In short

  • No single test is certain. Images from the major generators now carry several independent signals, such as signed metadata, an invisible watermark and the provider's own records, plus a visible label when someone adds one. Removing one rarely removes them all.
  • The law asks for layers. Since 2 August 2026, Article 50(2) of the AI Act requires generative-AI providers to mark their output in a machine-readable, detectable way (systems already on the market have until 2 December 2026). The Commission's Code of Practice asks its signatories for at least two layers and a free detection tool.
  • You can check for free. Google's SynthID Detector, open to everyone since 7 October 2026, reads the watermark used by Google, OpenAI, NVIDIA and Kakao. OpenAI, Meta and Midjourney run checkers of their own.
  • Every check can fail. A screenshot strips metadata, deliberate attacks can weaken watermarks, and AI detectors and human eyes both make mistakes, in both directions.
  • A missing mark proves nothing, and a stripped label doesn't make an AI image human-made. Labelling your own AI images is the honest choice: add the EU AI image label in your browser.

Can AI-generated images be detected?

Often, yes. Always, no. Each method works in some cases and fails in others.

Spotting AI used to mean looking for odd hands. Today's models show “hardly any obvious errors”, says Switzerland's National Cyber Security Centre, so the major providers now mark images when they are made, in several places at once: in the pixels (a visible label or an invisible watermark), in the file (signed Content Credentials or an IPTC tag) and at the provider (records its own detector can match).

These layers are made by different parties and break in different ways: a screenshot removes the metadata, while watermarks are designed to survive one. That is why an AI image is hard to hide. Hiding it means defeating every layer, and one surviving signal is often enough to raise the question.

Why this page doesn't say “always”

It would be easy to say AI is always detectable. The evidence doesn't support it: the EU Code of Practice itself says forensic detection of unmarked content was “not deemed mature enough” to meet the law's quality requirements when it was published. Several independent signals make detection likely; none makes it certain.

The layers that give an AI image away

Each layer lives in a different place and fails in a different way. None is enough on its own.

Layers of AI image detection compared
LayerWhere it livesWho reads itWhat defeats it
Visible label (the EU icon, a platform's “AI info”)On or next to the imageAnyone, with no toolCropping, painting over, or never adding it
Content Credentials (C2PA)In the file: a signed record with a hash of the imageC2PA viewers, LinkedIn, Google Search, the Gemini app, TikTokScreenshots, re-encoding, most social platforms; any edit breaks the signature
IPTC digital source typeIn the file: an unsigned “made with AI” fieldMeta's platforms, Pinterest, Google Search, metadata viewersScreenshots and re-saving; anyone can edit it
Invisible watermark (SynthID, Content Seal, TrustMark…)In the pixels, imperceptibleThe provider's own detector; SynthID Detector also reads its partners' marksDeliberate attacks; heavy cropping, for some; a detector only finds the marks it knows
Fingerprints and logsAt the providerThe providerOnly finds images registered when made; can be removed or forged
AI detectors (classifiers)Statistical traces in the pixelsAnyone with a detectorNew generators, compression, resizing; real photos get flagged too
Human judgementThe scene and its contextYouRealistic images; quick glances on a phone

Provenance works both ways: some cameras sign real photos. The Leica M11-P was presented in 2023 as the first camera with Content Credentials, and every JPEG from the Google Pixel 10's camera carries them.

Do AI-generated images have a watermark?

Most images from the major generators now do, usually an invisible one. Here is what each provider says it adds:

  • Google: SynthID, designed to survive “cropping, adding filters, changing frame rates, or lossy compression”, plus C2PA on Nano Banana Pro images.
  • OpenAI: C2PA since 2024, and since May 2026 SynthID as well, on images from ChatGPT, Codex and its API.
  • NVIDIA uses SynthID on Cosmos video, and Kakao announced it for its image and video models from the second half of 2026.
  • Meta: its “Content Seal” watermark in every Muse Image picture, with a preview detector.
  • Adobe: Content Credentials plus TrustMark, an open-source watermark, and a signed fingerprint.
  • TikTok: testing an invisible watermark that only TikTok can read, on content from its AI tools and on uploads carrying C2PA.
  • Midjourney: a hidden ID tag in every image, according to its help centre, with a free checker.
  • Black Forest Labs (FLUX): C2PA on its API and hosted tools, a watermark only optionally.

China has required visible and hidden labels on AI content since 1 September 2025. The gap: open-weight models run on your own computer may add no mark at all, so no checker can tell you an unmarked image is real.

What the providers say

What the EU law requires of AI providers

Article 50(2) of the AI Act (Regulation (EU) 2024/1689) requires providers of generative AI to make sure their outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated”, with solutions that are effective, interoperable, robust and reliable as far as technically feasible. It has applied since 2 August 2026; generative systems already on the market before that date have until 2 December 2026.

The Commission's Code of Practice on Transparency of AI-Generated Content (final version, 10 June 2026) makes this concrete. The Commission lists 95 signatories of its providers' section, who commit to:

  • at least two layers on images: signed metadata and an imperceptible watermark. Fingerprinting or logging alone “is not considered sufficient”;
  • no removal: intentional removal of marks banned in their terms of use;
  • a free detection tool, open to the public where the public may see the content, and always free and unlimited for regulators, media, fact-checkers, researchers and civil society;
  • robustness tests against recompression, screenshots, cropping, resizing, print-and-scan, and attacks that try to remove or regenerate the mark;
  • interoperable detection of watermarks by 2 February 2027.

The visible label is a separate duty, for deployers who use AI in a professional capacity and publish deep fakes (Article 50(4)), and a provider's machine-readable metadata doesn't satisfy it. Our guide to the EU AI labelling law explains who must label, and when.

  1. Code of Practice published

    At least two marking layers and a free detector, for the providers who sign it.

  2. Article 50 applies

    Providers mark AI output; deployers disclose deep fakes.

  3. SynthID Detector opens to everyone

    It reads the watermark used by Google, OpenAI, NVIDIA and Kakao.

  4. Marking deadline for older systems

    For generative AI placed on the market before 2 August 2026 (Digital Omnibus on AI, Regulation (EU) 2026/1744).

  5. Interoperable detection

    Watermark detection that works across providers, under the Code.

Official texts

How to check if an image is AI-generated

No single step gives a verdict. Note what each one finds, and weigh them together.

  1. Look for a visible label

    Check the corners and the caption for the EU icon or a platform label such as Meta's “AI info”. Our guides to the EU AI icons and to AI labels on social media show what they look like. No label tells you very little.

  2. Check the Content Credentials

    Open the original file, not a screenshot, in a Content Credentials (C2PA) viewer: it shows who signed the file and what they declared, such as generation with AI. Google Search's “About this image” and the Gemini app read C2PA too. Finding no credentials often means they were lost on the way or never added; it doesn't mean the image is real.

  3. Run SynthID Detector

    Open to everyone since 7 October 2026, Google's SynthID Detector looks for the watermark used by Google, OpenAI, NVIDIA and Kakao, with Apple announced. Google warns that a negative result doesn't prove an image is real: it only sees its own and its partners' marks.

  4. Try the generator's own checker

    OpenAI's Verify, a research preview, only recognises OpenAI images and makes no definitive claim when it finds nothing. Midjourney's checker reads its hidden ID tag, which screenshots, edits and most social sites remove.

  5. Find where the image came from

    A reverse image search can lead you to earlier copies, to the account that first posted the image, or to a version that still carries its label or credentials.

  6. Look at the scene and its context

    Text, shadows, repeated elements, merged objects and misaligned buildings still sometimes give an image away; hands, less and less. Ask whether the whole scene makes sense, who posted it and why. An AI detector can add a second opinion, never a verdict.

A result is a clue, not a proof

A watermark found by its own provider's detector is a strong signal. A missing mark, a detector's score or your own impression needs backing from another layer, especially before you accuse anyone.

Why every detection method can fail

Metadata is easy to lose. A screenshot removes Content Credentials and IPTC tags, and most social platforms strip them on upload: OpenAI called C2PA “not a silver bullet”. The C2PA's answer is “soft bindings”, watermarks or fingerprints that let a tool find the credentials again online.

Watermarks can be attacked, with effort. Regeneration attacks rebuild the image with an AI model. UnMarker, presented at the IEEE Symposium on Security and Privacy in 2025, needs a few minutes on a high-end A100 graphics card: against seven watermarking schemes, the best remaining detection rate was 43%, and an informal test reported by The Register cut SynthID detection from 100% to about 21%. Heavy cropping matters too: when Reuters cropped 40 Muse images to about a third to a half of their size, Meta's preview detector, which had caught all 40 originals, missed 55% of them. Removal takes deliberate effort, providers must test against it, and detection keeps improving.

AI detectors err both ways. In NewsGuard's audit of 8 May 2026, five detectors got 13.3% of their verdicts on real war photos wrong: one flagged 40% of the real photos as AI, and at least one tool disagreed with the others on 35 of 45 images. In a September 2026 preprint, 12 academic detectors flagged 17% to 91% of real smartphone photos.

People are not reliable judges. In Microsoft's “Real or Not” quiz, about 12,500 people got 62% of their answers right. A December 2025 study measured 54% on Midjourney images, barely better than a coin toss; another found 29% on FLUX.1-dev images. Portraits are easier than landscapes.

Even combinations fall short. A Microsoft Research report (January 2026) found that only 20 of 60 combinations of methods reached “high-confidence authentication”, and warns of attacks that make real content look fake.

What a check can tell you

  • A watermark found by its provider's own detector: a strong sign that provider's AI made or changed the image
  • Valid Content Credentials: who signed the file, and what they declared
  • A visible label: what the publisher says about the image
  • Several independent signals agreeing: a likely answer

What it can't tell you

  • No watermark: maybe another generator, a crop or an attack
  • No metadata: often lost on a screenshot or an upload, or never added
  • A detector's score on its own: real photos get flagged too
  • Your impression on its own: often close to a coin toss

Studies and tests

A missing mark proves nothing: label your own images

A missing watermark doesn't make an image real: it may come from an unmarked model, or have been cropped, screenshotted or attacked. And stripping a label doesn't make an AI image human-made. The provider may keep records, credentials may survive in an online repository, a platform may already have labelled the post, and careful readers still look at the scene.

Hiding where an image comes from is hard and easy to get wrong, and providers who sign the EU Code ban the removal of their marks in their terms of use. Saying it up front is the honest choice, and for deep fakes published in a professional capacity, Article 50(4) requires it: our guide to the deepfake label explains when an image counts as one.

The EU AI image label tool adds the European Commission's icon (AI, AI GENERATED or AI MODIFIED) to your images, in your browser, and can write the IPTC digital source type (“made with AI”, or “edited with AI”) into the file. The EU AI icons guide helps you choose the right one.

What this site does, and what it doesn't

AI Image Label adds the EU icon to your images; it doesn't detect AI. The icon tells your audience the truth, but using it doesn't by itself establish legal compliance: you decide whether and how to label, and you remain responsible.