Participants had just used AI to crack 12 problems when researchers suddenly pulled the tool away. Some were allowed to keep working unaided; others lost the crutch on the spot. The pattern was blunt: the group AI had lifted to high accuracy collapsed the moment the tool disappeared, while the never-used-AI control group held its line and even improved.

The work comes from a joint randomized controlled study across five universities — Carnegie Mellon, MIT, Oxford, UCLA, and UC Berkeley — under the title "AI Assistance Reduces Persistence and Hurts Independent Performance." The paper was presented on October 7, 2026 at the Conference on Language Modeling (COLM) and cataloged on arXiv as 2604.04721. Co-authors include Brian Christian, Grace Liu, Tsvetomira Dumbalska, Michiel A. Bakker, and Rachit Dubey.

Experimental Design: A Deliberate "Crutch Removal"

Researchers recruited participants online and split the work into several task types: some solved basic fraction problems, others tackled SAT-style reading comprehension. In every condition, an AI tool was allowed to help for the first batch of problems, then researchers removed the tool mid-session and watched what happened next.

The results were uniform. With AI on the side, users did score noticeably better than the unaided control in the moment. The moment researchers took the tool away, the AI group started flubbing answers or skipping questions almost immediately, while the control group held steady and even trended upward. AI boosted present-moment performance, but it did not train any capacity to keep solving once the crutch was gone.

The Cost of Efficiency: Offloaded Productive Struggle

Co-author Michiel Bakker, an MIT assistant professor, framed the motivation in a WIRED column: the experiment was designed to take a long-running essay about how AI may disempower humans over time and stress-test it in a controlled setting. Willingness to persist with hard problems, he noted, is the key to picking up new skills and a leading indicator of how much a person will keep learning. AI tools, the experiment shows, amplify the opposite — they make giving up more likely the moment the help vanishes.

The team is not calling for an AI ban in schools or workplaces. Bakker's exact words: AI can clearly help people perform better in the moment, and that is valuable — but we should be more careful about what kind of help AI provides, and when. In other words, an AI that defaults to direct answers and one that defaults to scaffolding, coaching, and challenging the user could have very different long-term effects. The "tutor path" is harder to commercialize, but it is the path the evidence points toward.

Commentary: The Other Edge of the Blade

Two questions dominate the discussion this study kicked off in the AI community.

First, is this "crutch-removal" effect specific to ChatGPT? The work only tested one conversational tool, but the sample size — hundreds of participants across three experiments — is statistically solid. WIRED, The New York Times Magazine, and UC Berkeley News have all surfaced similar conclusions, which makes the direction feel more like a class effect than a single-product bug.

Second, should schools ban AI tomorrow? The authors explicitly reject a blanket ban and instead push for the tutor path. That path is not easy: direct answers are a better user experience, drive higher retention, and ship faster. Asking a product to delay gratification and "teach" the user is, by Bakker's own admission, a paternalistic trade-off. Anthropic's own research has acknowledged that fully offloading coding tasks to AI erodes a programmer's baseline ability — only collaborative patterns, not pure outsourcing, leave human capability intact.

For working developers and researchers, the takeaway is a concrete yardstick: if you cannot solve a problem the moment AI is taken away, that collaboration is probably not augmenting you — it is replacing you. AI's real value is not how fast it answers for you, but whether it leaves you as someone who can still answer once it is gone.

Sources