The Essay Was Always a Proxy. AI Just Made It Obvious.

The Essay Was Always a Proxy. AI Just Made It Obvious.

Reading Time: 8 minutes

In late 2022, OpenAI released ChatGPT to the public. Within weeks, something uncomfortable became clear: a large portion of what we had been calling "student writing" could be produced, instantly and fluently, by a machine.

The immediate reaction across higher education was predictable. Prohibition, policy memos, detection tools. But beneath that surface response was a more unsettling question:

If AI can generate a passing essay in seconds, what exactly were we assessing all along?

The essay was never the thing itself

The traditional academic essay was not a direct measure of thinking. It was a stand-in for thinking: a proxy that assumed if a student could produce a structured argument, they must have engaged in the cognitive work required to generate it.

That assumption was always fragile.

Students learned quickly that success in essay writing did not require original insight. It required pattern recognition. The five-paragraph structure. The appropriate academic tone. Strategic use of sources. Transitional phrases that signaled coherence whether or not coherence was actually present.

This is not a critique of students. It is a description of rational behavior in the face of incentives. When an assignment rewards structure over substance, students optimize for structure. When a rubric emphasizes clarity, organization, and citation format, students learn to deliver those features efficiently.

Over time, the proxy became the goal.

We stopped asking: Did the student think? We started asking: Does the essay look like thinking?

For years, that distinction was easy to ignore. Then AI made it impossible.

What AI actually changed

The arrival of generative AI did not introduce a new problem. It removed friction from an old one.

Before AI, producing a passable essay required time, effort, and at least minimal engagement. A student had to gather sources, assemble sentences, and maintain coherence across paragraphs. Even when the thinking was shallow, the process imposed some cognitive cost.

AI eliminated that cost.

A student can now generate a structured, grammatically correct, citation-filled essay in seconds. Not a brilliant essay. Not a deeply original one. But one that is often indistinguishable, at a surface level, from the median submission in many courses.

This is why faculty report a consistent experience: they suspect AI use but cannot prove it. The writing is generic but not wrong. Polished but oddly empty. Competent but interchangeable.

The problem is not that AI produces exceptional work. The problem is that it produces work that looks exactly like what we have been rewarding.

That is the more uncomfortable finding. Not that students are using AI, but that AI can satisfy our assignments so easily. If a machine can complete the task convincingly, we have to ask whether the task was ever measuring what we thought it was measuring.

Research context. Studies cited in Academically Adrift (Arum and Roksa, 2011) found that 45% of students showed no significant improvement in critical thinking, complex reasoning, or writing skills during the first two years of college.[^1] AI did not create that gap. It made the gap impossible to ignore.

Why detection is the wrong solution

In response, institutions deployed AI detection tools. The reasoning was straightforward: if students are using AI, identify them.

The approach collapsed quickly under scrutiny.

Research from Stanford HAI and other groups has raised serious concerns about the reliability of these tools. False positives are common. Non-native English speakers are disproportionately flagged. Formal, structured academic writing, the very style higher education teaches students to produce, can resemble AI-generated text closely enough to trigger alerts.[^2]

Detection systems risk penalizing the students who are working hardest to meet academic expectations.

More fundamentally, detection misunderstands the problem. Even if AI use could be perfectly identified (it cannot), we would still face the same uncomfortable question: Why does AI-generated writing satisfy our assignments so easily?

If the answer is "because our assignments reward the appearance of thinking rather than thinking itself," the fix is not enforcement. The fix is design.

Detection is an enforcement solution to a design problem. The two are not interchangeable.

What assignments look like when thinking is visible

Redesigning for visible thinking is not new. Many approaches that make intellectual shortcuts harder have existed for decades. What AI has done is increase their urgency.

The common thread across effective redesigns: they shift evaluation from what was produced to how it was developed.

Staged drafts with revision requirements. A single submitted document reveals almost nothing about how it was produced. Requiring drafts, annotated revision memos, and documented changes creates a record of thinking. Students who skip the process cannot fake a credible revision history. More importantly, students who engage genuinely with revision learn something in the process.

Personalized and context-specific prompts. The more an assignment depends on shared, publicly available information, the easier it is to outsource. Assignments that require students to draw on course-specific discussions, unique datasets, personal observations, or locally documented events are harder to replicate generically. The specificity is not a loophole countermeasure. It is better pedagogy regardless of AI.

Oral defense components. Students who can write coherently about a topic are not necessarily able to discuss it fluently and responsively in real time. Brief oral components, whether structured question-and-answer sessions or informal check-ins, surface whether students can engage with the ideas they submitted in writing. This is not a surveillance mechanism. It is a fuller picture of understanding.

Reflective process components. Asking students to analyze their own thinking, describe what changed between drafts and why, or identify where their reasoning shifted turns the assignment into a metacognitive exercise. This kind of reflection is not easily outsourced because it requires actual engagement with the work.

Position-forward prompts. AI excels at summarizing existing information and presenting multiple perspectives with apparent balance. It is less reliable at generating and defending a specific, contestable position in a specific context. Assignments that push students to stake a claim, not just survey a topic, create more cognitive demand.

Ethan Mollick, who has written extensively on AI and pedagogy, frames the shift this way: design tasks that reveal process rather than just product.[^3] The goal is not to make assignments AI-proof. The goal is to make them learning-full: the kind of work you cannot complete without developing the skill you are supposed to be developing.

These approaches are not equally easy to implement at scale. Large lecture courses, under-resourced departments, and heavy teaching loads create real constraints. That is not an excuse to defer redesign indefinitely. It is an argument for starting with one assignment this semester, one change at a time.

The question that was always there

The conversation about AI in higher education has largely centered on what students are doing wrong.

A more productive framing is what we have been assuming without sufficient scrutiny.

The essay became dominant in higher education for understandable reasons. It is flexible. It scales across disciplines. It is deeply embedded in academic culture. When well designed, it can genuinely assess reasoning and argument.

But those same qualities made it vulnerable. When an assessment tool is convenient and culturally entrenched, it is easy to stop asking whether it is working. We conflated familiarity with validity.

If an assignment can be completed convincingly by a machine, the question is not whether to ban the machine. The question is whether the assignment was ever measuring what we believed it was measuring.

Writing still matters. The capacity to construct an argument, engage with evidence, and communicate clearly remains central to higher education and to most professional fields. AI does not eliminate the need for those skills. If anything, it raises the bar for what counts as meaningful writing.

What it does mean is that the essay has to earn its place in the curriculum in a way it has not had to for decades. Clearer alignment between the assignment and the learning goal. More emphasis on process and interaction. More intentional design of prompts and evaluation criteria.

Faculty who have already started redesigning their assignments, who have field-tested new approaches and know what actually works in specific disciplinary contexts, hold the most useful knowledge in this conversation. That is exactly the kind of tested, adapted material that moves faster when it circulates.

If you have assignments worth sharing, or are looking for redesigned approaches to adapt, the catalog is a place to start.