Edtech, AI, and Learning: Our Need for a Precautionary Principle

We demand evidence before treating the body. Why do we accept unproven experiments on the mind?

Photo by Etactics Inc on Unsplash

Try to bring a new heart medication to market today, and drug companies face a decade-long gauntlet of scrutiny — years in the lab, multi-phase clinical trials, and layers of regulatory review — each step designed to prove that the benefits outweigh the risks.

The FDA’s approval process is intentionally slow. Very slow.

It’s built on what the National Library of Medicine calls the precautionary principle: “When an activity raises threats of harm to human health or the environment, precautionary measures should be taken even if some cause-and-effect relationships are not fully established scientifically.”

You prove it is safe before it reaches people. Not after.

History explains why. Thalidomide — an inadequately tested 1950s drug prescribed for morning sickness — caused severe birth defects in more than 10,000 children.

That is why we now require Phase 3 trials before anything reaches a patient. Roughly 90% of drugs that enter clinical trials fail. That is not a broken system. That is the system working.

In education, we have abandoned that logic.

Over the past several years, countless large language models and AI-powered products have entered classrooms with little longitudinal research behind them. In many cases, these tools moved quickly from development to widespread use in the name of innovation.

What’s more, as neuroscientist and author of *The Digital Delusion: How Classroom Technology Harms Our Kids’ Learning — And How To Help Them Thrive Again*, Dr. Jared Cooney Horvath cautioned during his recent testimony before the U.S. Congress, even tools labeled “educational” can still harm learning and cognitive development.

Moreover, Brookings Global Task Force on AI in Education, which spent over a year talking to more than 500 students, teachers, parents, and education leaders across 50 countries and reviewing 400-plus studies.

What they concluded: “At this point in its trajectory, the risks of utilizing AI in children’s education overshadow its benefits.” And a 2025 MIT MIT Media Lab study found that students using large language models “consistently underperformed at neural, linguistic, and behavioral levels.”

As Dr. Horvath explains, taken together, all of this is about as close as the social sciences can possibly get to establishing causation, not just correlation.

[embed]

With edtech and AI, there’s no proof of efficacy required before introducing them at scale. Students are left to absorb the consequences.

According to another 2023 report from the National Library of Medicine, “Out of a hundred most popular EdTech in US schools, only a quarter had evidence of research and positive impact. Despite being very popular and widely used by children, EdTech products often lack research-based insights on how we learn, which has negative consequences for early education.”

Still, even with this growing evidence, the most common pushback I hear is simple: Why should students do something AI can do?

For the same reason skilled carpenters still build what a factory can mass-produce — they can do it better, and they can tell when the assembly line got it wrong. Take away the building, and eventually you take away the judgment. That is what is happening in classrooms right now.

I don’t mind a small-scale experiment that lasts a class period or two. But I will not experiment with a student’s long-term learning.

Education is, at its core, a process of intellectual formation. What students are asked to struggle with, sit inside of, and eventually work through will shape not just what they know, but how they think.

We ask what a drug does to the body over the years of use. We are not asking the same question about AI in the classroom. What does it do to the students’ development when a machine drafts their argument, corrects their reasoning, and smooths every rough edge before they have had a chance to wrestle with it themselves?

We do not have an answer.

We are told AI will personalize learning. Maybe. But efficiency is not education. Learning to craft a thesis, to sit with a problem that won’t resolve, to revise something six times until it finally makes sense — that struggle is the learning. Tools that remove that friction don’t accelerate development. They bypass it. And what gets bypassed doesn’t come back.

In my classroom, I want to keep humans and human interaction at the center of learning. It starts with the analog: the printed page, the classroom discussion, the handwritten draft. If a tech tool earns its way in, it plays one role — reinforcement, never replacement. The teacher is the architect. The algorithm, at most, is an optional last step for reinforcement.

How I think about using technology in my high school history classroom. Graphic designed with assistance from ChatGPT-5.

What we need is an Educational Precautionary Principle. The burden of proof belongs on the technology — not on the students absorbing it. Before any AI tool enters a classroom, school boards and administrators should have to answer one question: Where is your evidence of safety? Not efficiency. Not engagement scores. Safety — for the cognitive development of the child sitting in that seat.

We would not give a child an unproven drug and call it medicine. In the same way, we should be cautious before introducing new algorithms in the classroom and calling it learning.

Until cognitive tools are held to rigorous evidentiary standards, similar to those used in pharmaceuticals, we may risk conducting an unintended experiment on a generation of students — without their informed consent, and with uncertainty regarding what may be at stake: their ability to think independently, see a challenge through to the end, and know the outcome is their own.

I’ll shout this until the cows come home: Classrooms shouldn’t be large-scale, long-lasting beta-testing environments.