Caption
Submitted by a reader; content credentials name ChatGPT as the source.
It looks like a photo. The file says otherwise.
Staimp checks images and video before you publish: credentials, signs of AI and the AI Act duties that may apply.
Sample: an AI-generated image, scanned on 22 Oct 2026. Not proof of origin.
Article 50 already applies to what you publish.
Date Article 50 of the EU AI Act Article Status · today 28 Oct 2026
02 Aug 2026 Disclose realistic AI images and deepfakesWho: deployers who generate or edit the image with AI 50(4) In forcefor 87 days
02 Aug 2026 Machine-readable marking of AI-generated imagesWho: providers of the AI tool that made it 50(2) In forcefor 87 days
02 Dec 2026 AI tools launched before August must start marking their imagesWho: their providers, after a grace period under the Digital Omnibus on AI 50(2) 35 daysleft
Penalty Up to €15 million or 3% of worldwide annual turnover, whichever is higher; for SMEs, whichever is lower.Member States set the penalties within these limits. 99
Sources: Regulation (EU) 2024/1689 · Digital Omnibus on AI, (EU) 2026/1744
Provenance
The file says where it came from.
fox.png carries C2PA content credentials. Staimp reads them: which tool made the file and what was done to it. The signature itself is not verified yet. Now try the same picture as a screenshot.
fox.png · 1536×1024 · PNG · sent in by a reader
Content credentialsFOUND
statuspresent-valid
claim generatorOpenAI Media Service API
software agentChatGPT
actionscreated · converted · watermarked
sourcedeclared AI-generated
Metadata
No camera ever touched it.
No EXIF or XMP at all: no camera, no lens, no capture time. On its own this proves little, because many tools strip metadata. Together with the credentials, it is a second signal.
Image metadataNO EXIF / XMP
camera—
lens—
date taken—
xmp—
AI signals
Seven signals.
One likelihood.
Staimp runs up to seven signals (two are not active yet) and fuses the ones that return a result into a single likelihood with a confidence range. A signal with nothing to go on abstains instead of guessing. It is an automated indicator, not a certification.
- Content Credentials98%
- Camera metadata55%
- Generator tag97%
- Known-image matchno result
- Image artifacts100%
- AI image detectorno result
- External AI checkno result
AI likelihood98%confidence interval 93%–100% · 4 of 7 signals fused · scan of fox.png, 22 Oct 2026
Duties
Scan done.
Four checks before you publish.
The file's own Content Credentials declare it AI-generated: that is the AI provider's machine-readable marking, and readers cannot see it. Whether you also need a visible label depends on how you use the image. The checks come in order of priority.
- P0Check whether a deepfake disclosure is required.If readers could take it for a real photo of a real place or event, Art. 50(4) asks you to disclose it. If so, put a clear, visible "AI-generated" note next to it wherever it runs.Art. 50(4)
- P0Keep the Content Credentials when you publish.Check that your CMS or CDN does not strip them. Some editing and publishing workflows can strip embedded credentials.Recommended practice
- P1Show the Content Credentials to readers.Where your site supports it, show the badge (tool, author, date) next to the image.Recommended practice
- P2Archive this report.Keep the report ID and PDF in your compliance archive.record
Automated preliminary assessment only, not legal advice. Confirm obligations with qualified counsel before publishing.
Caption
The caption goes in. The page can run.
Your call, with a record of what was checked. Try it: switch the disclosure on.
Not just newsrooms.
Agencies and online shops ask the same question before an image goes out. Every image on this page was generated with AI. The originals carry content credentials; the web copies here are resized and lose them. We label our samples the way we suggest you label yours.
Publish
The page can run.
Scan before you publish.
From 2 December 2026, in 35 days, AI tools launched before August must mark their images too.
€39 per month, excl. taxes
100 scans a monthEverything in Free, plus downloadable PDF reports and email support. For professionals who publish regularly.
Paid plans renew automatically every month until cancelled. Cancel anytime in your account; the plan stays active until the end of the paid period. Prices exclude applicable taxes: VAT or sales tax, where applicable, is calculated at checkout. Scans reset monthly and don't roll over. Consumers can withdraw within 14 days after the contract starts.
Article 50, in plain terms
What the law asks, what Staimp does, and what it does not. Source: Article 50
Since 2 August 2026. Generative AI systems already on the market before that date have until 2 December 2026 to add machine-readable marking.
Disclosure may be required if the image is a deepfake under Article 50(4): content that resembles real people, objects, places or events and could falsely appear authentic (Art. 50(4) and 3(60)). Not every AI-edited image meets that definition; a retouched product shot, for example, may not.
Mainly the provider of the AI tool that generated the image (Art. 50(2)). Staimp checks whether that marking is present.
For evidently artistic, creative or satirical work, disclosure can be done in a way that doesn't spoil the work.
No tool can. Staimp shows what it found and which duties may apply, so you can act and keep a record. For legal decisions, consult a qualified lawyer.