When Cognitive Tests Break Large Language Models

Psychologists have used conflict-based attention tests for decades to measure human cognitive fatigue. When researchers subjected frontier language models to these sequential color-naming tasks, performance did not decline gracefully. On short lists, the systems maintained accuracy above 90%. As the sequences lengthened and contextual interference mounted, accuracy plunged below 10% in several benchmarked architectures.

The mechanism resembles cognitive interference in humans, but the systemic risk lies elsewhere. Large language models do not suffer from tired eyes; they suffer from degradation in their self-attention layers when long-range tokens compete for representational weight. In benchmark tasks where instructions conflict with entrenched prior patterns, long sequences dilute the system's ability to hold the governing rule in focus.

If you feed a five-year laboratory history into a general-purpose model to spot subtle trends, do not expect it to retain instructions given at the start of your prompt. Break your health exports into single panels, verify cross-year comparisons by hand, and treat narrative summaries as rough indexes rather than verified medical charts.