Prove it
before anyone asks.

How do I prove I wrote it before I'm ever accused?

You can't stop an AI checker from flagging clean, formal writing. But you can make the flag easy to answer. Keep a record of how you write, as you go. Numbered drafts. Your notes and outlines. Edit history left on. Then if a flag ever comes, the evidence that you wrote it already exists — instead of being rebuilt from memory, weeks later, when it counts most.

The difference isn't how much you have. It's when it was made. A contemporaneous record — in plain words, one captured at the time, before any charge existed — can't be dismissed as self-serving. That timing is the one thing you can only secure in advance.

One warning. A record only counts as contemporaneous if it genuinely is. Backdating files, or rebuilding a history after the flag, is the opposite of protection — it reads as exactly what it is, and hands the other side a second, heavier problem → Evidence that actually stands →

Why does a record of my process actually work?

Because a real first draft is hard to fake — which is exactly why integrity offices ask for progressive drafts and process documentation. Authentic writing leaves a mess. False starts. A thesis that changed. Paragraphs you cut. AI-generated text leaves none of that. And the asymmetry is the whole engine: to fake a believable draft history, you'd have to know what the finished paper says before you write the middle of it. Real writing can't know that yet. (Cornell ELSO · Harvard OAISC)

This is why the messy version protects you more than a polished early draft would. Incompleteness is the proof. A perfectly clean progression — perfect outline, perfect draft, perfect paper — is easier to doubt, because that's the shape a fabrication takes. The dead ends are what a person leaves behind.

But notice the load-bearing word: captured. The asymmetry only protects you if the messy path actually landed somewhere. A brilliant, false-started composition that lived only in your head, then got typed cleanly into one file, leaves the same empty trail as a fabrication. That's why where you write turns out to matter — the next question, and one most students never think to ask. It's also the line the cases turn on. Students kept their footing when a record of their process met a flag standing alone — not when they argued innocence in the abstract → The cases — coming soon

What habits actually build that record?

Six habits. None of them is hard. The only hard part is doing them before you need them — which is exactly why the reader who's still early has the advantage.

Which of these stands, which backfires, and the full forensic mechanics once you're already accused → Evidence that actually stands →

Where should I actually write — does the platform matter?

Yes. The platform decides how much your record is worth. Integrity offices tell you to keep drafts and document your process. Not one of them says where to do it. That gap matters more than it looks — so close it yourself.

What makes any record strong is the same across every habit above. It was captured by something outside your control. Captured continuously. Captured at the time — not curated by you afterward. Measure the options against that:

• A folder of manually saved files barely clears the bar. The dates live on your own device, where they can be changed. A batch saved in one sitting shows nothing about process. And the whole thing depends on you remembering to save at each step. Better than nothing — but the weakest version of "a record."
• A continuous, cloud-based history — the kind Google Docs keeps automatically as you type, stored on a server you don't administer — is tamper-evident in a way a folder of files is not. In plain words: it's recorded as the work happens, by the platform, and changing it later is hard to hide. That's the same quality that makes an outside-system timestamp strong. You couldn't have curated it to fit a story.

The honest limits, up front. An automatic history is not a guarantee. A thin one still reads against you. And a finished paper pasted in all at once defeats the point, no matter where you paste it. Which version histories stand, which get turned around, and exactly why → Evidence that actually stands →. The rule for here is simpler than all of it: write where the record makes itself — from the first word to the last, so the messy path actually gets captured.

What about the gray zone — Grammarly, translation? Get it in writing.

Schools can't agree where the line is — even inside one tool. So don't guess. Ask, in writing. The University at Buffalo says Grammarly for spelling and grammar is generally fine. Using it to "re-write or re-generate text" usually isn't. Translation tools like DeepL need a green light first — they now run on generative AI, not word-for-word lookup (University at Buffalo).

Why do schools contradict each other — and even themselves — on the same product? Because most policies were written around Grammarly's older, rule-based grammar engine. Then the same product folded in generative rewrite features. The tool quietly became "more AI" while the policies stood still. That's how one office can bless Grammarly while the next lists it beside ChatGPT. You can't out-argue that contradiction. But you can close it, permanently, for your own work.

The move: before you use anything borderline, get your instructor's answer in writing. One email removes an entire category of future ambiguity. Say exactly this:

"For [assignment], I'd like to use [Grammarly for spelling and grammar only / DeepL to check a translation / an outline tool]. I won't use it to re-write or generate text — only [to fix mechanics / to confirm a phrase, with the wording staying mine]. Could you confirm in writing whether that's within your policy for this course? I want to be sure I stay on the right side of it."

A dated reply that says "yes, that's fine" cannot later become an accusation about that use. You've turned a gray zone into a paper trail — on your side, before anyone is looking for one.

What am I supposed to show — that I didn't use AI?

Show what you did do. A real record answers the "didn't" by itself. This is the obligation most people get backwards. Cornell's writing-support office tells students preparing for a hearing to gather evidence of the ways they used and didn't use generative AI (Cornell ELSO). The obligation is double-sided. In plain words: proving a negative isn't enough, and it's nearly impossible by design anyway. So you affirmatively show what you actually did at each stage. A real, documented process is concrete — it's what a fair reviewer can actually weigh.

A record built in advance does both jobs at once. It documents your real process. And by existing, it answers the negative — because a genuine compositional trail is the one thing AI-generated text doesn't leave.

Will this guarantee I'm safe?

No — and anyone who promises that is wrong. A record helps in front of a fair reviewer, and it made the difference in the cases students won. It can't fix bad faith. It can't force a process that has decided to ignore it. What it does is real but bounded. It lowers the panic. It shortens the fight. It gives a fair process something solid to land on. You move from defending after the fact to having documented all along — a hearing-ready position, not a magic shield.

If the worst does happen later — who carries the burden, what a hearing actually decides — that's its own subject → What happens at the hearing →

The hard part this page leaves you with

Read back over every habit here and one thread runs through all of them. The record that protects you is the one that already exists — captured automatically, before any question, by something outside your control. Everything you can assemble after the flag is the weak version of that. That's why it's weak. So the real work isn't knowing what to keep. It's remembering to keep it. Every draft. Every session. By hand. On the one assignment that turns out to matter, months before you have any reason to think it will. That gap, between knowing and remembering, is the whole problem this page can't solve with advice alone.