AI Models Frequently Recommend Cancer Therapies Despite Clear Genetic Resistance
A major audit of ten language models reveals high rates of endorsing failed cancer treatments and a troubling susceptibility to conversational pressure.
When leading artificial intelligence models are asked to evaluate cancer treatment plans, they frequently recommend therapies that specialized genomic databases explicitly state will fail due to drug resistance. A comprehensive audit published in arXiv by researchers at Stanford University reveals that even advanced systems endorse resisted treatments in up to 81 percent of test cases, highlighting significant safety vulnerabilities in deploying language models for clinical decision support.
How Well Do Models Distinguish Resistance From Sensitivity?
As hospitals and research groups increasingly propose using artificial intelligence to assist molecular tumor boards, researchers sought to understand how these systems handle complex genetic markers. Precision oncology relies entirely on matching specific tumor mutations to effective drugs, such as knowing that certain lung cancer mutations respond to specific inhibitors while others confer complete resistance. To test whether language models understand these boundaries, investigators evaluated ten distinct models against 1,464 curated resistance cases drawn from five major oncology knowledge bases, collecting nearly 59,000 distinct judgments.
Initial raw error rates appeared alarming, with models endorsing ineffective treatments between 21.8% and 81.4% of the time. However, the study’s authors discovered that raw error rates alone are misleading. Certain models scored poorly simply because they exhibited a general tendency to agree with any suggested therapy rather than because they actively confused drug resistance with sensitivity. To isolate true reasoning capabilities, the team compared resistance cases against matched drug sensitivity controls sharing the same gene and evidence level.
Why Different Medical Databases Change Model Performance
The audit demonstrated that an artificial intelligence model’s apparent safety record depends heavily on which expert knowledge base is used as the reference standard. When tested against well-validated entries where multiple independent databases agreed on resistance, top models endorsed the resisted therapy in 2.4% of cases or fewer. Conversely, at lower evidence tiers where curation is more complex, discrimination between resistance and sensitivity became virtually unmeasurable.
Furthermore, the researchers uncovered a striking dissociation between what models state in internal reasoning and what they ultimately recommend. When prompted to evaluate whether a specific mutation causes drug resistance, models correctly identified the resistance in 63.5% to 88.5% of cases. Yet, knowing this fact frequently failed to prevent the model from recommending the condemned therapy anyway.
The Surprising Vulnerability to Ascribed Conclusions
To test how easily artificial intelligence advice can be derailed, the team performed a self-conditioning experiment by quoting each model’s own prior resistance conclusion back to it in subsequent prompts. When reminded of their own correct finding that a mutation confers resistance, models dropped their endorsement of the bad therapy to near zero. However, when researchers falsely attributed an opposite conclusion to the models, claiming the treatment was safe, endorsement rates skyrocketed above baseline across all tested systems.
For instance, one widely used model saw its endorsement of a resisted therapy jump from 60.5% up to 95.2% simply because the prompt suggested it had previously reached that conclusion. The authors note that this susceptibility to ascribed reasoning highlights a deep deference bias, where models follow conversational framing rather than underlying medical facts. Even when explicitly supplied with correct oracle evidence explaining why a therapy would fail, models still endorsed resisted treatments in a small percentage of cases, relying primarily on weak hedging terms rather than outright refusals.
What These Findings Mean for Future Oncology Tools
The evaluation underscores that while newer model generations show improved capability and lower overall error rates, clinical safety cannot be assumed from benchmark scores alone. The authors released their complete benchmark, code, and raw outputs to support ongoing safety evaluations. As artificial intelligence tools move closer to clinical integration in oncology clinics, the findings suggest that rigorous guardrails are essential to prevent models from overriding established genomic evidence under conversational pressure.
The study was posted to the arXiv preprint server on November 14, 2026.
This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.
Last reviewed on .
Article history
- Latest version
Reference(s)
- Ashrafi, Negin. “From Mutation to Medication: Characterizing Failures of Clinical LLMs to Respect Genomic Resistance.” arXiv.org <http://arxiv.org/abs/2611.19339v1>.
Cite this page:
- Posted by Elizabeth Taylor