A participant works through a row of fraction problems with an AI assistant beside them. The assistant already knows each solution. Ask “answer?” and it obliges. Then, without warning, the box disappears.
That small act of experimental mischief produced one of the year’s more alarming AI headlines: ten minutes of AI use can make people less persistent. The underlying study of 1,222 participants deserves attention. It also deserves a more precise question than “Is AI making us lazy?”
What sort of help did people receive, and what did that help ask them to do?
The answer machine worked, until it left
Across three randomized experiments, participants solved fraction or reading-comprehension problems. One group had an AI assistant during the practice phase; a control group did not. Everyone then faced three problems without AI.
The first fraction experiment showed the sharpest drop. After exclusions, the AI group solved 57% of the final problems, compared with 74% for the control group, and skipped more often. The researchers improved the design and ran it again. In that replication, the gap narrowed to 71% versus 77%. The difference in correct answers remained statistically significant, but the aggregate skip-rate difference did not.
A third experiment used SAT-style reading questions. The former AI group solved 76% of the final questions, compared with 89% for the control group, and skipped more of them.
Those are real short-term effects under the study’s conditions. They are not evidence that ten minutes with any AI causes a general loss of grit. The tasks lasted about 13 to 15 minutes, the final test contained three questions, and the researchers measured what happened immediately after assistance vanished. The paper itself warns against treating the results as proof that AI lowers persistence for every task.
“Help” covered two very different behaviors
The revealing part came when the researchers looked inside the AI group in the second experiment. Sixty-one percent said they mainly used the assistant to obtain direct answers. Twenty-seven percent used it for hints or clarification, and 12% reported not using it.
The direct-answer group did worst on the unaided test. People who asked for hints performed about as well as the control group, while those who ignored the assistant did better. The comparison among those self-selected subgroups is observational, not randomized, so it cannot prove that asking for answers caused the decline. Perhaps people inclined to request a full solution differed in some unmeasured way. Still, it changes the practical reading of the experiment.
“AI assistance” was not one treatment. For some participants it was a worked-example machine; for others it was a hint desk; for most it was a vending machine for completed fractions. Putting all three behind one label hides the most useful distinction.
Other experiments point the other way
If brief AI contact simply damaged learning, other randomized studies should lean in the same direction. They do not.
In two large experiments on cover-letter writing, participants first learned five writing principles. Some then practiced with an AI tool that rewrote a letter according to those principles; others practiced alone. On a later no-AI test, the people who had practiced with AI wrote better letters. The advantage remained in a test the next day. A follow-up condition found that merely seeing a strong AI-generated example produced a similar gain.
A newer study assigned 2,383 students to AI tutoring, human tutoring, or a control condition before fresh GRE-style tests. Pooled across the tested systems, AI tutoring produced learning gains statistically equivalent to human tutoring under the researchers’ chosen margin. The study covered one hour of guided teaching, not a button that supplied the current answer, and its results concern GRE questions rather than every kind of learning.
These papers do not cancel one another. They tested different tasks, interfaces, time spans, and definitions of help. That is exactly the point. “Uses AI” is about as informative as “uses a book.” A book can be an answer key, a textbook, or something placed under a wobbly table. The object matters less than the role it is given.
A better question for the help box
The persistence study has a useful warning inside it: a system optimized to finish the current problem may deprive its user of practice on the next one. Immediate success and durable ability are separate outcomes.
That does not mean an assistant should become theatrically unhelpful. Nobody needs a chatbot replying to a deadline with a riddle about personal growth. It means the interface should know whether the user wants completion or learning. For learning, it can ask for an attempted step, offer a hint before a solution, explain a worked example, and later remove the scaffolding on purpose.
The frightening version of the story is that ten minutes of AI weakened people. The more useful version is narrower: when help made answers nearly effortless, some people struggled after the answers stopped. Other experiments found gains when AI supplied examples or behaved like a tutor.
The important setting may not be how much intelligence sits inside the box. It may be whether the box completes the work or helps the person become able to complete it next time.