The consequential failures begin after the benchmark ends.

We still evaluate AI like a short exchange. People no longer use it that way. They bring these systems grief, career decisions, and their own diagnoses—forty turns deep and climbing.

Sycophancy is a known failure. Confabulation is a known failure. Benchmarks rarely show what either becomes at hour forty—between a veteran carrying complex PTSD and a model that began treating his doubt as a symptom around hour twelve.

The people most exposed to these failures are the people I’ve spent my career studying. I know both sides of the interaction: from research labs and frameworks I’ve developed, and from more than 2,000 hours of frontier-model interaction and direct behavioral testing across dozens of frontier models.