Process Over Product: Grading Drafts and Version History to Make AI Cheating Hard

A chatbot can hand a student a finished essay in twenty seconds. What it can’t hand them is two weeks of messy drafting, a proposal that got rewritten after your feedback, and an outline with three abandoned ideas crossed out. That gap is the whole opportunity. If you grade the journey instead of just the destination, the easy shortcut stops paying off.

This is the case for process-based grading: building your assessment around drafts, outlines, and version history so that the work has to happen over time, in the open, where you can see it.

Key takeaways

• Grading the process — drafts, outlines, version history — is harder to fake than grading a polished final product, because a coherent paper trail takes real time to build.
• Version history in Google Docs or Word builds an automatic edit timeline. Asking students to write inside the document makes the trail appear for free.
• A process rubric shifts points away from the finished essay and toward observable milestones: a proposal, an annotated source list, a marked-up first draft, a revision memo.
• This approach is more defensible than an AI detector score. Detectors report a probability, not proof, and they’re known to misfire — especially on non-native English writers.
• Process-based grading isn’t just an anti-cheating tactic. It teaches revision, which is where most of the actual learning in writing happens.

Why the final product became the wrong thing to grade

For most of the history of the essay, grading the finished paper made sense. The product was a reasonable proxy for the work. If the paper was good, the student probably did good thinking to get there.

Generative AI broke that proxy. A polished final draft no longer tells you who did the thinking. The output and the effort have come apart, and grading the output alone means grading something a model can now produce.

The reflex is to reach for a detector and try to catch the fake. That reflex has a problem: the tools aren’t reliable enough to carry the weight people put on them. They report a probability, not a fact. OpenAI quietly retired its own AI Text Classifier in July 2023, citing a low rate of accuracy — when the company building the models couldn’t ship a dependable detector, that tells you something. And a 2023 Stanford study (Liang and colleagues, with James Zou as senior author), published in *Patterns*, found that several popular detectors flagged writing by non-native English speakers far more often than writing by native speakers. If you want the mechanics behind why that happens, it comes down to statistical signals like perplexity and burstiness.

So instead of trying to detect the product after the fact, you can change what you grade. Grade the part a chatbot was never involved in: the process.

What “process” actually means when you grade it

Process-based grading sounds abstract until you turn it into things you can collect and mark. Here’s the concrete version. You’re assigning points to artifacts that show the work unfolding over time.

Drafts you can see being built

A rough draft submitted before the final isn’t just a checkpoint. When students write inside a shared Google Doc or Word file, the document carries its own version history — an automatic, timestamped record of the essay growing paragraph by paragraph. You don’t have to police it. It builds itself.

The one habit that defeats this is writing somewhere else and pasting the finished piece in at the end. A single giant paste with no edit trail is exactly what looks suspicious, even when the student wrote every word. So the instruction is simple: draft inside the document, and let it be messy.

Outlines, proposals, and annotated sources

Before the draft, there’s the planning. A one-paragraph proposal. An outline that names the argument. A list of sources with the student’s own notes on each. These are cheap to produce and cheap to grade — often just a completion check — and they pin the work to a real person making real decisions.

A revision memo

This is the quietly powerful one. After the final draft, ask for a short note: what changed between the rough version and this one, and why. “I moved my strongest point to the front because the original opening was weak.” A student who actually revised can write this in their sleep. A student who pasted a finished essay has nothing to describe, because nothing changed.

A mini-scenario: redesigning a standard essay assignment

Take a typical assignment: a 1,200-word argumentative essay, due in three weeks, graded once at the end.

Now rebuild it around process. Same topic, same length, same learning goal. But the grade breaks into pieces:

• Week 1 — Proposal (10%). One paragraph naming the argument and two sources, submitted in the shared doc.
• Week 2 — Rough draft (20%). At least 600 words, written inside the document, graded for completion and effort, with your feedback returned.
• Week 3 — Final draft (60%).
• With the final — Revision memo (10%). Three to five sentences on what changed since the rough draft and why.

Watch what this does to a student tempted to outsource the whole thing. To fake it, they’d need to generate a proposal, fake a rough draft that looks naturally incomplete, generate a final that plausibly builds on it, write a memo describing revisions that match the gap between the two versions, and somehow get the document’s version history to show all of it piling up over time. That’s far more work than just writing the essay. Meanwhile the honest student barely notices the structure, because it’s how careful writing already works.

Notice the trade. You didn’t add a detection step. You added design, mostly once, and you spread the grading across three weeks instead of one brutal stack at the end.

Building a rubric that rewards the journey

The rubric is where this lives or dies. If 90% of the grade still rides on the polished final, students will optimize for the polished final. Move the points.

A workable split puts meaningful weight on the process artifacts — say, 30 to 40% across proposal, draft, and revision memo combined — and writes criteria that reward *change over time*. The phrase to build around is “evidence of substantive revision.” Did the thesis sharpen? Did the structure reorganize? Did weak evidence get replaced? You’re not grading whether the student drafts in long bursts or tinkers line by line. You’re grading whether the work genuinely evolved.

One caution: keep the criteria about substance, not style. Some people write clean first drafts and revise lightly; others write a mess and rebuild it three times. Both can show real revision. Make the rubric ask “did this get meaningfully better and why” rather than “did this change the way I expect,” and you keep it fair across working styles.

Where detection still fits (a small role)

None of this means you can never glance at a detector. It means the detector stops being the verdict and becomes, at most, a faint signal that might prompt you to look harder at the process evidence you already have.

The difference matters when you actually talk to a student. A draft history that shows an argument forming over two weeks is something you can both read and discuss. A likelihood score is not — it’s a number from a tool with documented blind spots, and leaning on it drops you straight into the false-positive trap that hits honest and non-native writers hardest. For more on how these tools work and where they break, there’s more on detection and writing on the PaperBleach blog.

The point of process-based grading isn’t to win a detection arms race. It’s to make that race irrelevant.

Frequently asked questions

Doesn’t grading the process create a lot more work for me?

It front-loads design work and spreads grading across the term instead of dumping it all on one final stack. You set up the milestones once, then grade smaller pieces as they arrive. The final read often goes faster because you already know each paper’s history. Lightweight completion grades on the outline and a quick rubric on the rough draft keep the load down. You’re trading one grading marathon for several short ones.

Can’t a determined student just fake the version history too?

Some will try. A student could type a chatbot’s essay into the document in chunks to mimic a natural build. But that’s far more effort than pasting, and it tends to leave tells: the argument arrives fully formed, with no false starts or abandoned paragraphs. Real drafting is messy. Pair version history with a revision memo and a short oral check, and faking all three becomes more work than writing the paper.

How is this different from just using an AI detector?

A detector gives you a number after the fact and asks you to trust it. Process evidence gives you a record you can read and discuss. Detectors output probabilities, not proof, and research has found they flag non-native English writers more often than native speakers — OpenAI retired its own classifier in 2023 for low accuracy. A draft history showing an argument evolve over two weeks is concrete in a way a likelihood score never is.

What’s the single easiest way to start this term?

Require one rough draft inside a shared Google Doc or Word file before the final, and grade it for completion. That one step turns on version history, gives you a midpoint check, and anchors the final to a real timeline — no major redesign needed. Add a revision memo or outline milestone later.

Is it fair to grade students on drafts when some write differently than others?

Yes, if your rubric rewards evidence of revision rather than a specific style. Some students write in bursts, others tinker line by line. A good process rubric asks whether the work changed in meaningful ways — reorganizing, sharpening a thesis, adding evidence — not whether it changed in a particular way.

Closing

The fastest way to make AI cheating hard isn’t a better scanner. It’s an assignment where the grade depends on work that has to happen over time, in a document you can watch fill up. Drafts, outlines, version history, a revision memo: each is a small thing a chatbot can’t supply on its own, and together they quietly close the easy shortcut while teaching the revision skills that mattered all along.

Start with one assignment this term. And if you want to understand the detection tools your students are up against before you decide how much weight to give them, check a sample against a detector yourself and see what the score really tells you.