The Two-Stage Submission Model: Submit the Prompt History Alongside the Essay

Most academic integrity fights start at the wrong end. A teacher runs an essay through a detector, gets a scary percentage, and now has to prove a negative to a student who swears they wrote it. The two-stage submission model flips that whole script: instead of judging the finished essay in isolation, you ask for the essay *and* a record of how it was made.

It’s a quietly powerful idea. You stop guessing whether a machine was involved and start reading what the student actually did.

Key takeaways

• The model collects two artifacts: the finished essay and a record of the process, including any AI prompts and outputs.
• It moves grading from policing the product to reviewing the process, which sidesteps the false-positive trap of AI detectors.
• A prompt history is verifiable in a way a detector score is not — you can read what was asked and how it was edited.
• It’s gameable and adds some grading time, so it works best as one signal among several, backed by clear policy.
• Detectors give probabilities, not proof; a process record gives you something concrete to discuss.

What the two-stage submission model actually is

The structure is simple. Stage one is the essay you’d collect anyway. Stage two is a short companion document — a prompt log, a process note, an appendix — that shows the road the student traveled to get there.

If they used an AI tool, that’s where the prompts and outputs go, along with a line or two on what they kept, cut, or rewrote. If they didn’t use AI, stage two captures the human version of the same thing: an outline, a messy first paragraph, a note about where they got stuck. Either way, you end up with evidence of thinking, not just a polished result that arrived from nowhere.

Think of it as “show your work” for the writing process. Math teachers have asked for this forever. The essay equivalent just took longer to feel necessary.

Why “require AI prompt history with submission” is catching on

The appeal is that it dodges the problem detectors can’t solve. An AI detector reads the final text and estimates a probability. It can’t tell you *how* the text was made, only how statistically predictable it looks — which is why over-edited human writing and lightly-touched AI writing can score almost identically. If you want the math behind that, we’ve written up the statistics behind AI detectors elsewhere.

A prompt history doesn’t estimate anything. It’s a record. You read it, and either it tells a coherent story of a student wrestling with the assignment, or it doesn’t.

A mini-scenario: two essays, one good question

Picture two students who both turn in a clean, B-plus essay on the causes of the 1929 crash.

Maya’s stage-two log shows three prompts. The first asks ChatGPT to explain margin buying “like I’m in tenth grade.” The second pastes her own paragraph and asks whether her timeline is right. The third is a list of follow-up questions she answered herself. Her notes flag two places where she disagreed with the AI’s framing and rewrote them. You can see her thinking move.

Daniel’s log is one prompt: “Write a 900-word essay on the causes of the 1929 stock market crash.” The output matches his submission almost word for word.

Neither essay would necessarily trip a detector, and Daniel’s might even score *cleaner* if he ran it through a humanizer. But the process records tell completely different stories. You don’t need a probability to know which student engaged with the material. And critically, if you want to talk to Daniel, you have something specific and factual to point at — not a contested number.

What it catches, and what it doesn’t

It catches: wholesale generation with no engagement, essays that drift from the assigned prompt, and the gap between a student’s classroom ability and a suspiciously sophisticated submission. It also legitimizes disclosed AI use, so students who used a tool the right way aren’t lumped in with those who didn’t.

It doesn’t catch: a student who fabricates a believable log after the fact. Someone determined enough can write the essay with AI, then reverse-engineer a plausible-looking prompt history to match. The model raises the effort and the risk, but it isn’t tamper-proof. Treat it as a strong signal, not a verdict.

That limitation matters. The two-stage model is a process tool, not a lie detector. Its power is that it gives you a concrete artifact to discuss in a meeting or an oral check-in — and most students who try to fake a log fold quickly when asked a follow-up they didn’t anticipate.

Why this beats leaning on a detector alone

AI detectors have a documented reliability problem. OpenAI launched its own AI text classifier in early 2023 and pulled it that July, citing a low rate of accuracy. A 2023 study from Stanford researchers found that several popular detectors disproportionately flagged writing by non-native English speakers as AI-generated — punishing students for the very thing detectors are supposed to be neutral about.

A prompt history sidesteps both problems. It doesn’t care whether a student’s prose happens to be statistically smooth. It doesn’t penalize an ESL writer for clean grammar. It asks a different, fairer question: can you show me how you got here?

That said, detectors still have a place in the workflow — as a *spot-check* that tells you where to look more closely, not as the thing that decides a grade. If a section reads oddly, you might check the draft against the student’s prompt log. Students can run the same self-check before they submit, scanning a near-final draft to see which passages read as generic and then tightening them in their own voice.

How to roll it out without drowning in paperwork

The biggest objection from teachers is grading load, and it’s a fair one. A few practices keep it sane.

Give students a template

Don’t make them invent the format. Hand out a one-page structure: prompts used, key outputs, what you changed and why. A predictable shape is faster to skim and harder to pad. Keep the bar reasonable — you want evidence of process, not a forensic transcript.

Skim by default, read closely on a trigger

You don’t read every log line by line. You skim, and you slow down only when the essay raises a flag: a leap in sophistication, sourcing that appears from nowhere, a thesis that wanders off the prompt. Used as a targeted spot-check, the overhead stays small.

Be explicit about what’s allowed

The model only works if your AI policy is clear. If students don’t know whether brainstorming with ChatGPT is fine, their logs will be defensive and vague. Spell out what’s permitted, what must be disclosed, and what’s off-limits — then the prompt history becomes a disclosure tool instead of a confession booth.

Pair it with at least one in-person signal

A short oral check-in, a timed in-class paragraph, a quick “walk me through your second draft.” Any of these makes a fabricated log much riskier and gives you a second data point that’s far harder to fake than text.

A quick word on framing

How you introduce this to students changes everything. Pitch it as surveillance and you’ll get resentment and theater. Pitch it as “I want to grade your thinking, not just your final paragraph,” and most students relax — especially the honest ones who were nervous about being falsely accused by a detector in the first place.

That framing is also more truthful. The point isn’t to catch people. It’s to make the *process* visible and gradable, the way a lab notebook or a set of revision drafts always has been. AI just made the invisible part of writing easy to skip, and this model puts it back on the table. If you want more on the surrounding debates, you can browse more on AI detection and academic writing.

Frequently asked questions

What exactly should students submit as their prompt history?

At minimum, the prompts they typed into any AI tool, the relevant outputs, and a short note on what they kept, cut, or rewrote. Some instructors ask for screenshots or exported chat logs; others accept a clean appendix. The goal is to see the decisions, not collect a forensic transcript, so keep the bar reasonable. A page is usually plenty.

Doesn’t this just punish students who used AI at all?

It shouldn’t, and that’s the point of framing it as process rather than confession. If your policy permits AI for brainstorming or drafting, a clear log is evidence of legitimate, disclosed use. The students who struggle are the ones who pasted in finished text and can’t explain how it got there. That’s the gap the model is meant to surface.

Can’t a student just fabricate a believable prompt history?

Yes. A motivated student can reverse-engineer a plausible log, so this isn’t tamper-proof. It raises the effort required and creates a record you can ask follow-up questions about, but treat it as one signal among several. Pair it with in-class writing or oral check-ins rather than leaning on it alone.

How is this better than running essays through an AI detector?

Detectors give a probability, not proof, and they produce false positives — a 2023 study from Stanford researchers found several detectors disproportionately flagged non-native English writers. A prompt history gives you something concrete to read and discuss instead of a contestable percentage.

Does requiring a prompt log add a lot of grading work?

Some, which is why most instructors skim the log and read it closely only when the essay raises a question. Used as a spot-check rather than a full second grade, the overhead stays manageable.

The bottom line

The two-stage submission model won’t end every integrity question, and it isn’t immune to gaming. But it trades a guessing game for an evidence trail, and that’s a much better place to stand than arguing about a detector score nobody can fully trust. If you’d like students to see how their writing reads before they hand it in, point them at PaperBleach so they can check a draft and tighten the flat spots in their own words.