The ChatGPT Essay Problem Is a Symptom, Not the Disease

The conversation about AI in education tends to focus on the wrong end of the problem. Schools are investing in detection software, revising honor codes, redesigning assessments. All of that is reasonable. But the underlying question — why can ChatGPT pass these assignments at all — is getting less attention than it deserves.

What it means when AI can do your homework

If a language model can produce a passing essay on a topic, the assignment is testing something that language models can do. That's worth thinking about. Essays that ask students to summarize, explain, or give an overview of something are exercises in retrieval and organization — skills that AI genuinely has.

Essays that ask students to take a position on a contested question based on their own experience, or to connect course material to something specific that happened in class, or to defend an argument they made in a previous assignment — those are harder for AI, because they depend on inputs the model doesn't have.

The assignments AI can pass most easily are often also the assignments that do the least for learning. The overlap isn't coincidental.

The detection arms race

Schools that respond only with detection are going to keep losing. Detection tools have significant false positive rates, produce outcomes that are difficult to defend to students or parents, and can be circumvented with basic editing. The model producing an essay and the model editing an essay to defeat detection are the same model.

More fundamentally, playing catch-up with rapidly improving tools is an infrastructure that will always be one step behind.

Assessments that actually work

The assignments that hold up are ones that require something only the student can provide. Oral components where students have to explain and defend what they submitted. In-class writing that produces a baseline for comparison. Assignments where the process — annotated bibliography, revision history, peer review responses — is as graded as the final product.

None of this is new pedagogy. Teachers have been designing assessments against various forms of academic dishonesty for decades. AI changed the scale of the problem, not the nature of it.

The honest conversation

Some students were always going to take shortcuts. Some students are taking shortcuts with AI now who wouldn't have before, because the barrier is lower. And some students are using AI in ways that are genuinely useful — brainstorming, feedback, structural help — without submitting AI text.

Treating all three groups identically, with blanket prohibition and mass surveillance via detection tools, is both impractical and unfair. The better frame is probably: what kind of AI use builds understanding, and what kind substitutes for it? That question has a more defensible answer than "all AI use is cheating," and it's one that doesn't require pretending a technology doesn't exist.