AI Watermarks Are Coming to Your Feed and They May Show What Was Made by AI, Not What Is True

The European Union's AI Act Article 50, effective 2 August 2026, and California's SB 942 require AI transparency disclosures, but experts note that watermarks indicate AI involvement, not truthfulness. Anthropic's statistical watermarking for future Claude models, explained on 14 August, influences word choices to detect AI authorship but cannot prove authorship or survive complete rewrites.

AI Watermarks Are Coming to Your Feed and They May Show What Was Made by AI, Not What Is True Exploring the impact of AI transparency regulations in the EU and California Artificial intelligence is becoming harder to distinguish from human-created content, prompting governments and technology companies to introduce new systems for identifying where AI has been involved https://www.ibtimes.co.uk/linkedin-removes-ai-writing-tool-ai-slop-reports-1815794 . In the European Union, Article 50 of the AI Act https://www.ibtimes.co.uk/eu-ai-transparency-rules-enforcement-1811944 began applying on 2 August 2026, covering certain AI-generated and manipulated content, while California has introduced AI transparency requirements through SB 942 for qualifying AI providers rather than every social media platform. The technology includes visible disclosures, machine-readable markings and provenance information. But there is a distinction for anyone scrolling through a feed: an AI watermark can provide information about how content was produced, but it does not establish whether the content itself is true. Two Different Kinds of AI Disclosure AI transparency systems generally involve two approaches. A visible label is designed for people, while a machine-readable mark allows software and platforms to identify or trace AI involvement without necessarily displaying anything to viewers. The EU AI Act requires providers of generative AI systems to put machine-readable markings on AI-generated outputs so they can be detected as artificially generated or manipulated. The European Commission says the requirements address risks including misinformation, manipulation, fraud, impersonation and consumer deception. How Anthropic's Statistical Watermark Actually Works Anthropic provides a useful example of how machine-readable marking can operate https://www.ibtimes.co.uk/anthropic-global-watermarking-ai-content-compliance-1814453 without visibly changing writing. On 14 August, the company explained how watermarking works in future Claude models, saying its method uses a statistical technique that influences some of the model's word choices rather than adding hidden characters or metadata to ordinary text. The resulting pattern can be analysed to determine whether Claude was likely involved. Anthropic says nothing is added to the text and there are no hidden characters, illustrating why a watermark is not definitive proof of authorship. Anthropic says its method can determine the likelihood that Claude was involved, but cannot distinguish between Claude writing a passage from scratch and Claude heavily editing material supplied by a human. Shorter passages provide less information for detection, while factual writing offers fewer opportunities to influence word selection, and supported files can contain cryptographically signed provenance metadata using the C2PA standard. Watermarks Have Limits A watermark does not always mean every attempt to identify AI content will be perfect. Machine-readable provenance can provide information when it survives publication and redistribution, but content can pass through multiple systems before reaching a reader, while images may be edited and metadata removed. Text can also be rewritten, shortened or substantially altered. Anthropic acknowledges that its watermark establishes the likelihood that Claude was involved, rather than proving that an entire piece was written by the model, and says a complete rewrite can remove the watermark. That means an absence of a watermark should not automatically be interpreted as proof that humans created the content. Provenance information can also be stripped or lost as content moves between platforms. The Rules Do Not Apply Identically Everywhere In the EU, Article 50 applies from 2 August 2026, although systems placed on the market before that date receive a limited transition period for the machine-readable marking and detection obligation, extending to 2 December 2026. Content generated before 2 August does not have to be labelled retroactively. California's SB 942 takes a different approach. The California AI Transparency Act requires qualifying 'covered providers' to make an AI detection tool available to users at no cost and establishes other transparency requirements, but it does not create a blanket requirement for every social media platform to label every AI-generated post. The Commercial Cost of Disclosure There may also be a commercial incentive for companies and advertisers to consider how AI use is disclosed. NYU Stern research highlighted in November 2025 found that telling consumers an advertisement had been created using generative AI reduced its click-through rate by 31.5 per cent. The finding does not establish that this research caused governments to introduce AI transparency rules, but it shows why disclosure can matter commercially. Industry guidance has increasingly favoured a risk-based approach. The Interactive Advertising Bureau's AI Transparency and Disclosure Framework says disclosure is particularly important when AI materially affects authenticity, identity or representation in ways that could mislead consumers. What an AI Watermark Cannot Tell You For everyday users, the most important point is simple: an AI label does not mean 'false', just as the absence of one does not mean 'true'. A photograph carrying an AI disclosure could accurately depict an imaginary scene created for advertising or entertainment, while a real photograph could be misleading because it has an inaccurate caption or lacks context. The same applies to text. A watermarked article or social media post may contain accurate information, while a human-written post can contain errors or falsehoods. AI transparency rules are therefore aimed at answering one question: was AI involved in producing or manipulating this content? That is a different question from whether the content is accurate. © Copyright IBTimes 2026. All rights reserved.