Why AI is dangerous for kids

The question is usually asked as a straight yes or no: is AI dangerous for kids? That question has no useful answer, the same way “is sugar dangerous” has no useful answer. The honest question is narrower: what does this particular use, at this particular age, displace?

A five-year-old is not born knowing how to read, add or tie a shoelace. She builds each of those by attempting the task, getting it wrong, and trying again with slightly better information each time. That loop, attempt, fail, adjust, is not a slow or inferior way to learn. It is the only way a young brain builds the skill. An AI that supplies the finished answer before she attempts the task does not corrupt her. It just removes her turn.

The two questions that place any situation in this grid: has she mastered this specific skill yet, and is an adult in the loop.

This is not a new worry dressed up for the AI era. Developmental psychology has spent ninety years building a specific, well-tested answer to why children need to struggle, and the answer holds up.

In 1936 the Swiss psychologist Jean Piaget argued that children build understanding by acting on the world directly, not by being told the answer. He called the mental structures they build schemas, and the whole theory is that schemas form through doing, not through hearing. A young child who is handed the answer has nothing to act on, and so has nothing to build with.

Good teaching, a parent, a teacher, a well-designed hint, works entirely inside that middle ring: enough support that the child can reach the answer, not so much that the answer is handed over outright. An AI that answers a question outright does not operate in that ring at all. It reaches straight into the outer ring and hands back a solved problem, and the middle ring, the part supposed to do the growing, never gets used.

There is a second, separate piece of research that explains why this matters even for skills a child could technically be told directly. In 1994 the UCLA psychologist Robert Bjork coined the term desirable difficulty: certain kinds of effort during learning feel slow and inefficient in the moment but produce dramatically stronger long-term memory than being given the answer smoothly. Removing the difficulty does not make learning faster. It makes it look faster while quietly producing less of it.

Two studies, thirteen years apart, measured this directly rather than theorising about it.

In 2011 the psychologist Betsy Sparrow and colleagues published a study in Science documenting what is now called the “Google effect.” When people expect information to remain available through a search engine, they remember where to find it far better than they remember the information itself. Researchers call this transactive memory: offloading the fact to the tool and keeping only the address. Efficient for an adult who already knows how to think through a problem. For a child who has not built that skill yet, the address becomes the whole skill.

Figure 2: illustrative, not a real scan. The pattern shown (fewer engaged connections when a task is handed to an AI) is what a 2025 MIT Media Lab pilot study measured. That study was run on adult writers, not five-year-olds. Formal research on young children specifically is still catching up.

The 2025 study is the sharper of the two. Researchers at the MIT Media Lab, led by Nataliya Kosmyna, fitted adult writers with EEG headsets and compared three groups: one writing essays unaided, one using a search engine, and one using an AI assistant. The AI-assisted group showed measurably weaker neural connectivity while writing, and afterward struggled to quote or explain their own essays compared with the unaided group. The researchers called the effect cognitive debt: the cost of offloading the thinking is not visible on the page that gets produced. It shows up later, the next time that same person has to think without help.

These are the actual terms researchers use for what parents sense but often cannot name. Knowing them turns a vague fear into a specific, checkable question.

Nuance is not the same as permissiveness. Some uses cost a child almost nothing. Others cost her the exact skill the activity was supposed to build.

A tool becomes a parenting decision, not a technical one, the moment it changes what a child is asked to do for herself. AI crossed that line the day it became free, instant and sitting in the same device she already uses for cartoons.

Three consequences reach you directly.

None of that requires banning a device or becoming an expert in child psychology. It requires knowing which parts of her day are still building her mind, and protecting the attempt itself in exactly those parts.