LLMs approach expert-level clinical knowledge and reasoning in ophthalmology
Large language models like GPT-4 are reported to score near expert ophthalmologists on written exam questions, prompting debate over what this actually implies for real-world medical practice. Commenters question the study’s rigor, noting that reused textbook questions may already be in training data and that standardized tests emphasize memorized knowledge rather than hands-on skills, ethics, or nuanced clinical reasoning. Many see clear potential for AI to aid diagnosis—especially where access to specialists is poor—but warn that hallucinations, shallow reasoning, and existing problems in medical training and incentives must be addressed before relying on such systems in patient care.
Methodology and Validity of the Study
- Several commenters question the study’s design: the exam questions came from a standard ophthalmology textbook, not newly written items.
- People note that such textbooks are likely in common LLM training sets (via LibGen, “Books2”, or direct scanning), so test leakage is considered very plausible.
- Because of this, some argue the paper mainly tests memorization rather than reasoning, and might be “bunk” without stronger guarantees about training data and novel questions.
Limits of Exam-Based Benchmarks
- Multiple comments stress that passing written exams is not equivalent to real clinical practice.
- Standardized exams often emphasize recall and pattern-matching over true diagnostic reasoning, communication, or dealing with uncertainty.
- Some link this to broader critiques: LLMs exposing how professional exams and medical education overweight rote knowledge rather than real-world skills.
Capabilities and Risks of LLMs in Medicine
- The reported performance (GPT-4 near expert level on questions) aligns with many users’ personal experiences of LLMs being useful for symptom triage, blood test interpretation, or differential suggestions.
- Others, including at least one clinician, say LLM outputs look impressive to laypeople but break down on nuanced differential ranking and management decisions, sometimes suggesting dangerous treatments.
- Hallucinations are seen as a central blocker for direct medical use. Proposals include multiple sampling and self-consistency, but commenters note lack of solid evidence these fully solve the problem.
Broader Reflections on Healthcare Quality
- The thread contains many anecdotes of misdiagnosis, delayed diagnosis, and perceived superficial thinking by doctors, fueling optimism that AI tools could eventually outperform average practitioners.
- Others attribute problems to systemic constraints: 10-minute visits, per-visit payment models, long and expensive training, fragmented records.
- There are calls for better outcome-based metrics for physicians and recognition that LLMs’ exam success also highlights weaknesses in how clinicians are trained and evaluated.
Attitudes Toward Future Models
- Some are highly bullish that future models (e.g., a hypothetical GPT-5) will surpass top diagnosticians and be widely accessible.
- Others are skeptical, citing possible performance plateaus and unresolved safety, ethics, and reliability issues.