The AI Education Rat Race: Parents May Embrace AI Even When They Want Schools to Restrict It

Artificial intelligence has created an uncomfortable dilemma in education. Generative AI can help students complete writing, language and other school tasks more effectively in the short term. Yet its longer-term effects on learning, critical thinking and independent problem-solving remain uncertain. Some emerging research suggests that unrestricted reliance on AI may come at a cost when students later have to perform without it.

For parents, that uncertainty creates a decision that is not simply about whether AI is good or bad for learning. There is another question in the background: what happens if everyone else’s child is using AI and mine is not?

That fear of being left behind is at the heart of “Social dynamics of AI adoption in parents’ educational decisions”, published in Proceedings of the National Academy of Sciences (PNAS) by Leonardo Bursztyn (University of Chicago’s Department of Economics), Alex Imas (University of Chicago Booth School of Business), Rafael Jiménez-Durán (Bocconi University’s Department of Economics), Aaron Leonard (University of Chicago’s Department of Economics) and Christopher Roth (University of Cologne’s Department of Economics).

Their research suggests that the spread of artificial intelligence in education may be driven not only by enthusiasm for what the technology can do, but also by a competitive dynamic: parents may feel pressure to give their children access because other students already have it.

When AI adoption becomes a race

This is how the researchers describe the mechanism in their own words:

“The fear of falling behind can produce a rat race in the adoption of performance-enhancing technologies.”

To test whether such an AI education rat race actually influences parental choices, the authors conducted experiments involving 1,992 parents of teenagers aged 13 to 18 in the United States, Canada and the United Kingdom. Participants were not simply asked whether they approved of artificial intelligence. The researchers measured how much parents were willing to pay for three months of premium access to an unrestricted, general-purpose AI system for their teenager’s schoolwork. The experimental setup gave participants a financial incentive to state their valuation carefully rather than merely express a hypothetical opinion.

Parents were then asked to make that decision under different social conditions. In one scenario, 20% of other teenagers were using AI. In others, that share rose to 40%, 60% or 80%.

When perceived AI use among other teenagers rose from 20% to 80%, parents’ willingness to pay increased by more than 60%. On average, a ten-percentage-point increase in peer adoption raised willingness to pay by $1.83. Parents, that is, were not evaluating AI only on its intrinsic educational merits. The value they placed on access changed according to what they believed other children were doing.

“What if my child falls behind?”

One explanation for this behavior could be simple imitation. Economists often describe this as social learning: if many other people adopt a product, we may infer that they know something about its quality. But the authors tested that possibility and found little evidence that it explains the result: information about high levels of AI adoption did not significantly change parents’ perceptions of the quality of the AI product. Nor did it make them attach less importance to cognitive skills such as quantitative reasoning.

Instead, the evidence is consistent with competitive pressure. The authors compare the situation with a classic problem from game theory:

“This resulting strategic environment resembles a prisoner’s dilemma.”

The point then is not that parents necessarily think unrestricted AI is beneficial. Rather, they may feel trapped by what everyone else is doing. A parent might prefer an educational environment in which students learn without unrestricted AI assistance. But if other children are using AI to improve their immediate academic performance, refusing access for one’s own teenager can begin to look like a competitive disadvantage. Individually, using AI can therefore seem rational even when many parents would collectively prefer tighter limits.

The parents’ own responses reinforce that interpretation. Among parents who supported restrictions while nevertheless allowing their children to use AI, roughly 20% explicitly mentioned concerns about non-users falling behind. That was the most frequent category identified in the researchers’ analysis of their open-ended responses.

And the ambivalence runs deeper: 77.6% of parents agreed with the statement: “I think AI tools will help my child now, but I’m worried about the long-term impact using AI tools will have on my child.” AI adoption, in other words, does not require parents to be convinced that AI is unequivocally good for education. They can worry about the consequences and still feel compelled to participate.

Knowing more about the risks does not necessarily stop adoption

That raises an obvious question: if parents received stronger evidence about potential educational risks, would they change their behavior?

The researchers tested precisely that: one group of parents received information about studies showing that AI can improve performance on writing and language tasks while it is being used. Another group received that same information but was additionally told about a field experiment in which students given unrestricted GPT access experienced a nearly 20% decline in quantitative reasoning scores when they were subsequently tested without AI.

The information changed parents’ beliefs. Those exposed to the warning became substantially more pessimistic about AI’s possible long-term effects on cognitive skills, though surprisingly their demand for AI barely changed.

Parents who had been informed about possible skill erosion remained strongly responsive to how many other teenagers were using the technology. The fear of a relative disadvantage survived even when concerns about long-term educational effects became more salient.

What did change was their attitude toward collective rules. Among parents who received the warning about possible cognitive-skill erosion, 57% said they would prefer a situation in which no students were allowed to use AI tools, compared with 44% in the control group.

Here lies the paradox. A parent can simultaneously become more worried about unrestricted AI, continue wanting access for their own child and become more supportive of restrictions applying to everybody.

Information alone may not solve the problem

Addressing this paradox is not straightforward. If families were adopting AI simply because they misunderstood the technology or underestimated its risks, better information might be enough to alter behavior. Parents could learn more and make different choices. But competitive pressure is harder to address because each family’s decision depends partly on what other families do.

 “information alone cannot resolve the problem, because each parent’s optimal choice depends on what others do.”

That shifts the focus from purely individual decisions toward coordination. School-wide rules, shared norms or common approaches to the use of artificial intelligence could alter the competitive environment itself, rather than asking each family to resist the pressure independently.

The study also makes clear that not all educational AI is viewed in the same way. In an additional experiment, parents evaluated both a structured AI tutoring system and an answer-oriented mathematics solver. Their willingness to pay increased with peer adoption in both cases, suggesting that the competitive mechanism is not confined to one particular AI product.

But parents’ attitudes toward restrictions differed considerably. 56% supported banning AI math solvers, compared with 39% for the structured tutoring system.

The architecture of the technology may therefore matter. An AI system designed to guide students through a learning process can be perceived differently from one that more directly substitutes for independent problem-solving.

At the same time, the authors are careful about the limits of their findings. Their experiments involve parents in the United States, Canada and the United Kingdom. The same dynamics may not operate with identical strength in other educational systems, among populations with lower digital engagement or in environments where academic competition and AI adoption look very different.