When Cognitive Fatigue Breaks Health Chatbots

Psychologists have used the Stroop task since 1935 to measure how the human prefrontal cortex manages conflicting information under sustained demand. When computer scientists subjected leading large language models to this exact test, the outcome mirrored human cognitive exhaustion. While models maintained over 90% accuracy on concise queries, their performance collapsed toward near-zero percent as prompt length, sequence volume, and context depth expanded.

The mechanism is straightforward. Attention heads in transformer architectures allocate weights unevenly across extended token sequences. As the input grows longer and requires filtering irrelevant distractions, the model loses track of earlier constraints. It does not pause to reconsider; it simply generates plausible syntax with degraded reasoning.

A patient tracking autoimmune flares often maintains pages of overlapping variables: dietary changes, subjective fatigue scores, hormonal cycles, and blood markers. Expecting an off-the-shelf consumer interface to synthesize an 8,000-token medical background without dropping critical contradictions is currently unfounded. The software starts confidently, then silently drops earlier context.

Do not treat a single prompt thread as an infinite medical ledger. Break your health queries into small, discrete records: test one specific lab panel or one medication query at a time, check the citations manually, and verify the model retains the constraints you set three questions ago.