One thousand and thirty-two Chinese undergraduates, surveyed across universities in eastern and central China. Two kinds of AI reliance, measured apart from each other.
One of them cost three and a half times what the other did.
The researchers split what happens at a chat window into two behaviours. Tool dependence is handing the model the labour: draft this, fix the grammar, build the deck. Cognitive dependence is handing it the judgement: what should I make of this, what do I do next.
Both were scored against cognitive inertia, the habit of stopping at the first plausible answer. Tool dependence came in at β = 0.161. Cognitive dependence came in at β = 0.570.
The category of task moved the number, and the hours logged did not.
The category of task moved the number, and the hours logged did not.
Cognitive inertia then fed forward into lower innovation capability, at β = −0.111. Small, and consistent, and pointed at the one faculty a solo founder is selling.
At MIT’s Media Lab, 54 people wrote essays across four sessions wearing EEG caps. One group used ChatGPT, one used a search engine, one used nothing. Brain connectivity scaled down with the amount of help: brain-only strongest and most distributed, search engine in the middle, the model group weakest across the networks tied to memory and attention.
Session four is the one that should hold you.
Eighteen participants swapped conditions. The people moved off the model and made to write alone showed reduced alpha and beta connectivity, under-engaged at the task they were now doing unaided. The people who had worked alone first and got the model second recalled more of their own writing and lit up the prefrontal and occipito-parietal areas the other group had gone quiet in.
The under-engagement had followed them out of the tool.
The under-engagement had followed them out of the tool.
Same model, same task, same essays. What separated the two groups was the order they did it in, and the order is the whole finding.
That is the protocol, and its cost is not subtle. You pay one unaided draft, before the window opens: the position, the rough decision, the ugly version of the memo. It is slower than opening the window first, and the artifact you produce is worse than the one you would have had.
You are buying the connectivity with the bad draft.
Microsoft Research and Carnegie Mellon found the same shape from the other side, surveying 319 knowledge workers on 936 real tasks. Confidence in the model predicted less critical thinking. Confidence in your own read of the task predicted more.
Which means the thing being eroded is upstream of any output you could inspect.
Grade the evidence before you spend anything on it. Michael Gerlich’s work, 666 adults in the United Kingdom, is correlational: AI use against critical thinking at r = −0.68, offloading against critical thinking at r = −0.75. Correlations that size across that sample are a strong signal and not a mechanism, and the MIT cohort is 54 people over four months.
Level 2 evidence, converging from three directions, on a question nobody has run a decade-long trial on.
Fewer than 15 percent of businesses in the Philippines were using AI at all as of 2025, and the firms that do reach for chatbots and writing assistants rather than operations. Most of Southeast Asia is still standing in the cheap half of this, paying 0.161 for its typing.
The expensive half is the one that arrives quietly, on the morning you notice you have stopped forming the view before you ask for it.[^1]