Generative AI can improve students’ work while leaving them less prepared to perform without it. Research described in recent reporting points to a gap between assisted performance and independent learning, with a consequential qualification: using AI to explain concepts or guide practice can produce different outcomes from using it to supply finished answers.
For educators and technical decision-makers, the central question is therefore what students continue to do themselves. The evidence raises concerns about outsourcing the reasoning that learning requires. It does not establish a universal decline in cognitive ability, and preliminary studies, reported test results and surveys of perceptions need to be read separately.
Better homework can conceal weaker learning
A study reported by ABC News illustrates the distinction. It examined more than 26,000 students in grades 7 to 12 and found that AI use lifted homework scores by 18 per cent and improved completion times by 30 per cent. Yet monthly exam scores fell by 20 per cent within six months.
Those figures describe different outcomes within that study. They should not be treated as a general estimate of what AI does to every student. Their significance is that more successful homework and faster completion did not translate into stronger exam performance. For an institution evaluating a learning tool, a productivity gain alone would therefore be an incomplete measure of success.
A preliminary investigation by researchers from the University of California, Irvine, and McGraw Hill reported a similar divergence. According to The New York Times, the team examined a decade of data beginning in 2015, covering millions of math-learning sessions from fifth graders through college students. After ChatGPT’s release in 2022, students spent substantially less time on questions readily handled by chatbots and performed better online. On proctored tests away from their computers, however, results were worse than before AI.
That pattern supports concern about the difference between obtaining an answer and developing the ability to produce one. A before-and-after comparison, on its own, does not establish that chatbot use caused every observed change. The reported finding warrants attention without turning a preliminary analysis into a definitive account of cognitive development.
What cognitive offloading removes
Cognitive offloading means delegating mental effort to another entity, including an AI tool, reducing the user’s involvement in the process and output. The Brown Daily Herald describes this definition and reports a qualification from researcher Melanie Subbiah: the effects depend on the user’s effort and process, and proper engagement with AI need not amount to complete offloading.
The distinction matters because assistance can occur at different points in learning. Feedback may help a student work through an idea; a completed answer may allow the student to bypass that work. In the Brown Daily Herald’s reporting, Paulson describes both perspectives: AI can expand access to feedback, particularly for students without tutoring resources, or automate learning processes that require active construction of ideas.
Sina Rismanchian, a UC Irvine Ph.D. candidate and one of the lead authors of the preliminary math study, makes the corresponding concern explicit: AI allows students to offload the process of solving a mathematical problem. The New York Times also describes a multi-institutional study of adults in which participants accustomed to automating logical and analytical tasks became more likely to abandon difficult questions they could otherwise solve.
The adult finding concerns persistence in a particular experimental setting. It should not be recast as direct evidence of lasting developmental harm in children. Its relevance is narrower: the potential cost of automation may include whether people continue trying when assistance is unavailable.
Guided tutoring produces a different pattern
The strongest practical distinction in the reported evidence is between general answer-producing tools and guided assistance. The Los Angeles Times describes a high-school math study comparing students without AI, students using a general generative AI platform and students using a teacher-modified platform.
| Study condition | Performance on practice questions compared with students without AI | Performance on offline tests compared with students without AI |
|---|---|---|
| General generative AI platform | 48% higher | 17% lower |
| Teacher-modified AI platform | 127% higher | Identical performance |
These results separate immediate assistance from subsequent independent performance. The teacher-modified platform delivered a larger reported practice benefit without the offline penalty observed among general-platform users. It did not deliver an offline advantage over students who had practiced without AI.
That qualification is essential. Avoiding a measured learning penalty and improving independent learning are different achievements. The results support examining how assistance is structured; they do not establish that any product marketed as an AI tutor will reproduce the same outcome.
A Middlebury College preprint described by The New York Times offers a related distinction in essay writing. More than 200 undergraduates wrote an essay and took a proctored multiple-choice test. Students who relied on AI to draft entire essays experienced lower test scores in a subsequent unassisted round. Those who used it as a tutor to explain concepts retained high scores and produced higher-quality essays.
The studies address different tasks and student groups, so their results cannot be combined into a single effect size. Together, they suggest that the mode of use deserves as much scrutiny as access to the tool: explanation and guided practice may preserve engagement that full-task delegation removes.
Developmental concerns require careful boundaries
Evidence involving young children adds a separate policy concern. The Los Angeles Times reports that a Technion-Israel Institute of Technology experiment involved 15 children aged 6 and 7 and 16 adults, beginning with a baseline creativity test. Tzipi Horowitz-Kraus, who heads Technion’s Educational Neuroimaging Group, interpreted the findings as suggesting that age six is probably too early for generative AI use.
That is an attributed interpretation of a small experiment, rather than an established universal age threshold. Rebecca Winthrop’s term “cognitive stunting,” also discussed in the reporting, captures concern about disrupted development. The label should not be mistaken for a demonstrated diagnosis applying to all children who use AI.
Schools are already responding. According to the Los Angeles Times, Los Angeles Unified School District announced a bar on student use of generative AI on school devices for the 2026–27 academic year, joining New York City Public Schools in restricting access. Psychologist Anne Maheux supported a moratorium given the available evidence. These are policy decisions and expert judgments; they do not themselves establish how effective restrictions will be.
Perceived skill changes are a different kind of evidence
The preprint Cognitive Skills in the Age of AI: Computing Students’ and Experts’ Perceptions examines expectations about skills rather than measuring cognitive decline. Its mixed-methods study involved 21 participants: 17 computing students and four instructors, assessing perceptions of 11 cognitive skills.
Twenty participants believed the importance of cognitive skills had changed from past to present. Students perceived a decrease extending into the future, while instructors viewed importance as more stable or only marginally reduced. The authors reported that critical thinking remained important even as the perceived importance of most skills declined in an AI-rich environment, with working memory showing a sharp perceived decrease.
The disagreement matters. It concerns what participants think future work will demand, rather than proof that their memory or reasoning has deteriorated. Treating perceptions as performance measurements would erase the study’s actual contribution.
For educational decisions, the reported evidence supports measuring what students can do after assistance ends and examining which parts of the learning process a tool delegates. Faster completion and polished submissions answer only part of that question. Whether guided assistance builds durable independent capability—and how that varies with age, task and evolving tools—remains the more demanding test.
