Designing writing assignments when detectors fail: what teaching centers actually recommend

The short answer

The major teaching centers have converged on the same advice: stop trying to catch AI in the finished product, and redesign assignments so the process is visible. Cornell's Center for Teaching Innovation says it outright: "We currently do not recommend using automatic detection algorithms." Berkeley's framing is the positive version of the same idea: "assessment isn't just about what students submit, it's about how they got there."

This article collects what those recommendations actually say, the specific assignment-design moves they suggest, and the one gap they all share — the process evidence they ask for is easy to request and hard to verify.

Why the teaching centers gave up on detection

Cornell's guidance is the bluntest on the evidentiary problem: "Detection tools claim to identify work as AI-generated, but they cannot provide evidence for that claim." A score is an opinion with no cross-examinable basis — the full argument is in alternatives to AI detectors, and the list of institutions that acted on it is in universities that turned off AI detection.

The MLA-CCCC Joint Task Force on Writing and AI — the two major professional bodies for writing instruction — adds the equity dimension: its first working paper cautions against detector reliance in part because false accusations "disproportionately affect marginalized groups." And it warns against swinging to the opposite extreme, urging faculty to use technology to "support students' intellectual and social development… rather than to surveil, discipline, or punish."

So the professional consensus rules out two easy answers at once: don't trust the detector, and don't build a surveillance classroom. What's left is assignment design.

The design moves the centers recommend

1. Scaffold the process into the grade. Berkeley's core recommendation is to break writing into visible stages — "brainstorming, outlining, drafting, peer feedback, and reflection" — each carrying weight. A student who must show an outline, a rough draft, and a revision memo has produced process evidence as a side effect of doing the work. A student who pasted a finished essay has nothing to show for the intermediate stages.

2. Anchor assignments in local material. Berkeley suggests building prompts around class discussion, lab data, local readings, or campus events — "knowledge that GenAI tools don't have access to." A general-model essay on Hamlet is effortless; a general-model essay engaging with what your seminar argued about on Tuesday is visibly hollow.

3. Ask what the evidence would be before you need it. Cornell's guidance reframes integrity cases around comparison: suspicion becomes actionable when work is "inconsistent with previously submitted work or preparatory steps… such as outlines, drafts, or prior discussions." That only works if preparatory steps exist on the record — which is an assignment-design decision made in week one, not an investigation technique available in week ten.

4. If AI use is allowed, make the interaction the artifact. The MLA-CCCC paper suggests having students document their AI use — the transcript of prompts and responses becomes part of the submission, turning tool use from a secret into a citable step.

The honest limits of assignment design

Three gaps remain even when you do all of this.

• Scaffolding can be laundered. A student can generate the essay first and reverse-engineer the outline and "rough draft" from it. Staged submissions raise the effort of cheating; they do not verify anything.
• Process documents are self-reported. Version history is the usual backstop, but it records states, not authorship — Google Docs version history as proof walks through the ceiling.
• It doesn't scale down. Reading five process artifacts per student works in a 20-person seminar. In a 300-person survey course, the scaffolding becomes a grading burden that quietly collapses back to judging the final product.

Where a provenance record fits

Everything above asks students to produce process evidence. A writing-provenance record is process evidence — created automatically, with the student's consent, while they write. It shows typing, pasting, and revision over time, sealed so it cannot be edited afterward; reading one takes about two minutes once you know the pattern, which is covered in how to read a process record.

This squares the MLA-CCCC circle, because the record belongs to the student, not the institution — it is consent-based protection a writer carries, not surveillance an instructor imposes. The student seals their record; anyone can verify it without an account; and when work is "inconsistent with preparatory steps," the record either backs the student or it doesn't — no detector opinion required. Pair it with the scaffolding above and the conversation guide in how to talk to a student about suspected AI use, and the integrity process stops depending on a score nobody can cross-examine. It is free for individual students, so requiring nothing of your department budget.