Claude to start watermarking AI-generated text – but will it make quality worse?
Wednesday, 19 August 2026
Warm-up
- Where do you usually notice AI-generated writing online?
- What would be a fair way to label text written by a chatbot?
- Have you ever preferred a machine’s wording to a human’s? Why or why not?
Vocabulary
- watermark
- A hidden mark added to text to show where it came from.
- regulation
- An official rule made by a government or authority.
- detectable
- Possible to notice or find using a method or tool.
- decode
- To change a hidden signal into readable information.
- constrain
- To limit what someone or something can do.
- randomness
- The quality of happening by chance, not by a plan.
- pattern
- A repeated design or signal that can be recognized.
- disinformation
- False information spread to mislead people.
Reading
Anthropic says its Claude chatbot will add a watermark to AI-generated text to meet an EU regulation that starts in December. The tweak works at a granular, random level, shaping tiny choices and leaving a pattern detectable to Anthropic and to anyone who can decode it.
Some fear quality will fall. John Gruber, a veteran tech blogger, calls the plan an adulteration of writing and argues the watermark could constrain Claude, pushing it toward less precise word choices. However, Steven Murdoch, a computer science professor at University College London, expects little or no noticeable impact.
He notes that randomness is essential: it stops chatbots from looping and repeating. The watermark may nudge whether a sentence says “grey” or “overcast”, or picks “stream” or “brook”, but it should remain fluent. The rule covers AI firms operating in the EU and could deter students, lawyers and professors from passing off chatbot work as their own.
There is another motive: avoiding “model collapse” when AIs train on AI-written text. Clear labels can also fight disinformation. By the time the rule begins in December, AI firms in the EU will have added watermarks.
Comprehension
- Why is Anthropic changing how Claude writes?
- What hidden signal will the new system leave in the text?
- Who says watermarking could constrain Claude’s word choice?
- What benefit of randomness in AI writing does the text describe?
- By when will AI firms in the EU have added watermarks, according to the text?
Grammar focus
Future perfect
Use will have + past participle to say an action will be finished before a future time. It shows completion by that moment. (Example 2 constructed.)
- By the time the rule begins in December, AI firms will have added watermarks.
- By the time the rule starts, platforms will have added watermarks.
Grammar exercise
Complete each sentence using the future perfect (will have + past participle).
- By December, platforms (add) ___ watermarks to text.
- By the time the rule begins, developers (update) ___ models.
- Before the rule starts, teams (test) ___ watermarking methods.
- By then, students (notice) ___ more labelled work.
- Next year, AI firms (change) ___ training data practices.
Discussion
- Should AI-written text always carry a visible or hidden label? Why?
- Would watermarking change how you use AI tools for work or study?
- Who should be able to decode AI watermarks: everyone or only authorities?
- Is a small quality risk acceptable to reduce disinformation? Why or why not?
Vocabulary, reading, and exercises with instant feedback