Home/AI marking
Original research · AI Act Art. 50

Which AI tools mark their output?

From 02 December 2026, providers of generative AI must mark their output machine-readably as AI-generated. We reviewed 28 tools: 4 make a public statement about it, 24 do not.

Researched as of: 15 August 2026 · How we collect this · Check your own file

28tools in the dataset
4with a public vendor statement
3embed C2PA metadata
3embed a SynthID watermark (fully or partly)
Media type
28 tools
ToolMedia typeC2PA metadataWatermark (SynthID)Source
DALL·E 3 (ChatGPT / API)ImageYesYesVendor doc · checked 2026-08-15
ElevenLabsFree-tier text-to-speech only; expansion to all audio output announced.AudioYesPartlyVendor doc · checked 2026-08-15
Adobe FireflyImageYesNo statementVendor doc · checked 2026-08-15
Google Flow (Veo)VideoUnclearYesVendor doc · checked 2026-08-15
Canva AIImageNo statementNo statement
CaptionsVideoNo statementNo statement
DescriptAudioNo statementNo statement
Hailuo AIVideoNo statementNo statement
HeyGenVideoNo statementNo statement
IdeogramImageNo statementNo statement
InVideoVideoNo statementNo statement
Kling AIVideoNo statementNo statement
KreaImageNo statementNo statement
Leonardo.aiImageNo statementNo statement
Luma Dream MachineVideoNo statementNo statement
MagnificImageNo statementNo statement
MidjourneyImageNo statementNo statement
Murf.aiAudioNo statementNo statement
PhotoroomImageNo statementNo statement
PikaVideoNo statementNo statement
Play.htAudioNo statementNo statement
RecraftImageNo statementNo statement
RunwayVideoNo statementNo statement
Stable DiffusionImageNo statementNo statement
SunoAudioNo statementNo statement
SynthesiaVideoNo statementNo statement
UdioAudioNo statementNo statement
VEED.ioVideoNo statementNo statement
Free · no account

The AI labelling briefing

What Article 50 actually requires and what it does not, the verified vendor statements verbatim, and a practical guide up to 2 December 2026. We send you the link by email.

Both optional — you get the briefing either way.

We use your address to send you the link. More in our Privacy policy.

The raw data needs no sign-up: download the CSV directly.

Context

No, you don’t have to label every AI text

The AI Act’s transparency duties have applied since 2 August 2026 — and since then the idea has spread that every AI-generated sentence now needs a label. The law does not say that. Article 50 distinguishes cleanly between two addressees and several cases:

Providers of generative systemsmust mark their output machine-readably (metadata, watermark). Transition period for existing systems: 2 December 2026. That is exactly what the table above tracks.
Chatbot deployersmust disclose that people are talking to an AI — unless it is obvious anyway.
Deepfakesthat is, convincingly realistic image, audio or video content of real people, places or events, must be disclosed as artificially generated.
Text of public interestmust be disclosed when published without human review and without editorial responsibility.
Not coveredare internal documents and content where a human reviews editorially and takes responsibility — AI as a tool in a workflow does not by itself trigger a labelling duty.

Not legal advice. This page reflects the state of our research and does not replace legal assessment of an individual case. The text of the Regulation prevails.

Does your file carry a marking?

Our free checker reads the C2PA metadata from an image, video, audio file or PDF — no sign-up, no upload: the file is read entirely inside your browser.

Check a file
Methodology

How this dataset is produced

For every tool we look for the vendor’s public documentation — help centre, product page, blog announcement — and record what it actually says. Every entry other than “no statement” carries a source URL and a check date. Where nothing is findable, the field stays empty.

What the status values mean

YesThe vendor publicly states that it marks its output this way. Source linked.
PartlyMarking applies only to certain outputs, tiers or products — the limitation is stated in the row.
NoThe vendor explicitly states that it does not mark. Only assigned with evidence.
UnclearStatements exist but do not unambiguously cover the case.
No statementWe found no public statement. Not a verdict on the tool.

Limits of this edition

  • This first edition evaluates public vendor documentation. We have not yet measured the outputs ourselves — every row is flagged “vendor doc”, none as our own measurement.
  • We cannot independently verify SynthID watermarks: Google’s detector portal remains limited to journalists and researchers. The SynthID column therefore only reproduces vendor statements.
  • C2PA metadata does not survive many everyday paths — screenshots, messengers, re-uploads, format conversions. A file without metadata is no proof of “not made with AI”.
  • Vendors are changing their practice right now. Every entry carries a check date; older means less certain.

Related datasets: AI Tool Monitor · Price index · Open data

Correct an entry about your tool? We are happy to add evidence and correct quickly — send us the source via our contact page. The dataset is published under CC-BY-4.0 (KI-Radar (ki-radar.net)) and may be reused freely.

FAQ

Frequently asked questions about AI marking

Do I have to label every text I wrote with ChatGPT as AI-generated?

No. Article 50 of the AI Act does not require blanket labelling of all AI content. The disclosure duty for text covers AI-generated text published to inform the public without human review or editorial responsibility. Anyone who reviews a text editorially and takes responsibility for it is not covered. Internal documents such as emails or meeting notes are outside the scope anyway.

What exactly has to be marked from 2 December 2026?

Article 50(2) of the AI Act addresses providers of generative AI systems: they must mark their output machine-readably as artificially generated, for example via metadata or a watermark. For systems already on the market before 2 August 2026, the transition period ends on 2 December 2026. So this duty falls on OpenAI, Google or Midjourney — not on the people using those tools.

Is a visible note such as “created with AI” enough?

For a deployer’s transparency duty towards people, a clearly visible note can suffice. It does not replace the machine-readable marking that providers of generative systems must additionally implement. A visible label and a machine-readable mark are two separate requirements addressed to two different parties.

Why does it say “no statement” for so many tools?

“No statement” means we found no public vendor statement about marking its output. It explicitly does not mean the tool does not mark. We do not assert anything about a company that we cannot evidence — so the field stays empty until a source or our own measurement exists.

Does the marking survive when I send an image?

Often not. C2PA metadata is attached to the file and can be removed by platforms, image editors, format conversions or a screenshot. Watermarks such as SynthID are more robust in this respect because they sit inside the content itself. A file with no findable marking is therefore no proof that it was not made with AI.