A short test involving fractions and reading comprehension has produced one of the more unsettling findings in recent AI research. Not because it reveals something dramatic, but because it reveals something quiet: brief exposure to an AI assistant appears to reduce a person’s willingness to keep trying once that assistant is gone.
What the Experiment Actually Showed
Researchers led by Brian Christian of the University of California Berkeley designed a controlled study and presented it at the third annual Conference on Language Modeling. The setup was straightforward. Adult participants were asked to solve equations involving fractions. One group had access to a ChatGPT window throughout; the other worked without assistance. Then, partway through the test, the chatbot was removed from the first group entirely.
The results were clear. While ChatGPT was available, the group using it solved more problems and skipped fewer questions than the control group. That part is expected. The revealing moment came after the tool disappeared. The two groups essentially switched positions: those who had relied on the AI began skipping more questions and solving fewer, while the group that had worked alone maintained its performance.
The researchers then repeated the experiment using reading comprehension questions drawn from SAT exams. The same pattern emerged. Accuracy and persistence dropped in the AI-assisted group once the tool was removed.
Christian’s own summary of the finding is worth noting. He described it as a situation where systems built to be helpful end up being helpful “in ways that are kind of unhelpful.” That tension sits at the center of what this research is pointing toward.
The “Boiling Frog” Problem in Education
The study’s implications are sharpest when applied to younger learners, and the researchers themselves use a specific phrase to describe the risk: a “boiling frog” problem. The concern is not that a single test session causes lasting damage. It is that sustained AI use, over months and years, could gradually erode the motivation and persistence that make long-term learning possible. By the time those effects become visible, the researchers warn, they may be difficult to reverse.
This framing connects directly to a policy decision already underway. New York City schools restricted AI access for roughly 600,000 students from pre-K through eighth grade. The rationale offered by school chancellor Kamar Samuels was explicit: children learn best through productive struggle, through getting things wrong and trying again. No AI tool, in his view, can replicate that process.
What the Berkeley study adds to that argument is empirical weight. The concern is no longer purely philosophical. Even in short test sessions measured in minutes, the pattern of reduced persistence appeared. The question the research raises is not whether AI is useful. It clearly is, at least in the moment. The question is what happens to the cognitive habits that form through difficulty, repetition, and failure, when those experiences are consistently bypassed.
Why This Matters Beyond the Classroom
The study focuses on students and test conditions, but the implications extend further. The researchers themselves gesture toward this when they raise the question of AI agents designed to remove friction from everyday tasks, not just academic ones. If the productive struggle that builds persistence is valuable in learning, it is reasonable to ask whether it is also valuable in professional problem-solving, creative work, and decision-making under uncertainty.
This is what most coverage of AI capability tends to miss. The conversation around AI tools is almost entirely focused on output quality: does the model give a correct answer, does it write a coherent paragraph, does it solve the equation. The Berkeley study shifts the frame. It asks what happens to the person using the tool, not just in the moment, but afterward. Performance during AI-assisted tasks looked better. Performance after the tool was removed looked worse.
That asymmetry matters for anyone thinking about how AI fits into workflows, education systems, or individual habits. A tool that improves short-term output while reducing the underlying capacity that generates output is not straightforwardly beneficial. It is a trade-off, and one that tends to be invisible until the tool is no longer available.
The study does not argue that AI assistance should be avoided. It argues that the effects of that assistance on human cognition deserve serious attention, and that those effects may accumulate in ways that are slow to appear and hard to undo.
In Short
A controlled study presented at the Conference on Language Modeling found that adults who used ChatGPT during a test became less persistent and less accurate once the tool was removed, compared to those who had worked without it. The effect appeared in both math and reading comprehension tasks. Researchers describe the risk for younger learners as a “boiling frog” problem: gradual erosion of the motivation and persistence that drive learning, accumulating over years before becoming visible. The core issue is not whether AI produces good outputs in the moment. It is what repeated reliance on those outputs does to the human capacity that remains when the tool is gone.
Based on reporting from Nautilus.