Attention limits in LLMs threaten clinical record reviews
Long-context language models drop accuracy during extended cognitive tasks, creating silent failure points when parsing complex longitudinal patient records.
Health systems increasingly pilot large language models to summarize sprawling multi-year electronic health records. Cognitive psychologists at the University of Cambridge recently subjected leading commercial reasoning models to classic attention batteries, including extended versions of the Stroop test. In short sequences, the models maintained accuracy above 92%. Once lists scaled past 200 items, accuracy in several top-tier models collapsed to under 15%.
The degradation mirrors human cognitive fatigue, but without the physiological warning signals. In long input sequences, attention mechanisms encounter interference, prioritizing tokens unevenly and missing contradictions scattered across hundred-page histories. For a clinician scanning an automated summary, that arithmetic failure translates into vanished adverse reactions, omitted lab trends, or inaccurate medication histories.
Medical teams often trust long-context windows as neutral file readers. This trial confirms that context windows act more like porous sieves under sustained informational interference. When an automated copilot reads your decade of wearable logs alongside hospital notes, its probability of dropping vital anomalies increases directly with file size.
When you use consumer tools to parse your personal health records, do not paste years of laboratory sheets into one query. Split your files into distinct chronological batches, ask for specific extractions per visit, and verify that the output preserves dates, units, and drug names without redaction.
The longer view
One headline rarely tells the story. See how today’s news fits the bigger shifts on AI Trends, or learn to read your own data on How it works.