AI Content Inventory — Pre-filled Template
Regulation (EU) 2024/1689 (AI Act), Article 50 — transparency obligations Applicable from 2 August 2026 · Maximum penalty for Article 50 breaches: EUR 15,000,000 or 3% of worldwide annual turnover, whichever is higher.
This document is general information prepared from a self-assessment — it is not legal advice.
1. Why this inventory is the backbone of your Article 50 work
Article 50 does not apply to your company as a whole — it applies channel by channel, system by system. A chatbot triggers a different duty than an AI-generated product photo, and the same piece of AI text can be exempt or not depending on whether a named person genuinely reviewed it. You cannot demonstrate compliance, and you cannot fix gaps, until you know exactly where AI touches what you publish.
That is why A50-ORG-CONTENT-MAP sits underneath everything else: you are expected to know where AI generates or alters anything you publish — website imagery, social posts, chatbots, marketing copy, reports — and to determine, per system, whether you act as provider, deployer, or both. Every other requirement in this document attaches to a row in the table below. When a market surveillance authority asks "where do you use AI and what did you do about it?", this table, kept current and with owners assigned, is the answer.
Your self-assessment flags three findings that make this urgent:
- No visible labels are currently applied to AI-generated content.
- No machine-readable markers are embedded — yet you offer a generative AI product under your own brand, which makes marking a provider obligation for you under Art. 50(2).
- Chatbot disclosure is partial or inconsistent, against a duty under Art. 50(1) that has no "sometimes" setting.
2. Inventory
Legend for requirement IDs: A50-1 chatbot disclosure · A50-2 machine-readable marking · A50-4-DF deepfake visible label · A50-4-PI public-interest text · A50-5 disclosure quality · A50-ARCH archive/reuse · A50-MAP content map.
| Channel / system | AI involvement | Your likely role | Triggered requirement IDs | Current labelling status | Owner |
|---|---|---|---|---|---|
| Customer-facing chatbot / voice assistant | Third-party AI interacting directly with people in the EU | Deployer in practice; Art. 50(1) duty sits with the provider — confirm which you are for this system | A50-1, A50-5, A50-MAP | Gap — users told "partially or inconsistently"; no reliable first-interaction notice | ___ |
| Own-brand generative AI product / feature offered to others | You offer a generative AI system under your own brand (built on third-party models) | Provider — own-brand offering makes you provider regardless of who developed the model | A50-2, A50-5, A50-MAP | Gap — no machine-readable marking (watermark / metadata / provenance) on outputs | ___ |
| AI-generated / AI-edited images, audio, video | AI generates or substantially edits visual and audio assets | Deployer | A50-4-DF, A50-5, A50-MAP | Gap — no visible labels | ___ |
| Realistic depictions of real people / places / events | Assets that could misleadingly appear authentic (deep fake scope) | Deployer | A50-4-DF, A50-5 | Gap — no visible label on or with the asset; label must survive download | ___ |
| AI-generated / AI-assisted published text (general) | AI drafts or substantially edits published copy | Deployer | A50-4-PI where public-interest, A50-5, A50-MAP | Gap — no disclosure applied | ___ |
| AI text informing the public on matters of public interest | AI-generated text published to inform the public | Deployer | A50-4-PI, A50-5 | Gap / unresolved — human editorial review happens only "sometimes", so the editorial-responsibility exemption cannot be relied on across the board; no disclosure applied | ___ |
| Archive of AI content published before 2 Aug 2026 | Existing AI-generated assets held for possible reuse | Deployer | A50-ARCH, A50-MAP | Not required retrospectively — but reuse, republication or substantial modification after 2 Aug 2026 triggers labelling afresh; no flagging system in place | ___ |
| Emotion recognition / biometric categorisation | Not used per self-assessment | n/a | None triggered | n/a — re-check if this changes | ___ |
| (audit yourself) | ___ | ||||
| (audit yourself) | ___ | ||||
| (audit yourself) | ___ |
Suggested candidates for the three empty rows: social media posts and paid ad creative produced by agencies or freelancers; AI-assisted email, newsletter and CRM copy; AI-generated translations, subtitles, voice-overs or synthetic presenter video; AI-drafted product descriptions, support articles or internal reports that reach EU-based customers.
3. How to complete this inventory
List every channel that reaches people in the EU. Website, landing pages, social accounts, ads, email, in-product copy, support content, PDFs and reports, chat and voice interfaces. One row per channel or system — not one row per department.
For each row, record whether AI generates, substantially edits, or merely assists. Note the specific tool and who operates it. Include work produced by agencies and freelancers on your behalf: it reaches the public under your name and belongs in your inventory.
Determine the role from the actual distribution of tasks, not the contract wording. A supplier agreement calling you a "customer" or "licensee" does not settle the question. Ask: do we offer this system to others under our own brand, do we configure or fine-tune it, do we control the outputs and their release? If yes, you are likely a provider for that system — as you are for your own-brand generative AI product — and Art. 50(2) marking becomes your obligation, not your vendor's. You can be provider for one system and deployer for another, and both for the same one. Write the reasoning into the row, not just the label.
Map the requirement IDs to each row using the legend above, and record where the corresponding disclosure currently appears — or state plainly that it does not. "Partially", "sometimes" and "in the FAQ" are gaps, not partial compliance.
Close the chatbot gap first. Under Art. 50(1), the notice must appear at the start of the interaction ("You are chatting with an AI assistant") and not be buried in terms or FAQs. Record where and how the notice is displayed for each interface.
Decide, per publication type, whether you disclose or rely on editorial responsibility. Art. 50(4), second subparagraph exempts public-interest text only where there is genuine human review — accuracy, plausibility and source checking — with a named natural or legal person holding editorial responsibility. A formal skim-through does not qualify. Because your review happens only "sometimes", set a default of disclosure and carve out only those workflows where you can produce records naming the responsible person.
Design the visible label for media before you produce more of it. Under Art. 50(4), first subparagraph, the label must be recognisable at first glance, sit on or with the content, and persist on the asset after download. Metadata alone, footnotes, asterisks and site-wide notices are not sufficient. For evidently artistic, creative, satirical or fictional work, disclosure is still required but may be presented so it does not hamper display or enjoyment.
Specify the marking method for your own-brand generative product. Art. 50(2) requires outputs to be machine-readable and detectable as artificially generated, so far as technically feasible — watermarking, metadata or a provenance standard. Document the method and its robustness. Note the transition: systems placed on the market before 2 August 2026 must meet the marking obligation by 2 December 2026.
Run every completed row through the disclosure-quality check. Art. 50(5) requires disclosures to be clear and distinguishable, present at the latest at first interaction or exposure, in the audience's language, and conformant with applicable accessibility requirements. Test perceivability for users with disabilities, including screen-reader users for on-image labels.
Flag the pre-August-2026 archive rather than relabelling it. Under A50-ARCH, old content needs no retrospective labelling — but tag AI-generated assets in your DAM or media library so that any republication, substantial modification or reuse in a new campaign after 2 August 2026 triggers a labelling check. Add that check as a required step in campaign and republication workflows.
Assign a named owner to every row. Replace each
___with a person, not a team. The owner is accountable for keeping the row accurate and for the disclosure actually appearing in production.Put responsibility in writing with external partners. Define contractually with agencies and freelancers who reviews, who approves and who applies the disclosure before publication.
Re-run the inventory on a fixed cycle and on trigger. Review at least quarterly, and immediately whenever you add a tool, launch a channel, change a vendor, or start using AI for something new — for example emotion recognition or biometric categorisation, which you do not use today but which would bring additional obligations into scope if that changes.
Prepared with AI assistance; reviewed by the A50Check team.