Early use of generative AI can narrow student solutions and hinder metacognition. #
Posted July 25, 2026 [ Reviewed by Kaja Perina
](/us/docs/editorial-process)
Key points
- AI adds the most value when it comes after a first attempt at thinking, not before it.
- Groups with immediate AI access consistently converged on strikingly similar solutions.
- The deeper challenge is not only timing AI correctly, but keeping students the authors of their own thinking.
Co-authored by Paul Surlis, Ava Gilmartin, and Michael Hogan.
Imagine a group of postgraduate students who have just been handed a complex water management problem. Their challenge is to develop a sustainable water management plan for a busy tourist city, balancing environmental protection, economic growth, and increasing demand for scarce resources.
Working in small teams, they have ninety minutes to understand the problem, debate possible solutions, and agree on a plan. Within minutes, several groups begin consulting ChatGPT. By the end of the session, each group submits a polished proposal. Although the groups never interacted with one another, many of their solutions are strikingly similar: in their structure, in the assumptions they make, and in the possibilities they never explore. Nothing has been copied. Yet something important has been lost.
This is the scene that lies behind a recent study by Romero (2025). As generative AI becomes an increasingly familiar part of higher education, the most important question may no longer be whether students should use AI at all. A more useful question is when AI should enter the learning process. Could introducing AI too early erode the very creativity, discussion, and cognitive effort that universities are trying to cultivate?
Testing the Impact of AI Introduction Timing #
To test whether timing really mattered, Romero (2025) designed a simple but revealing experiment. Thirty-six graduate students were randomly assigned to one of three conditions while working in small groups on the water management challenge. One group had access to ChatGPT from the very beginning. A second group spent the first fifteen minutes working without AI, debating ideas, questioning assumptions, negotiating different perspectives, and building an initial solution together, before AI was introduced later as a tool for refinement. A third group completed the task entirely without AI. Although this was a relatively small quasi-experimental study involving a single task, its design isolated one variable that is often overlooked in discussions about AI in education: timing.
The results were striking. Groups with immediate access to AI consistently converged on similar solutions, suggesting that early reliance on AI narrowed idea diversity and encouraged familiar lines of thinking rather than independent exploration. By contrast, students who first grappled with the problem themselves generated a broader range of ideas before using AI to challenge, extend, and refine their thinking. Interestingly, the groups that never used AI produced the most diverse solutions, although their work was also more variable in quality. Romero (2025) draws an important design principle from this: AI appears to be most valuable when it enters as a finishing tool rather than a starting point.
Romero's (2025) study involved just thirty-six students completing one collaborative task, meaning the findings should be viewed as early evidence rather than definitive proof. Yet the pattern is a suggestive one. By delaying AI, educators may be protecting a period of productive cognitive struggle before technology begins shaping students' thinking. That naturally raises a deeper question: why should fifteen minutes without AI make such a difference?
When AI use Leads to Metacognitive Laziness #
Fan et al. (2025) offer a compelling explanation by examining what happens inside learners' thinking when AI becomes part of the learning process.
In a randomised experiment involving 117 university students, participants completed a two-stage English reading and writing task before revising their work with one of four forms of support: ChatGPT, a human expert, a structured writing checklist, or no additional support. Researchers then used trace data to map how students regulated their own learning during the revision process - what they monitored, what they evaluated, when they d to orient themselves.
The ChatGPT condition centred students' self-regulatory process on the AI. Interactions were extensive. Essays improved. But compared with the human expert and checklist conditions, the ChatGPT group showed relatively fewer of the metacognitive processes — evaluation and orientation — that involve stepping back, assessing where you are, and deciding what to do next. Human expert support triggered transitions between orientation and evaluation that ChatGPT simply did not. And crucially, when knowledge gain and transfer were measured, not just task performance alone, the ChatGPT group’s advantage disappeared.
This is what Fan et al. call metacognitive laziness: not laziness in the everyday sense, but the subtle displacement of the self-regulatory thinking that makes learning stick. The connection to Romero is direct: reaching for AI immediately risks skipping the period of struggle and genuine cognitive effort that precedes good ideas. Romero shows what this costs in terms of creative diversity. Fan et al. show what it costs in terms of metacognitive development. Together, the studies point towards the same educational principle: giving students time to think before introducing AI may help preserve both.
Artificial IntelligenceEssential Reads Creativity researchers have long recognised that good ideas rarely appear the moment a problem is presented. Instead, they emerge through something closer to Wallas's (1926) classic four-stage sequence of preparation, incubation, insight, and verification. The incubation stage, where people continue thinking without immediately reaching for an answer, is often where unexpected connections begin to form. The delayed access condition in Romero's study can be understood as a pedagogical attempt to protect that incubation space, which might otherwise never open at all.
From pedagogical timing to personal habit #
If delaying AI helps preserve the thinking that underpins creativity and learning, the next question is how educators should design for it. Romero's study cannot tell us whether fifteen minutes is enough, or whether the same principle applies across different subjects, learning activities, or assessment formats. Those questions remain open. Timing is only one design decision. Educators must also consider which tasks genuinely benefit from AI, how assessment can reward thinking rather than polished output, and how learning activities can scaffold the metacognitive processes that AI may otherwise displace. Ultimately, the deeper challenge is designing learning environments in which students remain the authors of their own thinking. Getting the timing right is one important step. Preserving learners' agency as AI becomes a permanent feature of education may prove to be the much larger task. However, agency cannot indefinitely depend on pedagogical structures that merely delay access to AI. As students encounter AI beyond carefully designed classroom activities, they will increasingly need to regulate their own use of these tools.
This means developing practical habits of good AI use: generating ideas before prompting, critically evaluating AI suggestions, and using AI to extend rather than replace their own thinking. In this sense, agency is not simply self-regulation, but self-regulation guided by an understanding of how to sustain learning over the long term. As AI becomes a permanent feature of education, the quality of learning may increasingly depend on the opportunities students are given to think before they prompt. Yet these habits do not emerge automatically. Educators play a central role in cultivating them through thoughtful pedagogical design, helping students gradually become capable of protecting their own agency even when external constraints are no longer present.