AI tools · AI tools

AI content detection and watermarking.

AI detectors estimate, and they err in both directions: human work gets flagged and edited AI output slips through. Provenance is the dependable approach. C2PA Content Credentials attach a signed history to a file, invisible watermarks survive some edits, and IPTC metadata labels AI-made media. The EU AI Act has required machine-readable marking of synthetic content since August 2026, with a grace period to December 2026 for systems already on the market, and other jurisdictions have similar rules.

Why detection is unreliable

Detectors look for statistical traces of generation and report a probability, not a finding. They produce false positives, flagging formulaic, heavily edited or non-native human writing and retouched photographs, and false negatives, missing AI output that has been paraphrased, re-compressed, resized or screenshotted. Two detectors often disagree about the same file, and at least one major model developer withdrew its own text classifier because its accuracy was too low. Use a detector score as a reason to ask questions, never as evidence on its own, and never as the basis for accusing a freelancer, supplier or employee.

Provenance: credentials, watermarks and metadata

C2PA Content Credentials attach a cryptographically signed manifest to a file, recording who or what created it, which tools were used and what edits followed; anyone can inspect it with a compatible viewer. The weakness is that it lives in metadata, and uploads, screenshots and some compression pipelines strip it. Invisible watermarks embed a signal in the pixels or audio itself, survive a fair amount of compression and resizing, and can be used to recover a stripped manifest, but usually only the watermark's owner can read them. IPTC's Digital Source Type field, with values such as trainedAlgorithmicMedia, is the simple label several platforms read.

What regulators expect

Article 50 of the EU AI Act requires providers of generative AI systems to mark synthetic audio, images, video and text in a machine-readable format so they are detectable as artificially generated, using solutions that are effective, interoperable and reliable as far as technically feasible; deployers must separately disclose deep fakes. The obligations have applied since 2 August 2026, with marking due by 2 December 2026 for systems already on the market, and a voluntary code of practice published in June 2026 shows how to comply. China's labelling measures, in force since 1 September 2025, require visible labels and embedded metadata. California's AI Transparency Act, operative since 2 August 2026, requires covered generative AI providers to embed latent disclosures and offer a free tool for checking content.

A practical set-up for brands

Brands are mostly deployers, so their job is to keep marking intact and disclose where needed. Check that the DAM, CMS, image optimisation and video transcoding steps preserve C2PA and IPTC metadata rather than silently stripping it. Decide per channel whether a visible label is needed or metadata suffices, and write that into the brief. Then keep an internal record that survives whatever platforms do to files. Synthetic White keeps a per-asset record of which model made what, so the answer to how an asset was made does not depend on metadata that a social platform may already have removed.

Updated 25 September 2026

Questions

AI content detection and watermarking, answered.

Can AI detectors tell whether an image or text was made with AI?

Not reliably. They output probabilities that vary between tools and fall apart after editing, compression or paraphrasing, and they flag some human work as machine-made. Treat a score as a prompt for questions, never as evidence. Provenance records and credentials attached at creation are far more dependable.

What are C2PA Content Credentials?

An open technical standard from the Coalition for Content Provenance and Authenticity for attaching a signed record to a file: who or what created it, which tools were used and what edits followed. The record is lost if metadata is stripped, so it is often paired with an invisible watermark that can restore the link.

Does the EU AI Act require watermarking?

It requires providers of generative AI systems to mark outputs in a machine-readable, detectable way from 2 August 2026, or 2 December 2026 for systems already on the market, without prescribing one technique; watermarks, metadata and fingerprinting are all candidates. Deployers must separately disclose deep fakes. Brands using a provider's tools should confirm the marking is applied and survives their own pipeline.

Get started

See your AI studio generating this week.

Thirty minutes. A live studio in your colours, your brand hub loaded, a real campaign brief.

See plans