Is the College Essay Dying? How AI Is Reshaping Assessment in Higher Education

The college essay is not dead, but the assumption underneath it, that a finished piece of writing proves a student did the thinking, no longer holds by default. That is the real disruption, and it is forcing a debate in higher education that is sharper, and more split, than most coverage of "AI and cheating" suggests.

Redesign or retreat: Higher Ed's split response to AI

One camp argues AI has simply exposed a weakness that was always there. EDUCAUSE Review's June 2026 piece on the state of AI in higher education makes this case directly: for decades, assessment relied on artifacts like essays and exams as imperfect proxies for learning, and those proxies persisted because they scaled to large cohorts, not because they measured understanding well. On this view, AI hasn't broken assessment. It has revealed that the essay was never a reliable signal of learning in the first place, and the honest response is to redesign assessment around what AI cannot substitute for: applied reasoning, defended arguments, demonstrated process.

A second camp is not persuaded that redesign is the answer, and is moving in the opposite direction, back toward the exact formats the first camp says are broken. The University of Glasgow announced a return to in-person examinations in 2026, part of a broader move across UK institutions toward verifying work by hand rather than trusting a written submission at face value. In the US, named faculty describe the same instinct in practice: a St. John's University computer science professor now requires handwritten exams and code, and a Michigan State associate dean who runs faculty AI workshops has faculty requiring students to show their work rather than submit a finished answer alone.

Both camps are responding to the same underlying failure, and neither position is really about the essay. It is about academic integrity: whether an institution can still trust that a submission reflects a student's own thinking, and whether that trust gets rebuilt around what AI has changed or defended by retreating to conditions where AI cannot reach.

This piece argues that the more durable path is the first one, but with a condition that gets skipped in most of this debate: redesigning assessment around AI only works if the AI involved is built for the pedagogy it is supporting, not adapted from a general-purpose tool after the fact. A model that can hold a conversation is not the same as one designed to support formative feedback, protect an instructor's judgment, or scaffold a task toward demonstrated understanding. Purpose matters as much as capability, and most of the tools shaping this debate were not built with assessment integrity as a design constraint from the start.

The redesign response includes hybrid, process-over-output models and faculty training programs, but faces real pushback that it risks piling on more assessment rather than fixing it.

What does redesigning assessment for AI actually look like?

Redesigning assessment for AI means shifting the graded artifact from a single finished document to something that shows how a student got there, whether that is a defended draft, a verification step, or a task that requires demonstrated reasoning rather than a polished answer. Several institutions are already building faculty training specifically around this shift, rather than leaving it to individual instructors to work out alone.

The clearest example of the redesign logic in practice comes from the same Times Higher Education reporting on UK universities that documented Glasgow's return to exams. The same piece describes institutions moving toward assessing process over output: third-year civil engineering students, for instance, might be asked to use AI to help design a building, then verify every AI output themselves through hand calculations and modeling, while English literature students continue submitting essays but also submit drafts or revision memos as evidence of authorship. Neither example abandons the underlying skill being taught. Both add a layer that makes the AI's contribution visible and puts the verification or authorship burden back on the student.

This is close to exactly the hybrid model description in circulation elsewhere: keep the essay, but attach a defense. Faculty Focus has published a two-part framework from Associate Professor Torrey Trust built around this idea directly, including the finding that banning AI outright risks worsening inaccessibility and digital divides among students, and that redesigning assignments to require process visibility is a more durable response than a ban. Her TRUST model, cited in the same piece, gives faculty a structured way to build that visibility into an assignment rather than treating it as an afterthought.

Institutions are also building this capability deliberately rather than expecting faculty to develop it independently. Penn State's Teaching and Learning with Technology unit launched a faculty cohort program for Spring 2026 specifically to redesign assignments for what it calls AI-resilience and AI-inclusion, requiring faculty to complete two full assignment redesign projects rather than a single workshop. Similar structured sessions exist at the University of Guelph's Office of Teaching and Learning and the University of Missouri's Teaching for Learning Center, both built around the same core move: examine an assignment that AI has undermined, and rebuild it around authentic, multimodal, or scaffolded tasks that lower the incentive to shortcut it.

What connects all of these examples is that none of them are chasing detection. They are changing what gets assessed and how visible the process is, which is a fundamentally different kind of institutional investment than buying a better detector, and one that requires faculty development, not just policy.

Is this assessment redesign wave creating a new problem?

Yes, according to a growing number of faculty: redesigning assessment around AI is adding assessment volume rather than replacing it, and the result is more work for both students and instructors without a clear net gain in learning. This is not a minority complaint. It surfaces directly in faculty response to the redesign trend itself.

Times Higher Education's reporting on this backlash captures the concern in blunt terms from faculty commentary: the number of formative assessments feeding into a single summative grade has grown, often through a one-size-fits-all policy that can work against longer-form assessment like the essay rather than support it. One frequently cited comparison from that discussion is a course that once ran on a handful of ungraded formative essays across a term, now redesigned into a stack of process checkpoints, revision memos, and verification steps layered on top of the original assignment rather than replacing any part of it. The complaint is not that redesign is the wrong idea. It is that redesign, done without discipline, becomes addition instead of substitution, and the instructor time this was supposed to protect ends up spent on more administration instead.

This tension shows up in how educators actually talk about AI, not just in published commentary. A large-scale academic study analyzed 270,000 AI-related posts and comments across 26 education subreddits, from November 2022 to April 2026, run by researchers from the University of Edinburgh, Durham University, and Middle East Technical University. It found that early discussion settled into a detection-and-enforcement mindset, treating AI mainly as an academic integrity threat to be policed, and constructive conversation about redesigning assessment only started to gain real ground partway through that period. One finding stands out. Threads about catching AI misuse get far more engagement than threads about redesigning assignments, and they last two to four times longer. The angrier the conversation, the more it spreads. That matters here because it means the loudest version of "responding to AI" in educator communities is enforcement, not redesign, even though redesign is the response with more staying power.

That pattern reinforces the over-assessment risk above: if enforcement dominates the conversation, it is the easier default to reach for, and redesign that is not grounded in clear pedagogical purpose tends to slide toward more checking rather than genuine rethinking. A revision memo requirement bolted onto an unchanged essay assignment is not the same thing as rethinking what the assignment is for. The real test of redesign is whether the new component replaces something the essay used to do inadequately, or simply adds a hurdle on top of it.

This is also where the tool doing the supporting starts to matter, not just the assessment design itself, which the next section takes up directly.

What does an AI tool built for assessment actually need to get right?

It needs to be built around the pedagogy it is supporting, not adapted from a general-purpose assistant after the fact. That distinction, more than any single redesign tactic covered in this piece, is what separates AI that helps assessment survive this moment from AI that adds to the problem.

For institutions evaluating tools against the redesigned formats above, the design criteria that matter reduce to a short list:

• It drafts, it does not decide. It produces a rubric-anchored draft for an instructor to review, never a grade or verdict issued on its own.
• It is built for pedagogy, not repurposed from a general-purpose assistant with assessment bolted on afterward.
• It understands the redesigned artifact. It can read a revision memo, a process checkpoint, or a verification step for what it is, not just as a block of text.
• It defers to instructor judgment on the parts a rubric cannot capture, rather than filling the gap itself.
• It supports the student's own preparation and practice without performing the task for them.
• It removes real workload without removing judgment.

One more design requirement follows directly from the redesign patterns in the previous section: assessment automation cannot be all-or-nothing. An institution running verification steps on one assignment and revision memos on another needs a tool that can be configured per assignment, not applied as a blanket policy across a course. A civil engineering task that already requires hand verification may need only light-touch support, while a discipline still relying on a traditional essay may need fuller feedback coverage. The instructor sets that boundary assignment by assignment, and the tool respects it. This is also the direct answer to the most common objection redesign efforts run into: that no two courses, or even two assignments within the same course, work the same way. A tool that assumes uniform application across a syllabus is solving a problem few instructors actually have.

That configurability matters most in exactly the formats where pacing is tightest: online courses, recertification and continuing-education programs, and MOOCs, where a cohort's ability to keep moving depends on fast, reliable turnaround rather than a single high-stakes submission. These are also the contexts where academic momentum, not just academic integrity, is the thing an institution is protecting.

The scale of the gap between adoption and readiness is well documented. EDUCAUSE's 2026 report on AI's impact on learning assessment, based on a survey of 438 faculty and staff, found real momentum toward using AI in assessment design, alongside genuine uncertainty about how to do it responsibly. That combination, momentum without settled practice, is exactly the condition in which a poorly designed tool does the most damage: it gets adopted quickly, before anyone has worked out what it should and shouldn't be trusted to do.

A tool designed for assessment has to hold a specific line that a general-purpose AI assistant has no reason to hold: it drafts, and a human decides. For example, our AI-assisted feedback and grading tool produces a rubric-anchored draft for an instructor to review rather than issuing a grade outright, and it applies with equal force to the redesigned formats covered in this piece. A revision memo, a process checkpoint, or a verification step is only useful if whatever reviews it understands what it is looking at and defers to the instructor's judgment on the parts a rubric cannot capture. A general-purpose AI has no built-in reason to know where that line sits. A tool built for this specific job does, because the line is part of its design, not an afterthought bolted on with a policy document.

This is also where the redesign tactics covered earlier connect back to something practical for institutions evaluating tools right now, not just assignment formats. A student preparing for a defended, process-visible assessment, the kind covered in this piece's second section, benefits from practice and preparation support ahead of the actual task, which is the role AI Student Tutor is built to play: supporting a student's independent work without doing the performance itself. The same discipline that should apply to redesigning an assignment, does this change what gets measured, or does it just add friction, applies to redesigning a tech stack. A tool added because it can hold a conversation is friction. A tool added because it protects instructor judgment while removing real workload is redesign.

Consider what this looks like for a single student rather than a policy. An online nursing student finishing a recertification course submits a graded assignment at 9 p.m. and is told review will take seven days, a routine turnaround for institutions running against the model this piece has been arguing against. Under a well-designed AI-supported model, that same student gets a rubric-anchored draft response within minutes, not a grade, but a starting point for revision, and can move on to the next assignment the same evening instead of losing a week of momentum. That gap, seven days versus same-night feedback, is not a hypothetical efficiency gain. It is the practical difference between an assessment model that fits how these programs actually run and one that does not.

That is the standard this piece has been building toward. The essay is not dying, and neither is the exam, the presentation, or the verification step. What is dying is the idea that any single format, or any single tool, can be trusted by default. What replaces it has to be chosen with the same rigor institutions are learning to apply to the assignments themselves.

Ready to see how AI-supported feedback and grading works without taking judgment out of faculty hands? Book a demo with the team.