EU AI Act Article 50 Provenance Compliance Kit
Watermarking, metadata and an audit-evidence pack so that smaller generative-AI providers and fintechs can show Article 50 compliance for text output.
Competitive
Written for: Small startup team
Why now
The biggest provider has shown that the obligation is real and enforceable in practice. Thousands of smaller EU-facing providers, including fintechs that send AI-written customer letters, now need a defensible approach. But detection interoperability is unresolved and buyers may wait for guidance, so this is a watch.
What became possible
A new paper reframes text watermarking from "impossible to make robust" to "auditable compliance" under Article 50(2), which ties the obligation to technical feasibility and the state of the art. OpenAI is now watermarking ChatGPT and Codex text in the EU, admits editing weakens detection, and is giving detection access to researchers first.
One way in
- Article50 Ledger
A middleware that adds a watermark or metadata tag to generated text, logs provenance receipts, and produces a technical-feasibility justification document that maps the controls used to the qualifiers in Article 50.
- Customer
- EU SaaS companies with genAI features; fintechs generating customer communications
- Model
- SaaS per million generated tokens plus a compliance-report subscription
Competition
Watermarking text is largely taken: Claude and Gemini watermark their output, Microsoft, Amazon, Mistral and Cohere signed the Code of Practice, and SynthID Text is free and open source. Selling your own text watermark is a weak pitch.
Full competition notes (10 competitors)
Watermarking text is largely taken: Claude and Gemini watermark their output, Microsoft, Amazon, Mistral and Cohere signed the Code of Practice, and SynthID Text is free and open source. Selling your own text watermark is a weak pitch. The opening is an evidence and paperwork layer that works with any model, for companies that may legally count as the 'provider' of their own AI system (fintechs, EU SaaS). It would record which upstream model produced each output and whether it was watermarked. It would add a metadata tag to text sent through channels where watermarks don't survive (email, PDF, templated messages). And it would produce a 'technically feasible' justification that follows the final Code of Practice clause by clause. Nobody found targets text-only fintech customer communications. A second opening: EU teams self-hosting Mistral or Llama models get no upstream watermark and need a managed SynthID setup plus detection plus the audit pack. Risks: text watermarks break easily under editing or paraphrasing, model providers keep expanding coverage, governance platforms could add Article 50 modules, and demand may peak around the 2 December 2026 grace-period deadline.
- Anthropic Claude text watermarking
Since 2 Aug 2026, text from new Claude models (including through the API, worldwide) carries an invisible watermark plus C2PA metadata on generated files. This covers marking at no extra cost for customers who build on Claude.
- Google SynthID Text (open source)
Open-source text watermark that plugs into the generation step, built into Hugging Face Transformers 4.46+. Already used on Gemini output, so anyone self-hosting can add marking for free.
- Resemble AI Watermarker
Watermarks text, audio, image and video through one API (PerTh model) and markets itself directly on EU AI Act Article 50 compliance.
- Numbers Protocol
Article 50 'evidence layer': C2PA-compatible provenance, persistent content IDs and inspectable receipts, pitched as 'the label is not the audit trail'. Mostly aimed at media files.
- CrawlQ Compliance
Tamper-evident (Merkle-chained) audit trail with a proof for each asset, plus human-review records for AI-generated marketing content under Article 50 and similar laws.
- MEOK AI Labs Article 50 MCP
Article 50 watermarking and transparency tool, delivered as an MCP server for AI agents, aimed at chatbot operators, general-purpose model providers and deepfake users.
- Disclosure Proof
Tool for recording proof that AI-use disclosures were made (evidence for transparency obligations).
- AIACTO
EU AI Act compliance SaaS that covers Article 50 obligations for businesses.
- eucompliance-ai-act (PyPI)
Early open-source Python package for EU AI Act compliance tooling.
- EU Code of Practice on marking and labelling AI-generated content
Final version published 10 June 2026. The free official reference for how to show compliance, which reduces the value of a proprietary 'mapping' document.
Risks
- Watermark standards and detector access are not settled yet
- Model providers may watermark at source, leaving little for middleware
- How strictly it will be enforced is unclear
Best on open space and buildability; tech readiness and market pull and team fit are the weak point (5).
tech 5 · pull 5 · space 7 · fit 5 · build 7
Sources
Status history
- 2026-10-07 Added to the watchlist