Who gets to check AI’s invisible ink?

The Borrowed Lamp is our fictional workshop notice: twelve places, borrowed tools, an entrance with two steps. It is ordinary enough to pass from a writer to an editor, then into a newsletter. Those details make the publishing decisions visible. Who checked the price? Who accepted the wording about access?

We retained the starting notice, openly drafted with Codex. It is an AI-assisted illustration, with an intentionally uncorrected grammar error. The proofreading and rewriting requests in our declared protocol remain unperformed; no current-provider text detector has checked it. This gives us a document to follow, without dressing an intended experiment as a result.

The useful distinction begins before any detector. The writer can keep the request and original output; the editor can keep the changes. A final paragraph can be beautifully tidy while its family tree is a mess.

A text watermark is a statistical pattern in word choices. It is not a visible stamp tucked into the paragraph. Anthropic’s method explanation describes using otherwise ordinary choices between plausible words to leave a pattern that a matching checker can recognise.

That idea has a research history. University of Maryland researchers demonstrated a word-selection watermark and a statistical detector in work first published in 2023. Their experiments used an older OPT model. The paper’s discussion of constrained wording and edits helps explain the problem; it does not validate the accuracy of today’s Claude, OpenAI or Grok systems.

Anthropic’s explanation of proofreading makes the authorship boundary concrete. A mark can appear after proofreading or translation, even when the underlying material came from elsewhere. Detecting processing would not tell our editor who devised the workshop, checked its entrance or decided what to promise. Those are separate parts of making something publishable.

Anthropic describes embedded text marks separately from signed file records called Content Credentials. Its free file checker checks Claude-issued credentials; text detection remains in private preview for eligible organisations. A public upload box for files does not provide a public test for pasted words.

The C2PA standard explains what a signed file history contributes: a check on the record’s integrity and its connection to the file. It does not certify that the event depicted is true. A valid record can help an editor follow a photograph’s history while leaving the reporting question open.

For an image, Full Fact has already followed a real version of this handoff. Journalist Sian Bayley reported five apparently genuine political images flagged by Google’s SynthID Detector. The team contacted the people who published them and Google. It described likely false positives, while leaving possible inadvertent minor editing unresolved.

Its retained tram-image comparison records two different findings. The first check detected SynthID; the later one returned uncertainty. The tool changed, but the need to establish the image’s history remained. Full Fact did not estimate how often false positives occur. This is its image investigation, separate from our fictional text example.

That leaves our editor with a practical access question. OpenAI’s 5 October announcement opened applications for a text detector initially limited to approved researchers and expert organisations. Selected API models could opt in to marking worldwide, off by default. Eligible EU ChatGPT and Codex output was described as a phased rollout over the following weeks. That dated plan alone does not establish which individual passage is marked today.

Anthropic likewise lists eligible organisations for its text-detection preview. Its marking support varies by model and surface; the current table still schedules completion of some Bedrock rollouts by 10 November. Eligibility is not a detector result, and a file checker does not remove the text-access gate.

The distinction also appears in Grok’s API documentation. It schedules default marking for listed models from 7 November, tomorrow at this reporting date. A response header can record whether generation used watermarking. That is useful evidence from the sender’s side; it is not a later detector finding on edited copy.

Our editorial view is that provenance becomes more useful when it travels as a record of decisions, rather than a verdict squeezed into a badge. The detector’s exact words, the version checked and the date belong together. “We could not run the check” and “the check was uncertain” tell the next editor different things.

For the fictional notice, the present state is simply not checked. We found no eligible retained current-provider text workflow to reproduce in the inspected sources, and obtained no detector result ourselves. The open question is what happens when an eligible publisher retains the original passage, its edited version and the actual returned findings.

An editor still has work to do in the meantime: follow the source, identify the changes and take responsibility for the copy that reaches a reader. Invisible ink may add a useful clue. It cannot sign off the workshop’s opening time.