3 billion parameters) showed statistically significant improvements in bioalignment through fine-tuning, indicating that responsible AI adjustments don't require massive computational resources and could be more widely implemented.

This approach may scale to larger systems: While the study focused on smaller models, the authors suggest the methodology could extend to frontier models, potentially offering a practical pathway for improving how advanced AI systems evaluate different technological approaches across critical domains.