EU AI Act: How to Prove Your Work Is Human-Made
To show your work is human-made, you need two things the AI Act will never give you: evidence of how the work was made (working files, drafts, version history, recordings) and a trusted date showing when you held it. Since 2 August 2026, the EU's transparency rules require AI-generated content to be marked and, in some cases, labelled. Human work gets no label at all. So when a client, a platform or a competition jury asks "did AI make this?", the answer has to come from you.
This guide explains what the AI Act actually requires, why "no AI label" will never count as evidence, and what to keep for each type of work: images, photos, music, writing, video and code.
This is general information, not legal advice for a specific dispute.
What the AI Act requires since 2 August 2026
The transparency rules sit in Article 50 of the AI Act (Regulation (EU) 2024/1689). According to the European Commission, they apply from 2 August 2026. They create duties for two groups, and neither of them is you as a creator.
- AI providers (the companies behind image, audio, video and text generators) must mark their output in a machine-readable format so it can be detected as artificially generated (Art. 50(2)). Tools that only perform "an assistive function for standard editing" are exempt.
- Deployers (businesses that publish AI content) must disclose deepfakes. For evidently artistic, creative, satirical or fictional works, the disclosure can be light, "in an appropriate manner that does not hamper the display or enjoyment of the work" (Art. 50(4)).
- AI-generated text published to inform the public on matters of public interest must be disclosed, unless it has gone through human review or editorial control and someone holds editorial responsibility (Art. 50(4)).
On 10 June 2026 the Commission published a voluntary Code of Practice on marking and labelling of AI-generated content, together with EU icons for those labels.
What the Act does not contain is just as important. It creates no "human-made" label, no register of human works and no duty for creators to certify that they did not use AI. The rules regulate the machines and the companies using them.
Why "no AI label" is not evidence
It is tempting to assume that once AI content is marked, anything unmarked must be human. That inference fails for three reasons.
- Marks can be lost. Machine-readable marks live in metadata or in the file signal. Screenshots, re-exports, compression and many upload pipelines strip metadata. A missing mark tells you nothing about where a file came from.
- Not every tool adds one. Tools that only do "standard editing" are exempt, a mark can be removed on purpose, and models run privately on someone's own computer are hard for anyone to police.
- Detectors guess. AI detectors classify patterns. They do not reconstruct how a work was made. OpenAI's own AI text classifier, by the company's own evaluation, caught only 26% of AI-written text and wrongly flagged 9% of human text; OpenAI withdrew it in July 2023 for low accuracy. A Stanford study (Liang et al., 2023) found that several popular GPT detectors flagged more than half of essays written by non-native English speakers as AI-generated.
The opposite mistake happens too. In 2024, photographer Miles Astray entered a real photograph of a flamingo into the AI category of the 1839 Awards. It took third place from the jury and the People's Vote before he disclosed it was a real photo and was disqualified. If a jury cannot tell a real photograph from a generated one by looking, your evidence has to come from somewhere other than the finished file.
When you will actually be asked
The law puts no burden on you. Other people will. These are the situations creators run into most:
- Competitions and open calls that require a declaration that the entry is human-made.
- Client contracts with an AI clause: "the contractor warrants the deliverables were not generated by AI".
- Platforms and marketplaces that remove or relabel content after a complaint or a detector flag.
- Publishers, labels and agencies that ask about AI use before signing.
- Copyright disputes. In the EU, copyright protects a work that is "the author's own intellectual creation" (CJEU, Infopaq, C-5/08) reflecting the author's "free and creative choices" (CJEU, Painer, C-145/10). A purely machine-generated output has no human author. If someone copies your work and argues "this was AI anyway", your human contribution becomes the point in dispute.
The two kinds of evidence you need
Think of the evidence in two layers. Each answers a different question.
Process evidence answers "how was it made?" Layered files, sketches, RAW files, stems, drafts, commit history and screen recordings show the steps a person took. A generated file arrives finished. A human work leaves a trail of intermediate states, mistakes and corrections.
A trusted date answers "when did you have it?" Your own files carry dates you can change. Your computer's clock can be set to any year. A qualified electronic timestamp under the eIDAS Regulation is different: Article 41(2) gives it a legal presumption that the date and time are accurate and that the data has not changed since, and Article 41(3) requires every EU member state to recognise it.
Neither layer is enough alone. Process files without a trusted date can be accused of being made after the fact. A dated file without any process behind it shows only that the file existed on that day. Together they let you say: "here is how I made it, and here is independent evidence that I had it, with my declaration, on this date".
What to keep, by type of work
This is the checklist we give StampR users. Keep the strongest items for every piece you would mind losing a dispute over.
| Type of work | Strongest process evidence | Also useful | What to date |
|---|---|---|---|
| Illustration, design | Layered file (PSD, Procreate, CLIP, Figma) and the app's timelapse recording | Sketches, rejected variants, reference board | The layered file and the final export |
| Photography | RAW files from the camera, with the frames before and after the shot | Contact sheet, edit history in Lightroom or Capture One | The RAW file and the edited JPEG |
| Music | DAW project with stems, MIDI and automation | Voice memos, early bounces, lyric drafts | The project file and the master |
| Writing | Drafts with track changes or version history | Notes, outlines, editor correspondence | Each major draft and the final manuscript |
| Video | Project timeline with raw footage | Storyboard, script versions | Project file and final render |
| Code | Git history with commits over time | Issue tracker, design notes, review comments | A ZIP of each release, or a manifest of fingerprints |
Three habits make this list work in practice:
- Turn on recording before you start. Procreate, Krita and most drawing apps can record a timelapse automatically. It is the single most persuasive piece of process evidence for visual work.
- Never delete the working file. Archive it next to the export. Storage is cheap; a lost PSD cannot be recreated.
- Date at milestones. The day you send a draft to a client, submit to a competition or publish, timestamp what you are sending. A record made before anyone asks carries more weight than one made after a dispute starts.
If you used AI for part of the work
Most working creators now use some AI: noise reduction, upscaling, autocomplete, a generated background, a code assistant. That is legal, and the AI Act does not ask you to stop.
What matters is accuracy. If you declare a work as entirely your own and a client later finds a generated element, the inaccurate declaration does more damage than the tool use ever would. Describe what you did yourself and what a tool contributed. Keep the process files that show your part. A precise declaration is easier to defend than a broad one.
Where a StampR certificate fits
A StampR certificate does one job in this picture: it gives your file and your declaration of authorship a qualified eIDAS timestamp from BORICA B-Trust, a provider on the EU Trusted List. The certificate records the SHA-256 fingerprint of your file, your name and your declaration, and anyone can check it on the verification page or through its QR code.
Be clear about its limits. A certificate cannot tell whether a work was made by a human or by AI. No certificate can. Your process files carry that. The certificate fixes the date on which you held the exact file and declared it as yours, which is what process files cannot do for themselves.
What a declaration of authorship is, and when it carries weight.
Related reading
- How to Prove My Art Is Mine, Not AI
- How to Prove When a File Was Created
- How to Protect Your Art From AI Training
- How to Timestamp Source Code and Git Repositories
Frequently asked questions
Does the EU AI Act require me to prove my work isn't AI?
No. Article 50 places its duties on AI providers and on businesses that deploy AI. It creates no duty for creators to certify human authorship. Clients, platforms and competitions can still ask for evidence under their own rules or contracts.
Will my human-made work get a "human" label under the AI Act?
No. The Act and the Commission's Code of Practice only define marks and labels for AI-generated or manipulated content. There is no official label for human work.
Can an AI detector show that my work is human-made?
Not reliably. Detectors estimate patterns and produce false positives, especially for non-native writers and compressed images. Process evidence such as working files and recordings is far more persuasive.
Can a timestamp certificate show that a work isn't AI?
No. A certificate shows that an exact file existed unchanged at a specific time, together with your declaration. Human creation is shown by your process files. The two together make a much stronger record than either alone.
I used AI tools for part of the work. Can I still declare authorship?
Yes, for the parts you created. Describe accurately what you made and what a tool did. An inaccurate declaration is a bigger risk than the tool use itself.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 50: transparency obligations for providers and deployers.
- European Commission, Code of Practice on marking and labelling of AI-generated content (10 June 2026).
- Regulation (EU) No 910/2014 (eIDAS), Article 41: legal effect of qualified electronic timestamps.
- CJEU, Infopaq, C-5/08 and Painer, C-145/10: originality as the author's own intellectual creation.
- OpenAI, AI text classifier: reported accuracy and withdrawal.
- Liang et al., GPT detectors are biased against non-native English writers (2023).
- PetaPixel, photographer disqualified from AI image contest after winning with a real photo (2024).
Sources checked 2026-11-10.
Certify your work with a trusted date.
Create a certificate →Your first certificate is €4.90 with an account.