AI Didn’t Create the Learning Problem. It Exposed It.
AI did not create the student who could complete an assignment without learning much from it. That student existed long before ChatGPT, Gemini, Claude, or any of the tools schools are now trying to figure out. Students copied from friends, searched for answers online, divided group work so efficiently that no one person understood the whole project, skimmed readings for keywords, memorized just enough to get through a quiz, and learned very quickly how to produce the kind of work a teacher was asking for. Most of us knew this was happening. We just did not always know how much of the finished product represented actual learning and how much represented a student becoming skilled at the mechanics of school.
Generative AI has made that question much harder to ignore.
A student can now produce a polished paragraph, a research summary, a presentation, an outline, or an explanation of a complicated historical event in a fraction of the time it once took. The quality of the finished product can look impressive. Sometimes it can look better than what the student would have created independently. That is where the problem becomes uncomfortable, because schools have spent a long time treating the visible product as evidence that something meaningful happened cognitively. A completed assignment, a well-organized essay, a finished project, a set of correct answers. We have often assumed that if the work looks good, then learning must have happened somewhere along the way.
Sometimes it did. Sometimes it did not.
That distinction matters more now because AI can produce the appearance of understanding without requiring the learner to do much of the thinking that normally builds understanding. If I ask a student to explain the causes of the French Revolution, an AI system can generate an organized answer about economic inequality, Enlightenment ideas, political instability, taxation, and social class. The response may be accurate. It may even be nuanced. But the quality of that response tells me almost nothing about what happened inside the student's mind. I still do not know whether the student can explain those causes without the tool, connect one cause to another, decide which factors mattered most, or recognize a weak explanation when they see one.
That is the part of the AI conversation that interests me more than the usual argument about cheating.
Cheating matters, of course. Schools need boundaries. Teachers need to know whether work is actually a student's own. But if the entire conversation stays focused on catching students who use AI, we may miss the larger issue sitting underneath it. AI has exposed how often schools have relied on the completion of a task as a proxy for learning. It has made it easier to see that producing something and knowing something are not necessarily the same event.
Learning is much less visible than schoolwork makes it seem. A student has to notice information, make sense of it, connect it to prior knowledge, hold parts of it in working memory, organize those parts into something coherent, and later retrieve the information when the original support is gone. That process is messy. It is often slower than we would like. It includes uncertainty, partial understanding, errors, forgetting, and reconstruction. A finished assignment may be the result of all of that, but the assignment itself is not proof that the process occurred.
This is where AI creates a problem, but also a useful one. It forces us to ask what part of an assignment actually carries the learning.
Take a traditional research project. A student may spend hours gathering sources, formatting citations, finding images, organizing slides, rewriting sentences, adjusting fonts, and figuring out what the teacher wants. There can absolutely be learning in that process, but there can also be an enormous amount of effort spent on things that have little to do with the concept we hoped the student would understand. If the goal is historical reasoning, then spending twenty minutes fixing a bibliography format is not the same kind of cognitive work as deciding which source is more credible or explaining why two historians interpret the same event differently.
This is why I do not think the answer is simply to remove AI from classrooms and return everything to the way it was before. Some kinds of cognitive effort probably should be reduced. If technology can eliminate unnecessary friction, that can be useful. A multilingual learner should not have to spend so much effort decoding unnecessarily complicated directions that there is little working-memory capacity left for the actual history lesson. A student should not have to fight with formatting when the goal is understanding a political system. There are plenty of places where tools can make learning more accessible without taking the learning away.
The difficulty is knowing when that line has been crossed.
There is a meaningful difference between asking AI to explain a confusing concept another way and asking it to complete every question on an assignment. One may help the learner gain access to the thinking. The other may replace the thinking entirely. Even then, the distinction is not always clean. A support that begins as a scaffold can become something a student waits for before attempting the task independently. The tool itself is not necessarily the problem. The question is what cognitive work remains with the student.
That question applies far beyond AI. Teachers already use guided notes, sentence starters, translation, graphic organizers, calculators, word banks, study guides, read-aloud tools, and other forms of support. I use many of those myself because they can remove barriers that have nothing to do with the learning target. But every support creates a second question that is easy to overlook: what can the learner do once the support is gone?
That may be one of the most important questions we ask in the next few years.
Schools are built around completion in ways we do not always acknowledge. There are pacing guides, deadlines, missing assignments, gradebooks, benchmark dates, make-up work, intervention lists, and course calendars that continue moving whether students have truly consolidated what they learned or not. In that environment, students quickly learn that finishing carries tremendous value. Submit the assignment. Clear the missing work. Get the grade. Move to the next thing.
AI fits that system very well.
If the goal is completion, it is extraordinarily efficient.
If the goal is learning, things become more complicated.
Real learning is not always efficient. Sometimes a student needs to try to retrieve an answer before seeing it. Sometimes they need to sit with a question long enough to realize what they do not understand. Sometimes they need to explain an idea badly before they can explain it well. Sometimes they need to return to information after forgetting part of it. Those moments can look unproductive from the outside because there may be no polished product appearing while they happen, but cognitively, they may be some of the most important moments in the entire lesson.
That does not mean struggle should become a virtue by itself. There is nothing educational about making a task confusing for the sake of difficulty. Poor directions create struggle. Overloaded slides create struggle. Unnecessary language complexity creates struggle. A badly designed assignment can exhaust a student's attention without building much knowledge at all. Good instructional design should remove the difficulty that gets in the way of thinking while protecting the difficulty that actually produces thinking.
AI can remove both.
That is what makes this moment so complicated.
The question for teachers is becoming less about whether students are allowed to use AI and more about what kinds of thinking we are willing to let a tool perform for them. There will probably never be one clean answer because the answer will depend on the purpose of the task. If I am assessing historical reasoning, I need to see the student's historical reasoning. If I am teaching a student how to revise writing, using AI to generate possible revisions may create an opportunity for comparison and judgment. If I am trying to determine what a student remembers independently, the tool needs to disappear.
Perhaps this will eventually change the way we think about assessment altogether. We may need to care less about how polished a final product looks and more about what students can retrieve, explain, apply, question, connect, or recognize as wrong. We may need more opportunities to watch thinking unfold instead of only evaluating what appears at the end of it. We may need to ask students why they made a particular choice, what changed in their understanding, where they became confused, or what they could still explain tomorrow without reopening the tool.
Those kinds of questions are harder to automate because they are closer to the learner.
And maybe that is the useful pressure AI is putting on education right now. It is forcing us to look more carefully at something that was already there. Students have always been able to complete work without learning as much as we hoped. They have always been able to mistake familiarity for understanding, performance for knowledge, and finishing for learning. AI did not invent any of that.
It simply made the distance between completing the work and doing the thinking impossible to miss.
If an AI system can produce an excellent assignment without learning anything at all, then the most important question may not be how to stop the tool from completing the assignment.
It may be what that assignment was ever telling us about learning in the first place.
Originally published on LinkedIn on August 26, 2026.
AI did not create the student who could complete an assignment without learning much from it. That student existed long before ChatGPT, Gemini, Claude, or any of the tools schools are now trying to figure out. Students copied from friends, searched for answers online, divided group work so efficiently that no one person understood the whole project, skimmed readings for keywords, memorized just enough to get through a quiz, and learned very quickly how to produce the kind of work a teacher was asking for. Most of us knew this was happening. We just did not always know how much of the finished product represented actual learning and how much represented a student becoming skilled at the mechanics of school.
Generative AI has made that question much harder to ignore.
A student can now produce a polished paragraph, a research summary, a presentation, an outline, or an explanation of a complicated historical event in a fraction of the time it once took. The quality of the finished product can look impressive. Sometimes it can look better than what the student would have created independently. That is where the problem becomes uncomfortable, because schools have spent a long time treating the visible product as evidence that something meaningful happened cognitively. A completed assignment, a well-organized essay, a finished project, a set of correct answers. We have often assumed that if the work looks good, then learning must have happened somewhere along the way.
Sometimes it did. Sometimes it did not.
That distinction matters more now because AI can produce the appearance of understanding without requiring the learner to do much of the thinking that normally builds understanding. If I ask a student to explain the causes of the French Revolution, an AI system can generate an organized answer about economic inequality, Enlightenment ideas, political instability, taxation, and social class. The response may be accurate. It may even be nuanced. But the quality of that response tells me almost nothing about what happened inside the student's mind. I still do not know whether the student can explain those causes without the tool, connect one cause to another, decide which factors mattered most, or recognize a weak explanation when they see one.
That is the part of the AI conversation that interests me more than the usual argument about cheating.
Cheating matters, of course. Schools need boundaries. Teachers need to know whether work is actually a student's own. But if the entire conversation stays focused on catching students who use AI, we may miss the larger issue sitting underneath it. AI has exposed how often schools have relied on the completion of a task as a proxy for learning. It has made it easier to see that producing something and knowing something are not necessarily the same event.
Learning is much less visible than schoolwork makes it seem. A student has to notice information, make sense of it, connect it to prior knowledge, hold parts of it in working memory, organize those parts into something coherent, and later retrieve the information when the original support is gone. That process is messy. It is often slower than we would like. It includes uncertainty, partial understanding, errors, forgetting, and reconstruction. A finished assignment may be the result of all of that, but the assignment itself is not proof that the process occurred.
This is where AI creates a problem, but also a useful one. It forces us to ask what part of an assignment actually carries the learning.
Take a traditional research project. A student may spend hours gathering sources, formatting citations, finding images, organizing slides, rewriting sentences, adjusting fonts, and figuring out what the teacher wants. There can absolutely be learning in that process, but there can also be an enormous amount of effort spent on things that have little to do with the concept we hoped the student would understand. If the goal is historical reasoning, then spending twenty minutes fixing a bibliography format is not the same kind of cognitive work as deciding which source is more credible or explaining why two historians interpret the same event differently.
This is why I do not think the answer is simply to remove AI from classrooms and return everything to the way it was before. Some kinds of cognitive effort probably should be reduced. If technology can eliminate unnecessary friction, that can be useful. A multilingual learner should not have to spend so much effort decoding unnecessarily complicated directions that there is little working-memory capacity left for the actual history lesson. A student should not have to fight with formatting when the goal is understanding a political system. There are plenty of places where tools can make learning more accessible without taking the learning away.
The difficulty is knowing when that line has been crossed.
There is a meaningful difference between asking AI to explain a confusing concept another way and asking it to complete every question on an assignment. One may help the learner gain access to the thinking. The other may replace the thinking entirely. Even then, the distinction is not always clean. A support that begins as a scaffold can become something a student waits for before attempting the task independently. The tool itself is not necessarily the problem. The question is what cognitive work remains with the student.
That question applies far beyond AI. Teachers already use guided notes, sentence starters, translation, graphic organizers, calculators, word banks, study guides, read-aloud tools, and other forms of support. I use many of those myself because they can remove barriers that have nothing to do with the learning target. But every support creates a second question that is easy to overlook: what can the learner do once the support is gone?
That may be one of the most important questions we ask in the next few years.
Schools are built around completion in ways we do not always acknowledge. There are pacing guides, deadlines, missing assignments, gradebooks, benchmark dates, make-up work, intervention lists, and course calendars that continue moving whether students have truly consolidated what they learned or not. In that environment, students quickly learn that finishing carries tremendous value. Submit the assignment. Clear the missing work. Get the grade. Move to the next thing.
AI fits that system very well.
If the goal is completion, it is extraordinarily efficient.
If the goal is learning, things become more complicated.
Real learning is not always efficient. Sometimes a student needs to try to retrieve an answer before seeing it. Sometimes they need to sit with a question long enough to realize what they do not understand. Sometimes they need to explain an idea badly before they can explain it well. Sometimes they need to return to information after forgetting part of it. Those moments can look unproductive from the outside because there may be no polished product appearing while they happen, but cognitively, they may be some of the most important moments in the entire lesson.
That does not mean struggle should become a virtue by itself. There is nothing educational about making a task confusing for the sake of difficulty. Poor directions create struggle. Overloaded slides create struggle. Unnecessary language complexity creates struggle. A badly designed assignment can exhaust a student's attention without building much knowledge at all. Good instructional design should remove the difficulty that gets in the way of thinking while protecting the difficulty that actually produces thinking.
AI can remove both.
That is what makes this moment so complicated.
The question for teachers is becoming less about whether students are allowed to use AI and more about what kinds of thinking we are willing to let a tool perform for them. There will probably never be one clean answer because the answer will depend on the purpose of the task. If I am assessing historical reasoning, I need to see the student's historical reasoning. If I am teaching a student how to revise writing, using AI to generate possible revisions may create an opportunity for comparison and judgment. If I am trying to determine what a student remembers independently, the tool needs to disappear.
Perhaps this will eventually change the way we think about assessment altogether. We may need to care less about how polished a final product looks and more about what students can retrieve, explain, apply, question, connect, or recognize as wrong. We may need more opportunities to watch thinking unfold instead of only evaluating what appears at the end of it. We may need to ask students why they made a particular choice, what changed in their understanding, where they became confused, or what they could still explain tomorrow without reopening the tool.
Those kinds of questions are harder to automate because they are closer to the learner.
And maybe that is the useful pressure AI is putting on education right now. It is forcing us to look more carefully at something that was already there. Students have always been able to complete work without learning as much as we hoped. They have always been able to mistake familiarity for understanding, performance for knowledge, and finishing for learning. AI did not invent any of that.
It simply made the distance between completing the work and doing the thinking impossible to miss.
If an AI system can produce an excellent assignment without learning anything at all, then the most important question may not be how to stop the tool from completing the assignment.
It may be what that assignment was ever telling us about learning in the first place.
Originally published on LinkedIn on August 26, 2026.