Blog
One watermark at a time.
Each post takes a single marking system, platform, or claim and asks the same two questions: what does it actually do, and can anyone outside the company that made it check?
3 November 2026
Can you check SynthID yourself?
Yes, for images, video and audio — but only inside Google's own tools, and still not for text.
Read27 October 2026
Does WhatsApp remove EXIF data?
Yes if you send it as a photo and no if you send the same file as a document — one choice at send time decides whether any of the metadata reaches the other end.
Read24 October 2026
Does Instagram strip image metadata?
Yes — the copy other people can download has no EXIF in it, because Instagram re-encodes every upload and nothing that is not a pixel survives that.
Read21 October 2026
How to check an image's metadata
Open the file in a reader that handles EXIF and XMP and every field it carries — camera, software, timestamps, GPS, credentials — comes back in seconds.
Read16 October 2026
Is there an AI image detector that works?
Pixel-based detectors are near-perfect on the generators they were trained on and much worse on everything else, which is why provenance beats detection on an image you did not make.
Read16 October 2026
How can you tell if an image is AI-generated?
Rarely by looking at it — the checks that work read the file itself, and knowing what each one can and cannot show matters more than any single result.
Read16 October 2026
What is a homoglyph, and why does it matter?
A homoglyph is a character that looks like another one but isn't it — Cyrillic “а” standing in for Latin “a” — and the mismatch quietly breaks domain names, filters and anything that compares text.
Read1 October 2026
Does paraphrasing beat an AI detector?
Usually it does — and what that proves is something about the detector, not about who wrote the text.
Read29 September 2026
Why does ChatGPT output contain U+202F?
ChatGPT sometimes emits a narrow no-break space instead of a normal one, which is a training artefact rather than a watermark and proves nothing about who wrote the text.
Read27 September 2026
How to remove invisible characters from text
Paste the text into a scanner and copy the cleaned version back out, or strip them yourself with one regex — and know what removing them doesn't fix.
Read26 September 2026
How do AI text detectors actually work?
Detectors don't find a mark left in your text — they score how predictable your writing is and report which pile of training data it resembles.
Read25 September 2026
Do em dashes mean AI wrote it?
No — and since November 2025 you can tell ChatGPT to stop using them, which is the end of the em dash as a tell.
Read23 September 2026
Can Turnitin detect ChatGPT?
It flags writing that reads like a language model, but it can't tell ChatGPT from anything else and its score is not proof of who wrote a sentence.
Read22 September 2026
Why was my own writing flagged as AI?
A detector did not find anything in your text — it guessed from your writing style, and plain, careful prose is what it guesses wrong about.
Read21 September 2026
Are AI detectors accurate?
No, and OpenAI withdrew its own for that reason. They guess from writing style, and they guess wrong most often about people who learned English as a second language.
Read12 September 2026
What is data watermarking?
It marks a dataset rather than a piece of content, and it answers one question: whether your data ended up inside someone else's model or someone else's copy of your table.
Read11 September 2026
What forensic watermarking actually means
It means hiding an identifier in each copy of a file so a leak can be traced back to the account it was issued to.
Read10 September 2026
What does the watermark on a check look like?
Tilt it to a shallow angle under a light and look at the back: the watermark on most checks is printed there in transparent ink, not made in the paper.
Read9 September 2026
Is C2PA a watermark?
No — it is a signed manifest attached to the file, which is why it behaves nothing like a watermark.
Read8 September 2026
How to check text for hidden characters
Paste it into a scanner that lists the actual codepoints, read what it found, then copy it back out clean — about a minute, start to finish.
Read7 September 2026
Is there an AI-proof watermark?
No: visible marks get painted over and invisible ones get regenerated away, so the defensible move is provenance you can show rather than a heavier logo.
Read6 September 2026
Why are they called watermarks?
No water is involved: the name was a guess made around 1708 about a mark that looked like a damp patch, and the guess has been confusing people ever since.
Read4 September 2026
What does an AI watermark look like?
Nothing you can see: a real AI watermark is a statistical pattern in the pixels or the word choices, and the badge in the corner is a separate, removable label.
Read3 September 2026
If you copy and paste from ChatGPT, is there a watermark?
No — there's no watermark in ChatGPT's text to copy across, though the odd invisible character can travel with the paste.
Read2 September 2026
An invisible character detector found something. Now what?
Hidden Unicode in pasted text comes from a short list of mundane sources, and working out which one you have is the difference between a useful clue and a false accusation.
Read1 September 2026
Watermarking means four different things
The word covers four unrelated technologies, and the one most people arrive looking for is the only one nobody outside a handful of companies can actually read.
Read31 August 2026
Does a screenshot destroy content credentials?
It destroys the credential entirely and may leave the watermark untouched — and the gap between those two facts explains most of the confusion about AI provenance.
Read