Proof behind the platform: research, benchmarks, and insights

Large language models have erased the surface signals that used to tell us when information was wrong — and that's a bigger problem than the hallucinations themselves.

Scaled Cognition co-founder and CTO Dan Klein joined the Gradient Dissent podcast to explain why large language models are plausibility engines, not truth engines — why the hallucinations you catch are only a fraction of the real problem, and why building reliable AI requires rethinking the architecture from the ground up, not layering more models on top of broken ones.

Most teams think they know their hallucination rate — but they're only measuring the errors that look wrong. The harder, more common failures are the ones that look right. Here's why your real number is probably five times higher, and why that changes everything about how you should evaluate AI reliability.