Why this happens to honest writers
Accusations usually start with a detector score or a hunch about style. Both are weak evidence. Independent testing by Weber-Wulff and colleagues found no detection tool accurate or reliable enough to support accusations, and Liang and colleagues found detectors flagging non-native English writers far more often. Clear, well-organised writing can look machine-like to both people and tools. The guide to whether an AI detector can judge a social post explains why short posts are the weakest case of all.
Evidence that settles it
| Evidence | Where to find it | Why it persuades |
|---|---|---|
| Version history | Your editor, CMS revisions, document history | Shows the text growing over time with real edits |
| Dated drafts and notes | Local files, note apps, email to yourself | Shows thinking before the final wording |
| Sources you used | Bookmarks, reading lists, citations | Shows research that shaped the piece |
| Original image files | Camera roll, exports with metadata or credentials | Shows capture and editing history |
| Explaining the work | A call or a reply | Shows understanding a copied text rarely brings |
WordPress keeps post revisions by default on most setups, which is often the fastest proof available: the revision screen shows how a post changed from first save to publication. For images, the original files with their metadata or Content Credentials do the same job, as checking Content Credentials explains.
How to respond
- Ask what the claim rests on. If it is a detector score, say so plainly and point to the published research on false positives.
- Offer the evidence that fits the setting: revision history for a client, a short public note for readers, the platform's appeal route for a label.
- Do not run your text through detectors until one says human; results from other tools prove as little as the first.
- If you did use a tool for part of the work, say exactly what it did. Precise disclosure ends most disputes faster than denial.
Accused of using AI writing: a worked example
Suppose a freelance writer delivers a 1,400-word feature and the client replies that a detector scored it 92 percent likely machine-written. The details are invented for this example. The writer's CMS shows 23 saved revisions over 6 days, the first a 300-word outline and the last the delivered text, with a 900-word middle draft that reads quite differently from the final. Her notes folder holds 4 dated files with interview quotes and a reading list of 11 sources, 9 of which appear in the piece.
She sends the revision screen as a short recording, the dated notes and the reading list, and offers a 15-minute call to walk through how the argument changed between drafts. She also points to a published study of detector accuracy and mentions that a high score on formal prose from a non-native writer is a known failure mode. The client withdraws the claim the same day. Had she run the piece through four more detectors until one called it human, the exchange would have lasted longer and settled less.
Pitfalls
- Writing in a tool that keeps no history, then pasting the final text into the CMS in one save. The revision screen then shows one version, which proves nothing.
- Deleting drafts once a piece is published. Keep them for at least a year; disputes arrive late.
- Answering with anger or a counter-accusation. The evidence does the work; the tone decides whether anyone reads it.
- Letting a house style or a heavy edit erase the writer's voice, so that every piece reads alike and looks machine-made to a detector.
- Denying any tool use when a tool was used for part of the work. One discovered omission costs more than an honest, specific disclosure.
What to record
Keep an authorship record for every published piece, written at the time and stored with the drafts. The rows below are examples. A record like this takes a minute per piece and turns a dispute into a lookup. The page-level signals that make accusations rarer are listed across the provenance section.
| Piece | Written in | Revisions | Tools used and for what | Evidence kept |
|---|---|---|---|---|
| Harbour feature, 1,400 words | CMS editor | 23 over 6 days | Spelling checker only | Revisions, 4 note files, reading list |
| Weekly roundup, 600 words | Desktop editor, pasted | 1 in CMS, 14 local versions | Language model for the first outline, disclosed | Local versions, outline prompt, disclosure line |
| Guest column, 900 words | Sent by the author | 2 | Unknown | Author's drafts requested at commissioning |
Make the next accusation less likely
Put a named author on every post with an author page, and carry authorship into the share card, as the guides to byline schema and authorship tags describe. Publish a short disclosure policy and follow it, using the wording in disclosure lines that survive a share. None of this proves authorship on its own, but it gives readers a person to trust and a record to check. The rest of the provenance section covers images.
Sources
- Weber-Wulff et al., Testing of Detection Tools for AI-Generated Text (2023)
- Liang et al. (2023), detectors of machine-written text are biased against non-native English writers
- C2PA, the Coalition for Content Provenance and Authenticity
Questions
How do I prove I did not use AI to write something?
Show the writing process: revision or version history, dated drafts, notes and sources, and be ready to explain the work.
Can I use an AI detector to clear my name?
Not convincingly. Detectors disagree and published tests found none reliable enough to serve as evidence either way.
Does WordPress keep proof of how I wrote a post?
Post revisions, when enabled, record each saved version with a date, which shows how the post developed.
Should I disclose if I used a tool for part of the work?
Yes, precisely. Say what the tool did and what you did. Specific disclosure is more credible than a general denial.
What if my client insists on a detector result?
Offer the process evidence anyway, and ask them to run a piece they know is human through the same tool. Published tests show false positives are common enough that the demonstration usually ends the argument.
Hannah Voss, Editor. Checks every guide against a working WordPress install and the networks' current documentation. Last reviewed September 2026.
Related reading
- Can an AI detector judge a social post?Why AI text detectors are least reliable on the short, informal text of social posts and captions, what published tests found, and what to rely on instead.5 min read
- Byline schema in WordPress: marking up a human authorMark up article authors with schema.org Person data in WordPress: the fields search engines read, author pages, sameAs profiles and common mistakes.6 min read
- AI disclosure lines that survive a shareWhere to put an AI assistance disclosure so it reaches readers who only see the share card: titles, descriptions, bylines and the first lines of a post.5 min read
- Authorship tags for shared links: author, article:author and fediverse:creatorAdd authorship signals to link previews: the author meta tag, article:author, twitter:creator and fediverse:creator, with WordPress code and testing steps.6 min read