How to prove authorship of a document: the evidence hierarchy for disputes

The short answer

Authorship disputes are decided by layers of evidence, not a single proof. The hierarchy runs from the weakest (unsubstantiated assertion) to the strongest (a sealed, independently verifiable process record created before any dispute existed). Most people facing an accusation have access to at least the middle layers; the question is knowing which to present first and what each layer actually establishes.

Why authorship disputes arise

The contexts differ; the evidentiary logic is the same across all of them.

In academia, the dominant modern dispute is AI-authorship accusations: an AI detector flags a submission and a student must demonstrate they wrote it. But the category is older — ghostwriting suspicions, plagiarism allegations, and "did you actually write this?" interviews long predate AI detection.

In employment and IP, disputes arise over who authored a contract, report, policy document, or creative deliverable. If a departing employee claims ownership of work done on company time, or a contractor claims credit for a deliverable, the question is the same: who fixed the expression?

In legal contexts, authorship questions arise in defamation (who wrote this?), fraud (was this document forged?), will contests (did the testator author this document or was it written for them?), and IP litigation. Courts look at chain of custody, access to tools, and evidence of the creative choices being made.

The hierarchy below applies to all three contexts, though the weight assigned to each tier varies by forum.

The evidence hierarchy

Tier 1 — Assertion alone (weakest)

Saying "I wrote this" is the universal starting point and also the weakest possible evidence. Everyone making a claim of authorship asserts it; the assertion distinguishes nothing. It is not zero — an uncontradicted assertion under oath carries some weight — but it is insufficient when challenged, in any forum.

What it establishes: that you claim authorship. Honest limitation: everyone claims authorship; it proves nothing contested.

Tier 2 — Stylometric comparison

Stylometry — using writing style as a fingerprint — has a legitimate history in forensic linguistics. But as the primary literature documents, its accuracy ceiling against a known author is around 76.5% in controlled conditions, and performance degrades on shorter texts, texts written in a second language, and texts that have been heavily edited. A stylometric analysis arguing "this matches your other writing" is corroborative, not conclusive.

What it establishes: consistency of style with attributed work. Honest limitation: ~76.5% accuracy ceiling; can be defeated by deliberate style-matching; fails on short texts. See stylometry and authorship attribution for the full treatment.

Tier 3 — Research trail and correspondence

Emails discussing the work, library records, research notes, drafts in personal folders, browser history from the period of composition, message threads asking questions about the topic — these together form a circumstantial trail that shows the author engaging with the subject before and during writing.

What it establishes: that you were thinking about, researching, and working on the subject at the relevant time. Honest limitation: correspondence can be fabricated; the trail shows engagement, not composition specifically; does not address paste-and-edit strategies.

Tier 4 — Document version history

For cloud-hosted documents, version history is a passively generated record created by a third party (Google, Microsoft) before any dispute existed. That pre-dispute creation is its main evidentiary strength: you had no motive to fake it at the time. Version history has cleared real academic accusations; a documented case involved a student whose hours of incremental Google Docs history was sufficient to end a GPTZero-based accusation.

What it establishes: that the document was built incrementally over time from an account associated with you. Honest limitation: snapshots are coarse (minutes apart, not keystroke-by-keystroke); does not prove who was at the keyboard; publicly available tools can simulate gradual composition history. See Word version history as evidence for the Microsoft-specific picture.

Tier 5 — Supervised demonstration

Can you discuss the work fluently? Can you reproduce a section, explain the choices you made, answer probing questions about the reasoning? In academic integrity proceedings, an oral examination on the submitted work is one of the most powerful tests available — not because it proves composition (someone could have studied the AI-generated text), but because authentic understanding of one's own writing is very difficult to fake in real time, under pressure, on specific passages.

What it establishes: that you understand and can reproduce the work at the level of authorship, not just recall. Honest limitation: a student who studied their AI-generated text thoroughly might pass; some students struggle in oral examination through anxiety rather than dishonesty; not available in every forum.

Tier 6 — Sealed, verifiable process record (strongest)

A process record — keystroke-level capture of the composition event, sealed cryptographically when writing ends, independently verifiable — is the only tier that is observational rather than inferential. It was created during composition, before any dispute existed, by a party (the writer, with consent) who had no motive yet to fabricate it. It cannot be silently altered because the seal breaks if it is. And it does not require trusting anyone: the reviewer checks the cryptographic seal directly.

Crossley et al. (EDM 2024) demonstrated that keystroke-level process data achieves 99% accuracy in distinguishing human-composed text from transcribed/pasted content — a discrimination that version history and stylometry cannot approach.

What it establishes: that writing occurred, at the observed pace and revision pattern, from a known account, and that the record has not been altered since sealing. Honest limitation: it records the process, not the intent; it does not prove no assistance existed (no system can); its strength depends on the recording layer (OS-level is stronger than document add-on). See what a writing record cannot prove.

Evidence type	What it proves	Main limitation
Assertion	That you claim authorship	Everyone does
Stylometric match	Style consistency	~76.5% accuracy ceiling; short-text fragility
Research trail	Engagement with the subject	Circumstantial; fabricable
Version history	Incremental document growth	Coarse snapshots; simulable
Supervised demonstration	Understanding of the work	Anxiety, coaching, not universal
Sealed process record	The composition event itself	Scope limited to what was recorded

A worked example: the accusation arrives

A student receives a written concern notice about a 2,200-word history essay. The notice says a detection tool flagged it with a high AI-probability score. There is no sealed process record — the student had never heard of them. This is the most common situation. Here is what the first 48 hours actually look like.

Hour 1: preserve first, panic later. The single most important action is to stop editing the document immediately and take screenshots or exports of every version-history state available. Google Docs version history, Word's version history, OneDrive snapshots — capture all of them now. Editing after receiving a concern notice creates a gap in the record that will be hard to explain. If the document is in Google Docs, go to File → Version history → See version history and screenshot the full timeline. The timestamps shown are server-generated; they are not under the student's control.

What they gather. The student works backward through the Tiers:

• Tier 4 (version history): Google Docs shows 23 saved versions over three days. The earliest version is 400 words. Versions 8 through 14 show a paragraph being written, deleted, and rewritten — the kind of revision that nobody adds to a fabricated incremental history. This is the strongest piece of available evidence and it was created before any dispute existed.
• Tier 3 (research trail): The student finds four browser history entries for the same Wikipedia article visited on different days, two library database searches, and a thread in the course discussion board where they asked a question about one of the sources. Together these show engagement with the specific topic during the relevant window.
• Tier 2 (stylometric): This is not worth commissioning independently for a student proceeding. It would be expensive and its ~76.5% accuracy ceiling would not settle the dispute.
• Tier 1 (assertion): The student writes a clear statement of their process. This matters less than the evidence above but establishes a narrative the evidence can support.

What is in their favour. The version history shows incremental construction over three days, not a document that appeared fully-formed. The research trail is contemporaneous. The concern notice came from a detection score alone — no other evidence of violation was cited. Under the OIA's burden-of-proof guidance (covered below), the institution must prove the violation, not the student prove innocence. A detection score alone, with contradicting version history, is a weak case for the institution.

The weak point. Version history does not prove who was at the keyboard. It proves the document was built incrementally from an account associated with the student, not that the student was the one typing. A sophisticated argument against Tier 4 is that gradual construction can be simulated. The student cannot rebut this argument with anything in their possession — they have no keystroke-level record, no OS-layer capture. If the institution presses hard on this specific point, the best response is to request an oral examination (Tier 5), where fluent discussion of specific sentences and choices is very difficult to fake.

The first 48 hours in order:

1. Stop editing. Screenshot or export all version history now.
2. Compile the research trail while memory is fresh — browser history, notes, library records.
3. Read the institution's academic integrity policy and identify who hears the case and what the burden-of-proof standard is.
4. Locate the detection tool vendor's published guidance on score interpretation (Turnitin's own documentation states the score "should not be used as the sole basis for adverse actions").
5. Write a factual narrative of how the essay was written and what the version history shows.
6. Request an oral examination if the institution does not offer one automatically.

The lesson is not that the outcome is guaranteed — it is that the evidence available, presented in the right order, tells a credible story that the detection score alone does not rebut.

What copyright law looks at

Copyright attaches to the author at the moment of creation — the person who fixed the original expression in a tangible medium. As the U.S. Copyright Office's Circular 1 states, copyright protection exists "from the time the work is created in fixed form." The WIPO guide to copyright uses the same language: the person "who created the work" is the author, and creation is about who made the creative choices, not who typed final keystrokes.

This framing is directly relevant to authorship disputes. Courts and tribunals are asking: who made the expressive choices? Who had access to the tools and the time? Who can demonstrate knowledge of the choices that were made? Process records answer these questions more directly than finished-text analysis because they document the choices being made, in sequence, at the time.

What the Adelphi ruling established

In February 2026, a student won an AI plagiarism lawsuit against Adelphi University. The court found the university had acted "without valid basis and devoid of reason" in pursuing the accusation. The significance for evidence: the court was not persuaded that a detection score alone constituted a valid basis for adverse action. The ruling does not establish binding precedent across all institutions, but it signals the direction: detector-only evidence is not sufficient to meet a reasonable evidentiary standard when challenged.

What the OIA says about burden

The UK's Office of the Independent Adjudicator — the higher education ombudsman — published casework guidance specifically on AI and academic misconduct. The OIA's consistent position is that the institution bears the burden of proof, not the student. A student is not required to prove they did not use AI; the institution must prove that they did. AI detection scores, the OIA notes, are not sufficient on their own to discharge that burden. The burden-of-proof framing matters: if the institution must prove the violation, the student's job is not to prove innocence but to present enough evidence to make the charge unsubstantiated.

Practical checklist by dispute type

Academic integrity proceeding:

• Gather all version history immediately and stop editing (preserve the record)
• Compile research trail: browser history, notes, library access
• Prepare to discuss the work in an oral examination
• Quote the vendor's own caveat against treating the score as sole basis for action
• If you have a sealed process record, submit it with the response

Employment or IP dispute:

• Establish timeline with email correspondence and meeting notes
• Gather all saved drafts and file metadata
• Identify any witnesses who saw you work on the document
• Check whether any collaboration platform (Slack, Teams, Google Workspace) captured access logs

Legal proceeding:

• Engage a forensic linguistics expert for stylometric analysis (Tier 2) if needed
• Preserve all digital metadata — file creation timestamps, cloud sync logs
• A sealed process record, if available, is the most direct evidence for discovery

How Authmarks fits in

If you submitted through Authmarks, a sealed process record is available at verify without an account — it covers everything in this hierarchy that version history and correspondence cannot: keystroke-level composition data, sealed on completion, independently verifiable without trusting the writer or the platform.

Related reading: what is writing provenance · how to prove you didn't use AI · tamper-evident vs tamper-proof · google docs version history as proof · process data in integrity cases