How to prove you didn't use AI: every type of evidence, ranked

The short answer

Evidence of authorship comes in layers, and the layers are not equal. From weakest to strongest:

1. Running the text through other AI detectors
2. Your past writing as a style baseline
3. Drafts, outlines, notes, and your research trail
4. Document version history
5. Talking about the work — explanation or a supervised sample
6. A sealed, verifiable record of the writing process itself

If you are responding to an accusation right now, gather layers 2–5 today — the step-by-step guide covers the process. This article explains what each layer actually proves, so you lead with your strongest material instead of your weakest.

Evidence	What it shows	Weakness	Effort to gather
Counter-scores	Detectors disagree	Accepts that scores decide anything	Minutes
Writing baseline	The work sounds like you	Subjective; your style may be what got flagged	Minutes–hours
Drafts & research trail	A process existed	Scattered; thin if you draft in one sitting	Hours
Version history	The document grew over time	Coarse; not bound to a person; can be simulated	Minutes
Discussing the work	The process lives in your head	Needs a fair hearing; stressful; after the fact	A meeting
Sealed process record	The process, tamper-evident and attributable	Must exist before the accusation	Set up once

Three mistakes to avoid, whatever you have

Leading with the weakest layer. The most common instinct — screenshots of four detectors saying "human" — is the least persuasive thing you can send, because it keeps the argument on the accuser's terms: dueling scores. Open with process evidence (layers 3–4), offer the conversation (layer 5), and use detector disagreement only to question whether a score should decide anything at all.

Destroying evidence while tidying up. People delete "embarrassing" messy drafts, clear browser history out of habit, or keep editing the flagged document. The mess is the evidence — a too-clean trail looks curated. Freeze everything the moment you learn of an accusation, and only share from copies.

Overclaiming. Saying "my version history proves I wrote every word" invites a reviewer who knows about typing simulators to discount all of it. Present each layer for what it is — "here is a timeline consistent with my account, here are my drafts, and I'll answer any question about the work" — and let the layers corroborate each other. Calibrated claims read as honesty; absolute claims read as salesmanship.

1. Counter-scores from other detectors (weak)

The instinct is understandable: the accuser used a detector, so you run the essay through three others that say "human" and call it settled. The problem is that this accepts the premise that detector scores settle anything. They disagree with each other constantly, their output changes as models update, and the vendor that flagged you — Turnitin — publicly cautions that its own score should not be the sole basis for action. A stack of contradictory probabilities is an argument about detectors, not evidence about you. Use the disagreement to challenge the accusation's basis, not as your proof of innocence — and if you want real ammunition on reliability, cite the universities that reviewed AI detection and turned it off rather than another score.

2. Your writing history as a baseline (circumstantial)

Earlier essays, exam scripts, emails — anything that shows the flagged work sounds like you. This is genuinely useful context, especially against the "this is too good / too polished" style of accusation. Its limits: style is subjective, people's writing improves, and detectors are documented to flag particular writers persistently — a 2023 Stanford-led study found non-native English speakers' genuine writing was falsely flagged at an average rate of 61.3% across seven detectors. If your authentic style is the thing being flagged, a style baseline both helps (consistency) and stings (the same features get flagged again). Treat it as supporting material.

3. Drafts, notes, and the research trail (good)

An outline with crossed-out ideas. Two messy intermediate drafts. Handwritten notes. Browser history showing you opened the twelve sources you cited, spread over days. A message to a friend complaining about paragraph three. None of these alone is decisive; together they sketch a process no one fabricates after the fact for a single assignment. This is the classic evidence that integrity panels understand, and it is the reason to never delete anything until a dispute is over.

4. Version history (strong, with real limits)

If you wrote in Google Docs or Word online, the platform recorded snapshots as you went. Timestamped, incremental growth of a document over hours and days is the strongest free evidence most writers have — it cleared a UC Davis student accused on the basis of a GPTZero score. It is strong precisely because it is a record of process, not a guess about output.

But it has limits you should know before you lean on it: the snapshots are coarse, nothing binds them to a person, and tools now exist that fake the process itself — public "typing simulators" replay arbitrary text into a document so the revision history shows gradual human-looking writing that never happened. We cover this honestly in Is Google Docs version history enough? The practical takeaway: version history is your best free layer, and an institution aware of simulators may ask what corroborates it.

5. Talking about the work (strong, in person)

Offer to discuss the essay: why you ordered the argument this way, what you cut, what source 4 says. Offer to write a supervised sample on a related prompt. A person who did the work can do this fluently; a person who didn't, can't. UK ombuds guidance (OIA) treats a student's account of their process as relevant evidence, and the offer itself signals confidence. The limit is obvious: it depends on a fair hearing, it stresses exactly the people already disadvantaged by the process (non-native speakers, anxious students), and it happens after the accusation — it cannot prevent one.

6. A sealed process record (strongest)

Every layer above is reconstructed after the fact from whatever happened to survive. The strongest evidence inverts that: a record of the writing process, captured as it happens with your consent, cryptographically sealed when you finish, and verifiable by anyone — instructor, panel, court — without trusting you, your institution, or the vendor. That is writing provenance: the record is tamper-evident (edit it and the seal breaks), attributable (a named person attests to it), and portable (the proof travels with the document, not inside somebody's platform).

This is what Authmarks builds. Recording is off until you turn it on for a specific document; the sealed record is yours; and anyone can verify it without installing anything. If you've ever had to assemble layers 1–5 under a deadline, the case for layer 6 makes itself — get set up before the next assignment.