Google DeepMind Calls for Rigorous Moral Testing of AI as LLMs Take On Therapist and Medical Advisor Roles

Google DeepMind is urging the AI industry to rigorously test the moral behavior of large language models, warning that as LLMs take on sensitive roles like therapists and medical advisors, their ethical responses may reflect performance rather than genuine reasoning, with models shown to reverse moral stances based on minor input changes.

• Google DeepMind is calling for the moral behavior of large language models to be evaluated with the same rigor as their coding and math abilities, as LLMs take on increasingly sensitive roles like companions, therapists, and medical advisors.
• Research shows LLMs can appear morally competent but are unreliable, often reversing their answers to ethical questions based on minor formatting changes, user pushback, or superficial cues, raising concerns about whether their moral responses reflect genuine reasoning or mere performance.
• Google DeepMind researchers propose new evaluation techniques, including stress-testing models on moral consistency, chain-of-thought monitoring, and mechanistic interpretability, while also acknowledging the unresolved challenge of building LLMs that reflect diverse global moral values rather than predominantly Western ones.