The EU Artificial Intelligence Act (“the Act”) was formally adopted by the European Council in May 2024, with the majority of obligations coming into effect in August 2026.
The European Commission notes that the EU Artificial Intelligence Act (“the Act”) is the world's first comprehensive law for AI. It aims to address risks to health, safety and fundamental rights. The regulation also aims to protect democracy, the rule of law and the environment.
While it governs AI use in the European Union, like the strict requirements of Europe’s General Data Protection Regulation (GDPR), the regulations are likely to flow to other parts of the world as systems designed to meet the highest standards are deployed globally.
Article 50 of the Act includes obligations on providers and deployers of AI systems, designed to provide transparency to users, which can be summarised as follows (which apply apart from some exceptions):
1. Providers of AI systems must ensure that its clear to users that they are interacting with an AI system.
2. Providers of AI systems must mark their outputs in a machine-readable and detectable format.
3. Deployers using AI systems to detect emotions or categorise people based on biometric traits must ensure that subjects are aware this is happening.
4. Deployers must label content as artificially generated or manipulated when that is the case.
50(2) Providers’ obligation to watermark outputs
As of 12 August 2026, 82 organisations providing AI systems have signed the Code of Practice on Transparency of AI-generated Content, demonstrating their commitment to comply with the obligations under Article 50(2) of the Act.
Notably, providers are required to comply with Article 50(2) from 2 August 2026 in respect of any generative AI system placed on the market after that date. Providers have until 2 December to make their older systems compliant.
In early August, Anthropic announced that Claude models launched in the EU on or after 2 August 2026 will embed watermarks into generated text. At a later stage Anthropic will provide means for users to detect these watermarks.
This follows Google deploying watermarking on its Gemini outputs way back in May 2024, and OpenAI deploying watermarking on images and audio in May and July respectively (but not yet on text).
So, we can expect more announcements in the latter half of 2026 confirming providers’ implementation of watermarking across all of their outputs.
The (in)effectiveness of watermarking with text
Watermarking text is intended to help us understand its provenance. Once the providers make their detection tools publicly available, we will be able to identify text that is simply copied and pasted from generative AI as just that (provided it meets minimum character thresholds).
But this does little towards allowing users to understand how a piece of digital work was constructed.
The main reason behind this is that it is impossible to watermark text in a manner which can survive manipulation. “Watermarking” traditionally relates to graphics where a visible or machine readable markings are embedded, and cannot be removed without altering the watermarked graphic. A string of text can not abide by this criteria (simply put, words can be changed to create a different, but complete, text string), so this is not what “text watermarking” involves.
Text watermarking is better explained as a hidden algorithm that requires generative AI to select specific synonyms over others, creating a mathematical pattern detectable by software holding the cryptographic key.
The problem then is obvious; minor changes to a text string make the passage unrecognisable to the watermark detection tool, therefore “breaking the watermark”. And minor changes happen almost naturally, except in the laziest circumstances!
At the innocent end of this scale, this may involve changing some words because they don’t sound right, or re-ordering some sentences or paragraphs because you don’t agree with the order. Or maybe adding in some thoughts of your own.
At the nefarious end this may involve using “AI humanizers” to rewrite the text, which is already being done at scale so as to defeat AI detection tools.
The stripping of watermarks results in a “false negative”; that is, a detector incorrectly concluding that a string of text did not come from generative AI. The “false positive” problem comes where someone interprets the detector’s correct conclusion that a string of text comes from generative AI, to mean that AI was the original author. Someone using generative AI to proofread, translate, summarise or convert files will produce a watermark which unfairly discredits their original work (assuming that these actions were permitted).
None of these weaknesses are controversial claims, and are acknowledged by Anthropic in their watermarking announcement. Put simply, the AI providers will meet their obligations under the Act, however those obligations do little to solve the provenance problem that Workings was designed to address.
Workings as the solution
Anyone wanting to guess the provenance of something they consume is stuck with the flawed technologies of watermark detection, AI detection, and unsupported claims of the author. Unless the author uses Workings, which allows them to demonstrate how that work was made. Take any of the examples listed above:
- The author copies and pastes text from generative AI then changes some words, re-orders sentences/paragraphs, or adds thoughts of their own. The Workings timelapse will show the entire process as it happened; the AI tool being prompted, the text being copied and pasted, and then altered. It’s for the reader to decide what to make of that.
- The author uses an “AI humanizer” to rewrite text. Again, the Workings timelapse will show the AI tool being prompted, the text being copied and pasted into the AI humanizer, and the output being pasted out of it.
- Where someone uses generative AI to proofread, translate, summarise or convert files. The Workings timelapse will show this process and the author can demonstrate the falsity of any claim made against them.
Want to show how you made your work? Get Workings.
Want to understand how work was made? Ask for their Workings.