4 November 2026 · 7 min read

Where AI Detection Is Heading: Watermarks, C2PA and Detector Limits

AI detectors give probabilities, not proof. How watermarks and C2PA provenance differ, what the EU now requires of generators, and what to do today.

AI detection appears to be moving from guessing to labelling. Detectors that examine text or pixels give an estimate, and they can be wrong in both directions. So more effort is going into provenance, which records where a file came from, and watermarks, which are built into content when it is generated. EU rules now push AI providers towards machine-readable marking too. None of this will settle every case, since many files carry no signal at all, so for now the sensible approach is to combine several checks and treat each result as evidence, not a verdict.

Why can't AI detectors give a definite answer?

A detector looks for statistical patterns that tend to appear in AI output. Nothing in a sentence or a pixel inherently proves how it was made. Human writing can look formulaic, AI writing can be edited, and image compression can wash out patterns. That is why a result is a probability, and a high score is not a finding of fact.

A 2023 paper shows the stakes: 'GPT detectors are biased against non-native English writers', by Weixin Liang and colleagues, found that several widely used detectors consistently misclassified non-native English writing as AI-generated, and that simple prompting could also help text evade them. Tools have changed since then, and results vary, but the authors' caution about using detectors in evaluative or educational settings still deserves attention.

What is the difference between detection, watermarks and provenance?

How do C2PA Content Credentials work, and what can't they prove?

Content Credentials are the C2PA standard's way of recording provenance. The C2PA describes them as a cryptographically secure way to capture and express where content came from. They can show that an image was generated by a certain tool, or that a photo has not been changed since capture. It is more like a signed label than a detector.

• They report origin, not truth. A real camera can photograph a misleading scene, and a signed file can still deceive through its context.
• They can go missing. The C2PA's own FAQ acknowledges that metadata can be removed, inadvertently or intentionally. A screenshot or re-save is often enough.
• They need support. Value grows if cameras, editors, generators and platforms all preserve and display them, and support is uneven.

The C2PA FAQ also describes 'soft bindings', such as invisible watermarks or fingerprints, which can help find a credential again after it has been separated from a file. That is why metadata and watermarks are often discussed together rather than as rivals.

Will watermarks solve the problem?

Probably not alone. Google's SynthID is a well-known example: Google says it embeds invisible watermarks in AI images, video and audio, and adjusts word probabilities to mark text. It also offers ways to check for that watermark. But a check like that can only find its own signal. A negative result does not mean a file is human-made; it may come from a generator that does not use that watermark. Open models that people run themselves may not add one at all.

So watermarks could make a growing share of content easier to identify, especially from major providers, while leaving a gap that will not necessarily close. How robust they will be against determined removal is an open question, and worth reading vendor claims carefully.

What does the EU require for marking AI content?

Article 50(2) of the EU AI Act requires providers of generative AI systems to mark outputs in a machine-readable format so they can be detected as artificially generated or manipulated. The Commission says the transparency obligations apply from 2 August 2026. Following the AI Omnibus amendments, its AI Act Service Desk shows a transition for systems placed on the market before that date: their providers have until 2 December 2026 to comply with Article 50(2).

• The Commission published a final Code of Practice on marking and labelling AI-generated content on 10 June 2026. It is voluntary, and the Commission reports about 190 signatories by the end of July.
• This duty falls on providers of the AI systems. Separate disclosure rules apply to people who professionally publish deepfakes.
• How well this works in practice is yet to be seen, including how providers outside the EU respond, how marks survive platforms, and how enforcement develops.
• This is a summary, not legal advice. Check the Commission's pages for the current position.

Where could AI detection go next?

These are directions to watch, not predictions:

• More layered marking, with metadata, watermarks and fingerprints used together so one lost signal does not end the trail.
• Provenance for real photos as well, so genuine captures can carry evidence of origin rather than AI content alone being labelled.
• More platform labels, which depend on signals surviving upload and on platforms choosing to show them.
• Checks that combine signals, reporting what was found rather than a single 'AI or human' score.
• A shifting burden of proof: 'show me your source' may matter more than 'what does the detector say?'

What should you do today?

Skrubly's two checkers deliberately do different jobs. The AI Content Detector estimates from patterns in text, so it can be wrong. The AI Image Checker does not judge pixels: it reports C2PA and generator tags embedded in the file, so it can only speak when those tags are there. Neither is a final answer, and the same goes for any tool.

Quick answers

• Will AI detectors become obsolete? We do not know. They may remain useful as one signal, especially for content with no marking at all.
• If an image has Content Credentials, is it trustworthy? It tells you about its recorded origin and edits. You still need to judge the context and the source.
• Can a watermark prove that something is human-made? No. A missing watermark only means that the check did not find that watermark.