Build1 distinct publisher3 min readUpdated
Article 50 and California's transparency act both switched on 2 August 2026. The watermarking work belongs to model vendors; a shipping team owes chatbot, deepfake and public-interest text notices.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Two AI labelling regimes switched on together on 2 August 2026: Article 50 of the EU AI Act and California's AI Transparency Act, SB 942 as amended by AB 853 [1][15]. For anyone shipping a product rather than running a model, the practical consequence is that the expensive part of compliance is a ticket on someone else's board.
The EU AI Act splits duties between the provider, meaning whoever puts a generative system on the market, and the deployer, meaning whoever uses it, and nearly all the engineering-heavy requirements sit with the provider [2]. Article 50(2) requires providers to design systems so that synthetic image, audio, video and text output carries a machine-readable mark identifying it as AI-generated: watermarking, C2PA-style content credentials, cryptographic signing [3]. If you generate an image in a mainstream tool, marking it is the tool's job [3].
What attaches to a deployer is three things [3]. If you deploy an AI system that interacts with people, they must be told they are dealing with AI unless that is obvious from context [4]. If you publish AI-generated or manipulated image, audio or video resembling real people, places or events, you must disclose it [5]. If you publish AI-authored text intended to inform the public on matters of public interest, you must disclose that as well [6].
The third duty has an exception written into it: it does not apply where the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication [7]. The dev.to explainer reads that plainly as the difference between a site that pipes model output straight to publication and one where a named human edits and stands behind the piece [7]. There is no equivalent escape for image, audio or video; a deepfake stays a deepfake after you edit it [8].
Note who the timeline favours. Article 50 kept its 2 August 2026 date while the Digital Omnibus, in force from 27 July 2026, pushed the standalone Annex III high-risk obligations to 2 December 2027 and Annex I embedded high-risk to 2 August 2028 [9][10] - a gap of sixteen months and twenty-four months respectively behind the transparency rules [2]. The single concession granted to Article 50 is a four-month conformity window, to 2 December 2026, for generative systems already on the EU market before 2 August; anything launched on or after that date gets nothing [11][1]. That window is a provider window. Deployer disclosures were due on 2 August [12].
Exposure is real but not top-tier. Article 50 breaches sit at up to 15 million euro or 3 percent of worldwide annual turnover, whichever is higher, against the 35 million euro or 7 percent reserved for prohibited practices, a cap 20 million euro lower [13][4]. The Act also reaches non-EU businesses whose AI output is used in the EU, so a US freelancer serving European clients is in scope [14].
California is narrower still. CAITA regulates covered providers only, defined as publicly accessible generative AI systems with more than one million monthly visitors or users in California [16]. Below that threshold it does not apply to you [16].
What to watch: whether the machine-readable marks actually show up in vendor output by the 2 December 2026 conformity deadline [11], because deployers relying on tool-side marking inherit the gap when it is missing. Watch, too, how the editorial-responsibility carve-out is treated where review is nominal rather than substantive; the text sets the test at responsibility, not effort [7].
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Two AI labelling laws took effect together on 2 August 2026: Article 50 of the EU AI Act and California's AI Transparency Act.
The EU AI Act splits obligations between the provider, whoever puts the generative AI system on the market, and the deployer, whoever uses it; nearly all the engineering-heavy requirements sit with the provider.
Under Article 50(2), providers must design their systems so that synthetic image, audio, video and text output carries a machine-readable mark identifying it as AI-generated; this covers watermarking, C2PA-style content credentials and cryptographic signing, and is the tool's job when a user generates an image in a mainstream tool.
Deployer duty one: if you deploy an AI system that interacts with people, those people must be told they are dealing with AI, unless it is obvious from context.
Deployer duty two: if you publish AI-generated or manipulated image, audio or video that resembles real people, places or events, you must disclose it.
Deployer duty three: if you publish AI-authored text intended to inform the public on matters of public interest, you must disclose that.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One secondary explainer, no primary citations
Every claim traces to a single dev.to community post that restates statutory structure, dates, thresholds and penalty tiers without linking or quoting the primary texts (Article 50, SB 942/AB 853, the Digital Omnibus) and without any legal, regulatory or vendor corroboration. The claims are internally consistent and specific, which raises them above rumour, but nothing in the cluster independently verifies the high-consequence particulars such as the 2 December 2026 conformity window or the EUR 15 million / 3% tier.
No adoption signal in cluster
The supplied material contains no observations of behaviour: no evidence that providers have shipped conformant machine-readable marking or detection tools, no deployer disclosure practices, no enforcement actions, and no usage or platform data. Legal application dates are obligations, not adoption, so no adoption value can be derived without inventing facts.
Deflationary framing, overconfident legal precision
The cluster's headline claim is deliberately deflationary — most of the burden is a provider problem and an individual's list is three disclosures — which understates rather than inflates stakes. The mild positive gap comes from the certainty with which uncited legal particulars are delivered (exact postponement dates, a four-month marking window, a clean 'it does not apply to you' threshold read), and from asserting that mainstream tools handle marking when no evidence of provider conformity is offered. The overstatement is in confidence, not in consequence.
Practitioner explainer with SEO-adjacent framing
The only source is a self-published developer-community post addressed to freelancers, marketers and small publishers, a format that rewards authoritative-sounding, shareable checklists. The piece also editorialises that the text carve-out 'neatly rewards the editorial workflow that helpful-content ranking already rewards', a search-visibility framing that indicates an audience-building angle. There is no disclosed vendor, law-firm or platform interest, so the incentive pressure is presentational rather than commercial.
Low: single publisher, uncorroborated legal detail
Confidence is capped by structure: one publisher, one author, zero primary or expert corroboration, and no adoption or enforcement observations. The duty-allocation logic (provider marking versus three deployer disclosures) is coherent and plausible, but every date, threshold and penalty figure would need independent verification before a team relied on it.
build
A 14,000-star watermark remover, and no detector to test it against1 distinct publisher
product
A five-hour script beats Claude's watermark, so stop treating it as provenance3 distinct publishers
science
Text watermarks land on 2 December. The detection they imply does not.1 distinct publisher
science
Claude's watermark is a compliance artefact, not a cheating detector1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 16, 2026