Psychology attention test degrades leading AI model accuracy below 10%

A classic continuous attention test demonstrates catastrophic failure modes in leading AI models, proving that long-context conversational tools cannot safely interpret lengthy patient charts.

By Sabin · Wellness & AI3 min read
AI News
Psychology attention test degrades leading AI model accuracy below 10%

When subjected to a classic psychology color-naming attention battery across extended token sequences, top-tier AI models suffered severe performance decay. Models that completed short-sequence color recognition tasks at over 90% accuracy degraded to near 0% accuracy as the sequence length increased. The underlying transformer architecture lost tracking coherence when forced to maintain continuous, interference-heavy attention across an unbroken prompt.

This degradation highlights a significant vulnerability for autonomous clinical processing. Healthcare enterprises routinely propose feeding decades of unstructured electronic health records, continuous glucose monitor streams, and telemetry exports into conversational AI windows to extract diagnostic summaries. If an architecture cannot sustain basic continuous attention across a uniform cognitive test, it cannot be trusted to spot contraindications in a 200-page medical history.

If you feed multi-year laboratory exports or personal wellness journals into consumer language models, do not paste complete archives at once. Divide your data into bounded, single-topic segments of fewer than three pages, and demand clear citations back to the source text for every summary sentence generated.

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 →