Orbrya
All posts
Orbrya2026-03-12

Why Punishing AI Use Backfires (And What to Do Instead)

Punishing a child for using AI doesn't rebuild the thinking skills they skipped. Here's what the research suggests actually works instead.

The moment lands differently for every parent. For some it is a teacher's email. For others it is a suspicious essay that does not sound like their child. For Renata, it was simpler: her thirteen-year-old left a browser tab open. Three paragraphs of a history report, sitting in ChatGPT, ready to copy.

Her first instinct was to ground him. Her second was to take his laptop for a week. She did both. The report got rewritten. Two months later she found the same tab open again.

The punishment had addressed the behavior. It had done nothing about the reason the behavior was appealing in the first place.

What Punishment Actually Does

When a child uses AI to bypass effortful work, the problem is not primarily a character problem. It is a habit and incentive problem. The tool made a hard thing easy. The child took the path of least resistance, which is exactly what human cognition is designed to do when the cost of effortful thinking feels high and the reward for producing the output feels immediate.

Punishment changes the cost-benefit calculation temporarily. It adds a penalty to the easier path. But it does not change the underlying pull of that path, and it does not build the cognitive habits that would make the effortful path feel more natural or more rewarding. As soon as the immediate penalty fades from memory, the original incentive structure reasserts itself.

This is why school bans have not meaningfully reduced student AI use. The RAND Corporation's 2025 nationally representative survey found that 54% of middle and high school students used AI for schoolwork, up from roughly 39% the prior year, during a period when most schools were actively discouraging or prohibiting the practice. The prohibition did not change the behavior at scale. It changed where students were careful about being seen.

Punishing a child at home operates on the same logic. The child learns to hide AI use more carefully. Independent thinking goes unchanged.

The Capability Gap Punishment Cannot Close

There is a second problem with punishment as the primary response, and it is more important than the first. A child who has been using AI to bypass effortful cognitive work has, over that period, not been practicing the thinking skills they bypassed. Those skills have not been maintained in reserve, waiting to be accessed once the tool is removed. They have been underexercised.

Removing the tool from a child who has been cognitively offloading for six months exposes the gap that practice would have filled, rather than closing it. The child now faces the effortful work without the tool and without the developed capacity to do it comfortably. The frustration that follows is not defiance. It is a genuine struggle with a task that has become harder than it should be.

This is what the Brookings Institution's 2026 review of the research called the doom loop: dependency makes independent work harder, which makes the tool more appealing, which deepens dependency. Punishment alone interrupts the loop without breaking it. The child complies under surveillance, loses access to the shortcut, finds the work genuinely difficult, and returns to the shortcut when the opportunity reappears.

What the Conversation Looks Like Instead

The more productive response to discovering AI misuse is structured differently. It begins with a question rather than a consequence, and it aims to make the student's relationship to their own work explicit.

Start by asking them to explain the work. Not accusatorially, but genuinely: walk me through your argument. What did you find most interesting? What would change if this part turned out to be wrong? A student who engaged with the material, even if they used AI as a tool in the process, can answer these questions. A student who wholesale delegated the thinking cannot. The conversation itself becomes diagnostic.

From there, the goal is not to establish that they did something wrong, but to establish what they actually know. If the answer is very little, the next step is not punishment. It is doing a portion of the work again, this time with the thinking visible: talk me through your outline before you write. Show me your notes. Tell me one thing you learned from this source. These requirements are not punitive. They are structural. They make the cognitive work a required part of the process rather than an optional one.

Reframing the Rule

Many families have an AI rule that focuses on the output: you cannot submit work written by AI. This is a reasonable starting point, but it addresses the wrong thing. The goal is not to control the output. The goal is to protect the learning process.

A more durable framing: AI is a research and review tool, not a thinking tool. You can use it to find sources, check your draft, generate counterarguments to your own position, and identify gaps in your reasoning. You cannot use it to generate the reasoning itself. What you submit needs to reflect how your mind engages with the material, not how the AI does.

This framing does not prohibit AI. It redirects it. A child working within this boundary is learning to use AI the way professionals in high-skill careers use it: as a layer over their own judgment, not as a replacement for it. And a child who understands the reason for the boundary, who sees it as protecting something valuable about their own developing mind, is more likely to internalize it than one who sees it only as a rule to navigate around.

The Longer View

The families who handle AI use most effectively are the ones who make the value of independent thinking legible to their children early enough that the child develops some investment in it.

That is a longer project than a grounding. It requires ongoing conversation, the visible modeling of how to work through a hard problem rather than outsource it, and genuine curiosity about whether the child understands what they submitted. But it builds something that punishment cannot: a child who sees their own mind as worth protecting.

The shortcut is always going to be available. The question is whether the child wants to take it.

For a practical framework on what to ask when reviewing AI-assisted work, see our post on the Explain It to Me test.

Sources cited in this post: