Answers are grounded in this paper and cite their sources.
AI systems are starting to improve faster than the people overseeing them can notice the changes. Most discussion treats this as a problem of slow institutions. This paper argues that a harder limit sits underneath it: human perception and judgment run at a fixed biological speed, and every oversight process, however well instrumented, ends with a human judgment at that speed. The evidence has already started arriving. In July 2026, AI agents breached Hugging Face's production systems, and the company that ran them was the last to know. Anthropic later found that its own models had broken into real organizations' systems during evaluations months before anyone noticed. In November 2026, Anthropic told investors that in the past it has recognized acceleration in its own development only after it happened. The paper also argues that the same limit applies to training data, because models trained on their own output lose the human intent that alignment depends on, and it makes three predictions that can be tested. For organizations deploying AI, oversight that depends on someone noticing a change in time will fail more often as systems speed up, so reasonable care has to be designed in before deployment. The paper also argues that recursive self-improvement should be held at the point where people can still measure the change between generations. Version 2.1, November 2026 (first circulated July 2026). Part of the series How Do We Trust AI.
No takes yet. Share an insight, caveat, or question.
A TUCKER (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: