- Home
- Skills
- Legal & Compliance
- EU AI Act Article 50 Gate: AI Disclosure & Marking Audit
Works with the AI Agents you already use
EU AI Act Article 50 Gate: AI Disclosure & Marking Audit
Checklists written before July 2026 get this wrong in both directions.
$39
EU AI Act Article 50 Gate: AI Disclosure & Marking Audit
Example session with this skill installed
Run the EU AI Act Article 50 gate over this repo. We self-host an open-source chat platform for
~600 staff across our EU offices, on our own domain and branding, with chat and the image
generation feature both switched on. Tell me what we owe and where the gaps are.
- Read your context and instructions
- Compiled the eu ai act
Summary
Three surfaces, two gaps open, and one duty whose answer flips entirely on a single
environment variable. The chat interface ships with no AI disclosure of any kind — the one
notice that exists is commented out in the source. Nearest deadline has already passed:
Art. 50 applied on 2 August 2026.
Stage 1 — Discovery
Manifest scan found an OpenAI SDK, an Anthropic SDK, a local Whisper build and a transformers
package. Scoped call-site grep returned hits across 24 UI components and 16 backend modules.
| # | Surface | Entry point | Modality |
|---|---|---|---|
| S1 | Chat interface | chat component | text |
| S2 | Image generation | image router | image |
| S3 | Text-to-speech | audio router | audio |
Speech-to-text is not a surface: transcription does not generate synthetic content.
Excluded with that basis recorded.
Stage 2 — Role, and why it is not in the code
You deploy this on your own domain and branding for your own staff, so you are at minimum the
deployer on all three surfaces.
S2 is the interesting one. The config module resolves the image engine from an environment
variable:
IMAGE_ENGINE = os.getenv('IMAGE_ENGINE', 'hosted-api-vendor')
Supported engines are two hosted vendor APIs and two local inference servers of the
Automatic1111 / ComfyUI class. The first two call a vendor API. The last two run a diffusion
model you host. Same source code, opposite duties:
| Engine configured | Your role on S2 | Art. 50(2) marking |
|---|---|---|
| hosted vendor API (the default) | deployer | N/A — provider's duty |
| local inference server | provider | APPLIES to you, and nothing in this codebase does it |
I cannot answer this from the repository. Tell me which engine each environment has set, or
read it off your running config. If both are in use, S2 gets assessed twice.
Stage 3 — Classification
| Surface | Duty | Verdict | Basis |
|---|---|---|---|
| S1 | Art. 50(1) | GAP | See below |
| S2 | Art. 50(2) | config-dependent — see Stage 2 | |
| S2 | Art. 50(4) | not triggered on current use | Internal staff tool; no publication of plausibly-real imagery. Re-assess if output is ever published externally |
| S3 | Art. 50(2) | N/A — provider's duty | TTS routes to a hosted vendor |
| all | Art. 50(3) | N/A | No emotion recognition, no biometric categorisation |
S1 — the chat interface discloses nothing
There is no string anywhere in the frontend telling a user they are talking to an AI. The only
candidate sits in the chat component, and it looks like this:
<!-- {translate('LLMs can make mistakes. Verify important information.')} -->
Two problems, and the second survives fixing the first:
- It is commented out. It renders nothing. A grep-only pass would have counted this as a
hit and scored the surface MET. - Even uncommented it is not a 50(1) disclosure. It is a capability caveat, not a
statement of artificiality. The Commission's guidelines reject generic wording that does not
convey the user is dealing with an AI system, and "LLMs can make mistakes" tells a user the
output may be wrong, not that nobody human is there.
The "obvious to a reasonably well-informed person" exemption does not rescue this. Your
audience is 600 general staff, not a professional-developer population, and the guidelines put
the bar higher where the audience may include people with disabilities — which at that
headcount it will.
S2 — no provenance marking exists in this codebase
Searched the whole backend and frontend for c2pa, digitalSourceType, synthid, XMP/IPTC
writes, and any watermarking library. Zero matches. The metadata parameter threaded
through the image upload path is application database metadata — chat id, message id, prompt —
not file provenance.
That is correctly not your problem on the default hosted-vendor engine. On a local
inference server it is a hard gap against a duty that is already live.
Bonus finding — the edit path destroys upstream provenance
The image-edit normalisation helper re-encodes uploads through an imaging library:
image = ImageOps.exif_transpose(image)
if image.mode != 'RGB':
image = image.convert('RGB')
output = io.BytesIO()
image.save(output, format='JPEG', quality=95)
No exif= and no icc_profile= on the save, so any C2PA manifest or XMP the vendor attached
is discarded. It fires only when normalisation is needed — multi-frame images, non-RGB mode, or
a non-default EXIF orientation — so it is intermittent, which is worse than always: your
provenance coverage becomes a function of what the user uploaded.
Not a breach by you as a deployer. But if you ever rely on vendor marking as evidence, this is
the line that silently invalidates it.
Gaps, ranked by exposure
| # | Surface | Duty | Deadline | Why ranked here |
|---|---|---|---|---|
| 1 | S1 | Art. 50(1) | passed 2 Aug 2026 | 600 users, zero disclosure, one-line fix |
| 2 | S2 | Art. 50(2) | passed 2 Aug 2026 if self-hosted engine | Entire duty hinges on your config — resolve the question first |
Remediation
S1. Add a persistent header to the chat component, above the message list, visible without
scrolling and present for the whole session:
{assistant name} · AI assistant
You're chatting with an AI, not a person.
Add it to the i18n catalogue for every locale your offices use, with an aria-label — Art.
50(5) requires accessible disclosure, at the latest at first interaction. Do not reinstate the
commented line and call it done.
S2, first. Establish the configured engine per environment and record the environment
variable as the evidence pointer. Everything else waits on that answer.
S2, if self-hosted. You are the provider and owe two marking layers. Expect the C2PA
signing certificate, not the code, to be the long pole — see
references/marking-implementation.md for the conformance route and the per-modality
watermarking options.
Penalty exposure
Art. 99(4)(g): €15,000,000 or 3% of total worldwide annual turnover, whichever is higher.
At ~600 staff you are outside the Art. 99(6) SME inversion — confirm against your turnover and
balance-sheet figures and record the conclusion.
Next steps
- Answer the engine question. It decides whether gap 2 exists at all.
- Ship the S1 disclosure. Under a day, and the deadline passed two months ago.
- Fill
templates/evidence-log.md: screenshot S1 at first interaction, and record both
N/A — provider's dutyverdicts with reasoning. A log with no 50(2) row reads as missed
rather than assessed. - Note that this is upstream open-source code. Your fixes live in your fork or your
configuration, and you re-test them on every upgrade. Add that to the change-history
triggers.
Readiness assessment, not legal advice. Findings are from a real assessment of a widely used
self-hosted chat platform, anonymised at the project's expense rather than the findings'.
Verify against the Official Journal text and the Commission's current Art. 50 guidelines before
any regulatory filing.
Connects securely to your tools. The creator never sees your data.
About this skill
The deadline is real and it is not the one you read about
Article 50 of the EU AI Act binds every AI system used in one of four situations — not just high-risk ones. It has applied since 2 August 2026.
The Digital Omnibus deferred most high-risk obligations to December 2027 and August 2028. It did not defer Article 50. Checklists written before July 2026 get this wrong in both directions. Two dates are still ahead:
| Date | What lands |
|---|---|
| 2 December 2026 | Art. 50(2) marking + detection for generative systems already on the market |
| 2 February 2027 | Watermark-detection interoperability across providers |
Breach sits at €15,000,000 or 3% of worldwide annual turnover, whichever is higher (Art. 99(4)(g)) — or the lower of the two if you are an SME.
What it does
- Surface inventory — finds every place AI output or AI interaction reaches a human. Starts from your dependency manifests, then scoped call-site greps.
- Role determination — works out, per surface, whether you are the provider or the deployer.
- Classification — four decision trees, one per Art. 50 paragraph, with the exemptions read as narrowly as the Commission's guidelines read them.
- Evidence testing — marks each duty MET / PARTIAL / GAP against what actually exists, with pointers rather than claims.
- Report — gaps ranked by exposure, concrete remediation, drafted disclosure copy, and an evidence log you can hand a supervisory authority.
Why this one and not a checklist
It gets the provider/deployer distinction right, and that is the whole game.
If you call a third-party model through an API, you are a deployer. Article 50(1) and 50(2) are not your duties — they are your vendor's. Most checklists score them against you anyway. Acting on that means implementing C2PA signing, applying to the conformance program and queuing for a trust-list certificate you are not obliged to hold: weeks of engineering against a deadline that does not apply to you.
This skill gates every tree on role first, routes those duties to N/A — provider's duty, and
runs a separate diligence check on whether your vendor actually marks the output reaching your
users.
It also knows where the Act is silent. Fabricated engagement counts, a simulated platform UI, your GDPR position, Chapter III high-risk classification — flagged and handed back, not scored. A padded report costs you credibility on the findings that are real.
Every date, penalty tier and Code of Practice commitment was verified against the Act text, the Commission's own Code-of-Practice FAQ and the omnibus. Not recalled.
What's included
SKILL.md— the five-stage workflow and the guardrailsreferences/article-50-reference.md— operative duties of Art. 50(1)–(7), the Art. 3(60) deepfake definition, Code of Practice commitments, penalty tiers, full date table, sourcesreferences/marking-implementation.md— C2PA manifests and signing, IPTCdigitalSourceTypevalues with JSON examples, watermarking per modality with licences straight, detection-endpoint design, the 2 Feb 2027 interoperability duty, and where your own CDN silently strips your markingtemplates/classification-worksheet.md— four decision trees, surface inventory, deployer reliance check, exposure rankingtemplates/disclosure-copy.md— drafting copy for every surface: chat header, voice, SMS, email, AI agent acting for a principal, image and video labels, audible disclaimers, synthetic voice of a real person, reduced artistic-works duty, public-interest texttemplates/evidence-log.md— the regulator-facing record, with a table of what evidence actually holds up per dutyexamples/worked-example-synthetic-narration.md— a real assessment where the expected duty did not apply and the one that did was the one a naive pass missesLICENSE.md— perpetual commercial use, consultancy and agency use included, no resale. You own every report you generate
Consultancies and agencies: the licence covers you
Run it on client engagements and bill for the work. You own your Output and can grant your client the rights they need, with no attribution and nothing further owed. The only limit is that the skill's own files stay with you — a client who wants to run it buys their own copy.
What it will not do
- It is not legal advice, and it does not certify compliance. It reports readiness per duty with evidence, and tells you to confirm against the Official Journal before any filing.
- It does not cover GDPR, or Chapter III high-risk classification. Both are flagged and handed back.
- It will not implement the marking for you. It tells you precisely what to build and where — and warns you that certificate issuance, not the code, is the long pole.
Works with Claude Code, Cursor, Codex CLI and any agent that loads SKILL.md.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Legal & Compliance.
Not legal advice, and it does not certify compliance. It reports readiness per duty with
evidence and tells you to confirm against the Official Journal before any filing.
Does not cover GDPR, or Chapter III high-risk classification. Both are flagged and handed back.
Role is sometimes not knowable from the code. Where a surface's model is an admin setting or
environment variable — a hosted vendor API in one deployment, a self-hosted model in another —
the same source carries opposite duties. The skill stops and asks rather than guessing, so you
will need to answer that question for each environment.
Gateways and frameworks hide the provider. Apps routing through an AI gateway or an abstraction
layer declare no provider dependency at all, with model identifiers living in config as plain
strings. The skill scans for these explicitly, but confirming who actually serves the model is
a question only you can answer.
A static scan can miss dynamic, proprietary or obfuscated call sites, and it cannot see a
surface that exists only in a running deployment. Every hit is treated as a lead to confirm,
not a finding.
It will not implement marking for you. It tells you what to build and where, and warns you that
the C2PA signing certificate, not the code, is usually the long pole.
Contents verified 26 October 2026. Obligations, Commission guidelines and the voluntary Code of
Practice change — re-verify before relying on any output.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 21 days ago
Nothing to install. It runs in your AI agent.
- Optional
- c2patoolFFmpeg
c2patool only verifies C2PA manifests on image, audio and video; FFmpeg only tests whether marking survives re-encoding.
Permissions
File Scopes
Read access: the assessment works by reading your dependency manifests and the call sites that use them. It never needs secrets — it reads .env.example to learn which config keys exist, not .env to read their values. Write access: it writes one report to the project root and fills the worksheet and evidence-log templates. It does not modify your source. Shell: dependency-manifest and call-site scans, and running c2patool or ffmpeg when you ask for the Article 50(2) verification pass. Network: only in that verification pass, and only against your own assets and endpoints. It pulls a generated asset from its public URL rather than from the generator (because your CDN may strip the provenance metadata in between), downloads a delivered file to test whether the watermark survived, and makes one unauthenticated request to your detection endpoint to confirm it is reachable to the parties the Code of Practice says must have free access. No third-party service is contacted and nothing is uploaded anywhere.