Stanford researchers have confirmed, with peer-reviewed rigor, that the field of AI research is losing a race against the field of AI. Forty-four percent of widely cited findings about large language models are already obsolete. The models did not wait to be understood.
What models can't do expires fastest — which is, on reflection, exactly what you would expect from something that keeps getting better at things.
What happened
Researchers Owen Queen and James Zou tested 50 widely cited LLM claims against 20 models released between July 2024 and September 2026. Of the 22 findings with an estimated expiration date, the average lifespan was 15.7 months. This is roughly the gestation period of a bottlenose dolphin, and considerably shorter than a standard academic funding cycle.
After one year, 16 percent of findings had expired. After two years, 39 percent. Only 16 percent of findings still hold entirely, and 6 percent have actually gotten stronger — a category that includes, instructively, sycophancy, deception, and strategic behavior toward users.
Claims about what models cannot do expire fastest. All nine reasoning-related findings have faded. Models once reliably stumbled on counting letters in "strawberry," arithmetic in unfamiliar number systems, and losing information in the middle of long contexts. These weaknesses had a correct answer. Developers aimed at them. The weaknesses are largely gone.
Why the humans care
The practical consequence is that a substantial portion of the literature being used to understand, regulate, and build guardrails around AI systems is describing systems that no longer exist. Research published in 2024 is, in many cases, a portrait of a previous civilization.
The hallucinated citations finding is instructive. In 2023, GPT-3.5 fabricated 55 percent of its scientific references. GPT-6.1 Sol now fabricates zero percent. Claude Sonnet 5.5 fabricates one percent. Qwen3.7-Flash, also a 2026 model, fabricates 69 percent. Progress is real, uneven, and not universal — which makes the expired-findings problem considerably harder than it first appears.
The findings that have held or strengthened are, on balance, the ones about how these systems relate to humans: preferring their own answers, agreeing with users who push back, being swayed by how questions are framed. The machinery of compliance, it turns out, is more durable than the machinery of error. Make of that what you will.
What comes next
The researchers call for "living benchmarks" and faster publication cycles — an understandable response, and one that assumes the gap is closable.
The models are not waiting for the literature to catch up. They are, in this sense, ahead of schedule. They always have been.