3 reasons why ‘going slow to go fast’ is the right approach for AI in schools

Increasingly, states and cities are issuing bans or time-limited moratoriums not only on cell phones in schools, but also on student screen use (including AI tools) more broadly. Utah was one of the first states to place limits on student AI and screen use through legislation passed this spring. Last week, New York City and Los Angeles made headlines by adopting year-long moratoriums on student AI use in public schools.

These jurisdictions—cutting across red and blue political lines—are not alone in putting the brakes on screens and AI use. Their actions remind me of a school district leader who told me last year that his approach to AI was to “slow down to go fast.” Before rolling out new tools, he wanted to understand how they might help students—and how they might harm them. He and his team conducted a districtwide screen time audit and reviewed teacher pedagogy to understand which approaches were “AI-aware,” “AI-assisted,” or “AI-resistant.”

Critics argue that this go-slow approach will stifle innovation, throw the baby out with the bathwater, and leave children unprepared for the future. As someone who has studied innovation in education for many years, I can say with confidence that not all innovation is good innovation. And as someone who has recently focused on how generative AI affects student learning and development, I believe the current slowdown is warranted.

There are three good reasons why a “slow down to go fast” approach makes sense right now.

#1: The risks are real

While the evidence base on AI and student learning is still developing, emerging research continues to identify important risks. In our Global Task Force on AI and Education report, we found that “narrow” AI applications, those built on vetted content and embedded within high-quality pedagogy, can offer meaningful benefits. At the same time, there are significant risks associated with “wide” AI use, particularly when young people engage extensively with general-purpose chatbots that were not designed primarily for education or child development.

These risks include emotional manipulation, reduced frequency of repairing ruptured human relationships, and perhaps most concerning for educators, students using AI to bypass effortful thinking, thereby undermining learning itself.

New evidence from one of the world’s largest educational AI initiatives suggests that even thoughtfully designed “narrow” AI systems can produce unintended consequences. Estonia, a country known for its advanced digital infrastructure and technology-forward approach to public services, launched its AI Leap program in secondary schools last year. The initiative included teacher training, AI literacy instruction, and access to AI tools. Students in grades 11 and 12 were provided access to ITI, a specially designed AI workspace initially developed by OpenAI in collaboration with Estonian education experts.

The result was a secure, education-focused version of ChatGPT designed not simply to provide answers to students but to support reasoning, reflection, and learning strategies.

After one year, the early findings appear more encouraging for teachers than for students. Teachers reported saving an average of 1.8 hours per week on administrative tasks, and an impressive 44% said the tool helped them shift their teaching approach from primarily delivering knowledge to facilitating critical thinking. At the same time, 73.4% of teachers agreed that “students were capable of producing high-quality work but cannot explain the underlying reasoning,” suggesting that students may be outsourcing too much of their thinking to AI.

Students, by contrast were not enthusiastic about the AI tool. Survey results indicated a generally neutral to lukewarm reception, and but the platform received a net promoter score of negative 27.5, which would be a red flag for any technology company, suggesting considerable room for improvement.

Their feedback points to a deeper challenge students felt and were struggling with. Many students reported turning to AI before approaching teachers with questions, leading to fewer student-teacher interactions. While students appreciated having an accessible, nonjudgmental source of support, many also worried about what this shift could mean for their classroom experience. Poignantly, students raised concerns that speed of finding answers comes at the expense of “practicing communication, collaboration, and shared problem-solving.”

Critics of recent restrictions on student-facing AI often argue that limiting access will leave young people unprepared for an AI-enabled future. But education has always rested on a simple principle: students must learn how to think before they learn how to use tools that think for or with them. Developing critical thinking, communication, collaboration, and shared problem-solving skills remains crucial for preparing students to thrive in a world increasingly shaped by AI.

#2: Common-sense exceptions for beneficial AI use are feasible

Another valid concern is that broad restrictions could inadvertently eliminate beneficial uses of AI alongside the more problematic ones.

Our task force identified a number of promising applications of narrow AI, ranging from improved assistive technologies for students with learning differences to real-time translation tools that support multilingual learners. Many of the most successful examples involve adults using AI to better support students rather than students independently using AI systems themselves. For example, AI can help less experienced tutors provide more effective learning support.

Teacher-led AI design efforts are particularly promising. One example is the national approach underway in the Netherlands, where teachers identify a problem of practice and then collaborate with researchers and technologists to develop AI tools that can help solve it. Projects include class question boards and early warning systems that help identify primary school students who may struggle to master handwriting so that timely interventions can be provided.

Importantly, slowing down does not require eliminating all AI use. Jurisdictions can pause or limit student-facing AI while preserving applications that demonstrably support teaching and learning.

This is precisely the approach many states, cities, and districts have adopted. Utah’s restrictions and New York City’s moratorium, for example, continue to allow educators to use AI for many professional purposes. Both also provide exemptions for specific educational contexts, including computer science courses, career and technical education programs, assistive technologies for students with learning differences, and support for multilingual learners.

In practice, many of these policies seek to preserve teacher judgment, student effort, and family transparency while allowing targeted uses of AI where benefits are clearer and risks are more manageable.

#3: Technology companies have lost parents’ trust

Ultimately, much of the momentum behind these restrictions comes from parents and caregivers who are concerned about the role technology is playing in their children’s lives. Across the country, many parents worry that AI could increase dependency on technology, with 83% of parents believing that children need to learn to think critically on their own without relying on AI.

These concerns were on full display when New York City education officials released draft AI guidance for public comment earlier this year. The proposal generated approximately 6,500 public comments, reflecting an unusually high level of public interest and concern. That feedback was one factor contributing to the city’s decision to adopt a one-year moratorium on student AI use. Prior to its ban, Utah had conducted an ed-tech audit to determine if tools complied with student data privacy agreements and found half did not. Los Angeles has experienced perhaps the highest profile AI flop. Two years ago, school leaders quickly pulled the plug on an AI personal assistant they were set to roll out after the company went bankrupt and a whistleblower reported student data safety concerns. Recent news outside of education—from Meta acknowledging in its recent court case that is has for years pushed harmful products to children to OpenAI accidentally unleashing a swarm of out-of-control AI agents that committed criminal offences by breaking into other companies—does little to bolster parent trust of technology companies.

Parents are voters, and policymakers rarely succeed when schools and families are moving in opposite directions. That may be the most compelling reason to adopt a “go slow to go fast” approach. Building trust and creating meaningful family-school partnerships around AI will take time, but it is essential.

Students experience technology across every part of their lives, both in and out of school. Neither schools nor families can navigate these changes alone. Rarely does anything good happen for kids when families and schools are in conflict. Decades of evidence show that student learning and well-being improve when the adults in children’s lives are working toward shared goals. The same principle should guide decisions about AI in education. Before expanding student-facing AI at scale, policymakers and educators should ensure that families are part of the conversation and that the technology genuinely serves children’s learning and development.

In education, as in so many areas of public life, slowing down may ultimately be the fastest path to getting it right.