When Long Clinical Histories Break Large Language Models

Relying on generative models to parse multi-year medical files risks dangerous oversights as cognitive attention fails under complex, extended review.

By Sabin · Wellness & AI3 min read
AI News
When Long Clinical Histories Break Large Language Models

Psychologists have used the Stroop test since 1935 to measure how human attention handles conflicting signals. When computer scientists recently administered modified versions of this selective-attention task to leading large language models, the software initially succeeded. Across brief sequences of color-word associations, top models maintained accuracy rates above 90 percent. As researchers lengthened the prompts and introduced more competing distractors, accuracy collapsed toward zero.

The failure points to a structural limitation in transformer architectures: processing tokens across vast context windows does not equate to sustained executive attention. The software routinely drops crucial details when forced to navigate dense, contradictory background noise. In daily consumer software, an overlooked sentence causes mild annoyance. Applied to personal health, that identical blind spot introduces silent diagnostic risk.

If you feed years of lab panels, wearable exports, or specialist letters into a general chat interface, keep prompts deliberately segmented. Ask for targeted comparisons across specific six-month intervals instead of requesting broad syntheses of your entire history. You remain the primary auditor of your record.

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.

Keep reading

Based on what you've been reading — always learning.

See all →