Neural Newscast

Watson for Oncology did not drift in a single dramatic failure. It drifted through a quieter substitution: IBM marketed “artificial intelligence” that sounded like global, data-driven discovery, while the system’s treatment recommendations were trained by a couple dozen physicians at one U.S. hospital—Memorial Sloan Kettering Cancer Center—and could reflect their preferences even when evidence was thin. In a two thousand seventeen investigation, reporters found hospitals and doctors often did not understand that the recommendations were derived from Memorial Sloan Kettering training, not generated as new insights from the broader literature. The system could surface articles and background material, but the recommendation itself came from the human training. As Watson was deployed across countries with different guidelines, drug availability, insurance coverage, and patient populations, complaints followed: American-method bias, confusing or unsupported recommendations, and “concordance” research that mostly measured agreement rather than improved outcomes. The unresolved question is not whether this tool can be useful, but who is responsible for proving safety, efficacy, and fit when a product is already in clinical use—and external, third-party evaluation is effectively optional.