The Unmarked Truth: Anthropic and the Great Watermark Debate
Anthropic announced that Claude will start watermarking the text it generates. This means an invisible, machine readable mark is woven into the text itself. It applies to every Claude model launched on or after 2 August 2026, and it applies globally rather than only to users in Europe.
The trigger for this is Article 50 of the European Union Artificial Intelligence Act, which became applicable from 2 August 2026. This law requires providers of generative AI systems to ensure their outputs, including text, are marked in a format that is machine readable and detectable as artificial intelligence generated. Breaking this rule carries fines of up to fifteen million euros or three percent of global annual turnover, whichever is higher.
The duty under Article 50 is binding. However, how a company proves it meets that duty involves a voluntary element. The European Commission published a Code of Practice on Transparency of Artificial Intelligence Generated Content. Signing this Code gives companies a pre approved route to compliance and legal certainty. Anthropic signed it, alongside other major labs such as Google, Meta, Microsoft, and OpenAI. OpenAI previously acknowledged that text watermarking at scale remains difficult to deploy. Anthropic appears to have shipped what competitors have not yet managed.
What the Watermark Actually Does
Anthropic describes the watermark as a statistical signal embedded directly into generated text. It travels with the text when copied and pasted, and it may persist through some editing. However, Anthropic notes that it is not yet clear how much editing is required to remove it, and they have promised to publish technical details so third parties can build detection tools.
The mark shows that Claude had some hand in a piece of text. It does not show what that hand did. Asking Claude to proofread a paragraph or translate a sentence leaves the same trace as asking it to write the whole thing from scratch. The watermark cannot tell the difference between authorship and assistance. At a technical level, there is no difference to find.
Anthropic chose to apply this mark everywhere, not just inside the European Union. Article 50 governs artificial intelligence systems used within the European Union and does not require a global rollout. Anthropic chose a single global standard because building two versions of a product is more expensive than building one to the highest common standard. If other labs follow this logic, the European Union law will shape product design worldwide.
The Problem With Binary Labels
When I am working on a piece I usually follow a fairly standard pattern. I dictate my thoughts into a voice note. Claude takes that voice note, removes the parts where I lost the thread, and structures what was left into something I could edit. I use artificial intelligence as a sounding board and a brainstorming partner. I instruct it to challenge and question me, which suits my background as a former lawyer. I then take that newly acquired thinking and the note Claude has produced as the raw materials for me to write and edit from.
Whether that makes this piece artificial intelligence generated is a difficult question. I think out loud because my ADHD brain needs to get ideas out of the noise and into a shape I can work with. Talking a problem through is my thinking process. The analysis, the argument, and the judgement about what matters belong entirely to me. That’s my process, but everyone will use AI tools differently depending on their own personal capacity and circumstances. There is no right or wrong way and gatekeeping who and how tools should be used and the outputs that they generate has become a loaded conversation in some circles, which can be exclusionary or judgmental.
The current conversation about artificial intelligence generated content relies on a binary label. It assumes either a human wrote it or a machine did. This binary fails to describe how people actually use these tools. It stigmatises specific groups of people through narratives found on professional platforms.
Non native speakers use artificial intelligence to match their writing to their thoughts.
Dyslexic writers use tools to catch errors their brains miss during proofreading.
Autistic people use technology to translate their thoughts into a register that avoids being misread as blunt.
People with attention deficit hyperactivity disorder use tools to structure their spoken thoughts into written form.
Individuals with chronic fatigue use assistance when they lack the physical stamina to get ideas onto a page unaided.
People with motor differences use supported tools as a bridge between knowledge and physical production.
Collaboration is a continuous process rather than an on and off switch. A sentence drafted entirely by a human sits next to one tightened by a tool, which sits next to a paragraph restructured and then edited back into a personal voice. Treating this as a binary generated label throws away useful context about where human judgement was exercised.
The European Union law recognises this nuance. Article 50 states that obligations do not apply where content has undergone human review or editorial control, and where a natural or legal person holds editorial responsibility.
When Labels Become Weapons
Binary labels quickly turn into tools for gatekeeping and policing who gets to speak with authority.
A teacher in the United States embedded hidden text in an assignment, instructing any artificial intelligence tool reading it to output a specific phrase if a student used it. Most of the class failed. The exercise unfairly punished students who used the technology as a legitimate support for learning rather than to cut corners.
Dr Sam Illingworth ran his old doctoral thesis, written years before these tools existed, through a detection system. It came back as seventy percent artificial intelligence generated, proving that detection systems can confidently give wrong answers on writing that predates the technology.
I was once accused in a group chat of using artificial intelligence to write a post because my argument was inconvenient to another person. The accusation was a deliberate attempt to dismiss my argument without engaging with it.
In another exchange, I replied to someone using only my thumbs on a phone. The response was a complaint that I was starting to sound like a language model. That exchange was a patronising attempt to put a woman who writes clearly into a smaller box.
These experiences do not excuse genuine failures. When professional firms pass off unverified machine outputs as due diligence to clients, accountability fails. That is a real problem. However, treating that failure as identical to legitimate human collaboration damages people who are using tools correctly.
Wanting It Both Ways
Some critics argue that watermarking itself is unethical, calling hidden marks a surveillance practice or an attempt to claim improper ownership. Given that these models are trained on unauthorised data, the practice can feel extractive. However, that criticism clashes with demands from the same people for more transparency about what is artificial intelligence generated.
A watermark is a transparency mechanism. Objecting to the signal while simultaneously demanding the information the signal provides is contradictory. This is an opening move in a long conversation, and future iterations of these tools will look different.
Even within days of the announcement developers responded by building software tools to wipe those marks away. Known as "watermark strippers" or removers, these programs automatically tweak the text, stripping out metadata and rearranging word patterns so that detection algorithms can no longer tell the content came from an AI. Many creators use these tools because they worry that a blanket AI tag will wrongly brand their work, especially when they only used the AI to proofread or polish human-written ideas. This rapid game of cat-and-mouse shows just how difficult it is to enforce rules around labeling digital content, as technology intended to make AI use transparent is almost immediately met with new software built to bypass it.
If a watermark cannot distinguish full generation from light assistance, and a detection tool misclassifies old academic work, then the presence of an artificial intelligence signal is not the right place to build accountability. The important question in academic work, journalism, and government reports is whether meaningful human judgement, oversight, and editorial responsibility guided the final result. Provenance matters, but the assumption that a binary label can do that work is false.
