How to Prove You Wrote Something: A Practical Evidence Playbook

It's 2am and your phone buzzes. You roll over, squint at the screen, and see an email from a client you've been writing for since last summer. Subject line: "Concerns about your latest deliverable." Your stomach tightens before you even open it.

Inside, three sentences. They ran your article through an AI detector. It came back 87% AI-generated. They're holding your payment.

You wrote every word. Three days of research, an outline scrawled across two pages of a legal pad, a first draft hammered out in Google Docs over a long Saturday, a rewrite of the intro because it felt flat, then a final pass where you trimmed 400 words and tightened the transitions. Your partner even read it aloud to check the rhythm.

But right now, none of that matters. Because you can't prove any of it.

I've talked to dozens of people in some version of this situation -- a graduate student in Toronto whose thesis chapter got flagged two weeks before defense, a freelance journalist who lost a $4,000 contract over a Copyleaks score, a high school senior whose college application essay was questioned by an admissions officer who'd never met her. The specifics change. The core problem doesn't: they did the work, and when challenged, they had nothing concrete to point to.

This guide exists so you don't end up there. Not through paranoia -- through simple habits that take minutes and produce evidence worth its weight in gold.

The Uncomfortable Timing Problem

Here's the thing nobody wants to hear: the time to build your defense is before anyone questions you.

I know that sounds obvious. But think about what happens after an accusation lands. Everything you produce looks reactive. A draft you "find" the next morning looks fabricated. A process you reconstruct from memory sounds like a convenient story. And honestly? The people evaluating you are right to be skeptical. If someone accused of plagiarism suddenly shows up with a perfectly organized evidence folder 24 hours later, that raises questions, not confidence.

That asymmetry is why this guide focuses on habits, not crisis management. The goal is building evidence as a natural side effect of how you already work. No extra apps, no daily rituals, no surveillance of yourself. Just small adjustments to the tools and workflows you already use, so that if the question ever comes, you can pull up a folder and say: "Here. All of it."

How I Think About Evidence Layers

Not all proof carries the same weight, and different types of evidence answer different questions. I find it helpful to think in layers -- each one useful alone, but dramatically stronger when stacked together.

Your draft history is the foundation. It shows your writing evolving through time, with all the false starts, deleted paragraphs, and ugly first attempts that characterize real human work. If you only do one thing from this guide, do this.

Independent timestamps answer the "when" question. A file modification date on your laptop can be fudged. An email sitting on Google's servers with their timestamp? That's a different story. Simple, old-school, surprisingly hard to argue with.

Voice fingerprinting tackles the "who" question. This one changed how I think about the whole problem. Stylometry -- the statistical analysis of writing style -- has been used in forensic linguistics for decades. The idea is that you have measurable, stable writing patterns: your vocabulary choices, sentence rhythms, punctuation habits, how you transition between ideas. A voice fingerprint captures those patterns across multiple dimensions and compares them against a disputed text. It doesn't ask "is this AI?" It asks "is this consistent with how this person writes?" Much better question.

Provenance trails go deeper than snapshots. Instead of periodic saves, provenance captures the continuous act of creation -- typing speed, pauses, the order sections were written, where you stopped to look something up, where you deleted a paragraph and started over. Think of draft history as a series of photographs. Provenance is the video.

Third-party attestation means someone else saw your work in progress. An editor who reviewed an early draft. A colleague you sent your outline to. A professor you discussed your thesis with three weeks before the deadline. Their testimony, backed by timestamped communication, is surprisingly powerful.

You don't need all five layers for everything you write. An email to a friend? Obviously not. But a client deliverable worth thousands of dollars? A dissertation chapter? A piece you're submitting to a major publication? Stack as many layers as the stakes justify.

Draft History: The Thing You're Probably Losing

Most writers save the final version of a piece and that's it. The outline goes in the trash. The rough draft gets overwritten. The research lives in browser tabs that get closed when the laptop restarts. From an evidence perspective, it's like a painter destroying every sketch and study and then asking people to trust they didn't trace the finished work.

Here's what to do instead.

Write in something that tracks revisions automatically. Google Docs is the easiest path. It logs every change with a timestamp -- you can see the full history through File > Version history > See version history. Notion keeps page history too, though less granularly. If you're comfortable with a text editor, use Git. I mean it. Even if you're not a developer, learning git init, git add, and git commit takes about ten minutes and gives you an ironclad revision trail that nobody can tamper with after the fact.

Name your versions at turning points. Google Docs lets you label versions (File > Version history > Name current version). I label mine things like "outline done," "rough draft complete," "post-editor revisions," "final pre-submission." Those labels become your narrative later. They show intention and process, not just text accumulation.

Keep the mess. That paragraph you cut because it was going nowhere? The three versions of your opening you tried before one clicked? Don't delete them. Human writing is messy and iterative. AI output isn't. Your mess is genuinely valuable evidence -- it shows the kind of wrestling with ideas that no text generator replicates.

Save your research alongside your drafts. Bookmarks, annotated PDFs, screenshots of source pages, exported highlights. If you can demonstrate that you read and engaged with the material that informed your arguments, you've shown intellectual process -- not just text production.

A graduate student I spoke with last year kept what she called a "reading log" -- a simple text file where she'd paste a quote from whatever paper she was reading and jot a one-line reaction underneath. When her committee questioned a chapter, she opened the log and walked them through her thinking, source by source, reaction by reaction. The inquiry lasted five minutes.

Timestamps: Boring, Independent, Powerful

The word that matters here is "independent." Evidence that lives entirely on your own machine, under your own control, is easy to question. Evidence that sits on someone else's servers -- with their timestamps, under their control -- is a different thing entirely.

Email yourself at milestones. Old trick. Still works. Finish your outline? Email it to yourself. Complete a first draft? Forward it. Before final submission? One more email. Each one creates a dated record on Gmail's (or Outlook's, or Yahoo's) servers that you can't retroactively alter. Don't overthink the subject lines. "outline" or "draft1" is fine. The point is the timestamp and the attached content, not the metadata.

Export PDFs at key moments. After a major revision pass, print to PDF. The file embeds a creation timestamp in its metadata. Store these in a dedicated folder in Dropbox or Google Drive -- somewhere backed up that you won't accidentally clean out. On Mac, just use File > Print > Save as PDF. Windows has Microsoft Print to PDF built in.

Keep a writing log. This is maybe the most underrated piece of evidence I've encountered. A simple spreadsheet -- date, project name, what you worked on, rough word count. I use a Google Sheet because it's automatically timestamped and cloud-backed. A typical entry: "Feb 8 -- Drafted sections 2 and 3, rewrote the opening, +1,400 words." It's mundane. That's the point. Nobody fabricates boring documentation. Nobody invents a fake writing log with realistic word counts across six weeks of sessions. The sheer tedium of it makes it credible.

Block writing time on your calendar. If you schedule focused writing sessions (and you should, for reasons beyond evidence), those calendar entries become data points. "2pm-5pm: Draft client article" is small, but it adds texture to the overall picture.

None of this is about building a surveillance system for your own work. It's about leaving breadcrumbs. When someone asks "when did you write this?" you want specifics, not "I think it was sometime last week."

Voice Fingerprinting: The Layer That Changed My Thinking

Draft history shows process. Timestamps show timeline. Voice fingerprinting answers a harder question: does this text sound like the person who claims to have written it?

The science is called stylometry, and it has real pedigree. Researchers have used it to attribute anonymous pamphlets from the American Revolution, identify the authors of disputed Shakespeare plays, and provide evidence in criminal cases involving threatening letters. The core finding is that writers have stable, measurable patterns across vocabulary, syntax, sentence length variation, punctuation usage, paragraph structure, transitional habits, and more. These patterns are consistent enough to distinguish between individuals and stable enough to persist across years.

Building a voice fingerprint with WritersLogic works in three steps. First, you analyze 3-5 samples of writing you know are yours -- ideally from different contexts and time periods. An email, a report, a blog post, an academic essay. The variety captures your real range, not just one register. Second, when a dispute arises, you run the questioned text against your baseline. The comparison measures consistency across multiple dimensions simultaneously. Third, you get a report showing exactly where the new text matches your patterns and where it diverges, with dimensional scores that can be examined and explained.

Why is this stronger than draft history alone? Because drafts can theoretically be manufactured. Someone could paste AI-generated text into a Google Doc and type over it gradually to simulate a writing process. Difficult, but conceivable. But faking a consistent voice profile that matches an established baseline across vocabulary distribution, syntactic complexity curves, punctuation frequency, semantic coherence patterns, and argument structure? That's not a thing people can do. The dimensionality is too high. You'd have to simultaneously replicate someone's word choices, their rhythm, their comma habits, and how they build paragraphs. It's like trying to forge a fingerprint -- the more you know about one ridge pattern, the more you realize you're getting the others wrong.

I've seen voice fingerprinting resolve disputes that would have dragged on for weeks. A professor runs a comparison between a student's previous essays and the flagged paper, looks at the dimensional match scores, and the conversation shifts from accusation to apology.

Provenance: The Video, Not the Photo

Provenance captures what snapshots miss: the continuous, real-time act of writing. Typing speed and rhythm. Pauses between sentences. The order sections were composed. Where you stopped mid-paragraph to check a fact. Where you deleted three sentences, sat for forty seconds, and wrote them differently.

This matters because human composition and AI generation look fundamentally different at this level. A person writes in bursts. They hesitate. They jump between sections. They type a sentence, backspace through half of it, rethink, try again. They stop to reread what they just wrote. An AI produces text sequentially, at consistent speed, without the starts and stops of someone actually thinking.

These patterns are essentially impossible to fake retroactively. Even if someone tried to "perform" human writing behavior while copying AI text, the irregularity wouldn't be right. Real thinking pauses are unpredictable and context-dependent -- longer before a complex sentence, shorter in the middle of a familiar argument. Simulated pauses follow their own detectable rhythm.

Witnesses: The Overlooked Layer

This one's optional, but powerful. If someone else encountered your work in progress, that's attestation.

Share an early draft with a colleague and ask for quick feedback. Their reply -- with a timestamp in their email client -- becomes evidence. Use collaborative editing in Google Docs; Google logs who accessed the document and when. If you're a student, talk to your professor about your topic or thesis weeks before the deadline. That conversation itself is a form of attestation -- if they remember discussing your argument with you on a specific date, that's context no AI detector can generate.

What the Morning Looks Like When You're Prepared

Back to that 2am email. Imagine you'd spent 30 minutes building these habits six months ago. When you wake up, here's what happens.

You open a project folder. Inside: your outline in Google Docs with version history stretching back eight days, three named draft versions showing the piece growing and changing, two emails to yourself with intermediate drafts attached (timestamped by Gmail), your writing log showing three sessions totaling nine hours, and a voice fingerprint report showing your article matches your baseline across ten stylistic dimensions.

You bundle these into a PDF, write a short cover note -- "I take this seriously; here's my complete process documentation" -- and send it by 9am.

In my experience, that's where disputes end. Not with arguments about detector accuracy or emotional appeals. With evidence. Organized, multi-layered, quietly overwhelming evidence. Nobody assembles a trail like this to cover up pressing a button.

The Mistakes That Make Things Worse

Starting after the accusation. Evidence assembled after the fact looks like evidence assembled after the fact. Evaluators know this. Opposing counsel knows this. Build the habits before you need them.

Keeping only the finished product. A polished final draft proves nothing. In fact, high polish can actually work against you -- AI output tends to be clean and structured, so pristine text without any messy history behind it looks suspicious, not professional. Your ugly first drafts are worth more as evidence than your final submission.

Going all-or-nothing. You don't need five evidence layers for every email you send. Scale your documentation to the stakes. A text to a friend: nothing. A $5,000 client project: full stack. A dissertation chapter: everything you've got.

Leaning on a single proof. One screenshot of a draft isn't enough. One email timestamp isn't enough. One voice analysis isn't enough. But combine all three? Now you have a position that's very difficult to argue against.

The 30-Minute Setup

If you're starting from zero, here's the sequence.

Pick a tracked writing tool. Google Docs requires zero configuration -- just start writing there. If you prefer plain text, spend ten minutes learning Git basics.

Create a writing log. Open a Google Sheet with four columns: Date, Project, Work Done, Word Count. Bookmark it. Update it at the end of each session. Thirty seconds per entry.

Build a voice baseline. Gather 3-5 samples of your writing from different periods and contexts. Run them through WritersLogic. Save the report. Fifteen minutes.

Set up an evidence folder. One folder per project, somewhere cloud-backed. This is where PDF snapshots and draft copies go.

Start emailing yourself drafts. Finish a section, forward it. Five seconds. An independent timestamp you'll be grateful for later.

That's 30 minutes of setup and maybe two minutes per writing session going forward. Within a week, it's automatic.

The Question That's Not Going Away

"Did you actually write this?" is going to get asked more often, not less. Models get better. Detectors stay unreliable. The stakes keep climbing -- in academia, in publishing, in freelancing, in journalism, in college admissions.

The writers who navigate this well won't be the ones who avoid AI tools or the ones who argue endlessly about detection accuracy. They'll be the ones who can open a folder and show their work. Not out of paranoia. Out of preparation.

Start building that folder now. Future you will be grateful.

Build your voice baseline or see what organized evidence looks like in a sample report.