- When AI Contradicts Biology: How Should Toxicologists Respond?
-
Room 302‚ Convention Center
Chair: Patricia Ceger, RTI International
Co-Chair: Mark Nelms, RTI International
Primary Endorser: Ethical, Legal, Forensics, and Societal Issues Specialty SectionToxicologists are increasingly using machine-learning (ML), large language model (LLM), and other artificial intelligence (AI) tools in their day-to-day work, from basic research to hazard identification and decision making. As these tools become more common, it is equally clear that their predictions do not always align with biological evidence. Models may flag potential concerns where mechanistic or in vivo data suggest little risk, predict low concern where biological data raise red flags, or identify statistically robust patterns that lack clear biological explanation. This Engagement Session uses real-time audience polling, focused case studies, and a live diagnostic exercise to help attendees work through AI/biology conflicts as problems of scientific judgment. These disagreements do not necessarily indicate that either the biology or the model is “wrong.” Instead, they often arise from recognizable sources, including differences in data provenance (e.g., standardized in vitro data versus heterogeneous legacy studies), mismatched endpoints across levels of biological organization, limitations in model applicability, lack of consideration of study quality, differences in background/training knowledge (including the statistical methods that AI and LLMs use for predicting the next “correct” word when creating answers), and the tendency of some models to identify statistically strong but biologically non-informative patterns. This session approaches AI–biology disagreement as a practical challenge of scientific judgment rather than a problem to be eliminated. Through case-based examples, speakers will highlight common sources of discordance, including training data bias, proxy features that appear mechanistically relevant but are not, missing biological context (e.g., metabolism or tissue-specific responses), and apparent conflicts driven by subtle but consequential endpoint misalignment. Building from these examples, the session will introduce a practical six-step framework for working through conflicting evidence based on analytical tradecraft perfected in the intelligence community: 1) Pertinence (data relevance), 2) Source Reliability (study/data quality), 3) Orthogonal Corroboration (triangulation), 4) Mapping the Evidence’s Reliability x Credibility, 5) Analysis of Competing Hypotheses, and 6) Confidence Judgement. Attendees will then apply this framework during a 25-minute interactive “Live Diagnosis” workshop. Using a blind toxicological case study, participants will vote in real time on how to triage, diagnose, and resolve conflicting evidence. The expert panel will facilitate discussion of audience decisions, model competing interpretations, and demonstrate how experienced toxicologists navigate uncertainty when AI-derived predictions and biological evidence diverge. Discussion will focus on how to distinguish meaningful biological conflicts from expected differences across data types; how data quality, provenance, and modeling assumptions shape apparent disagreement; how to recognize the limits of model applicability; and how missing biological context or underlying heterogeneity can influence interpretation. Emphasis will be placed on practical approaches for weighing multiple lines of evidence, documenting uncertainty, and communicating AI-informed judgments in ways that support scientific credibility, regulatory decision-making, and stakeholder trust. Attendees will come away with a clearer understanding of why AI and biological evidence sometimes diverge, along with practical mental tools for recognizing, evaluating, and communicating these disagreements in real-world toxicological decision-making. The session is designed to be accessible to a broad audience, requiring no prior expertise in AI while offering meaningful insights for both early-career scientists and experienced practitioners. This Engagement Session will utilize real-time audience polling, focused case studies, and a live diagnostic exercise to move beyond describing these conflicts and instead ask attendees to evaluate them directly. As attendees vote on how to triage, diagnose, and resolve the conflict, the panel will facilitate a dynamic, bidirectional debate with the room to model real-world scientific judgment.
Session and Activity Type
Enter Note
Go to previous page in this tab
Event
