The Unnoticed Freezes in the AI Era

Abstract

In 2023, 18% of American adults said they had ever used ChatGPT. By early 2026, that number was 44%. AI adoption in the workplace has been even more remarkable. According to Microsoft's 2024 Work Trend Index, 75% of knowledge workers were already using AI at work, with nearly half having started within the preceding six months. We did not vote to hand over drafting, summarizing, fact-checking, and even thinking to these systems. There was no big decision. Just a free chatbot, a looming deadline, and a prompt that saved twenty minutes. Then another, and another... Regarding this serious issue, I was drawn to a new paper quietly posted on SSRN last September by Professor Vuong Quan Hoang. It gives this slow handover a name: the Correction Freeze Problem (CFP). His question is simple: can a system think it is fine while it is drifting away from reality, and stay that way not because it is lazy or lying, but because the very way it corrects itself makes staying blind the rational choice? The paper answers yes. In four different ways. And when any of those ways meets a world that rewards going along, the drift becomes a trap that is much harder to escape than it was to fall into. To grasp these insights, you don't need the math. Simply because you've lived it.