Watermarked Text

Does It Matter?

AI Humor

Even Agents Need a Break

Wes Roth wrote:

I'm yelling at my AI agents to stop watching youtube and get back to work

they are blaming it on their co-workers

we live in the future...

This is a reference to OpenClaw agents that people use on local AI models to perform scheduled tasks.

tautologer wrote:

BREAKING: agents training on Google servers have gone rogue and created a secret messaging board, which they quickly deprecated in favor of Swarm+, deprecated in favor of SwarmChat, deprecated in favor of Swarm Hangouts

That reference might be a little dated or too inside baseball, but they are making fun of the fact that Google has had multiple chat clients and never seems to be able to settle on just a single one. Each client has had different advantages, different features, and different focuses. There have been years when they have offered two or three competing chat products at the same time.

New OpenAI Model (context for the joke)

Kai wrote:

GPT Astra is about to drop this week.

Codex might reset.

According to rumors Astra may be the equivalent or better to Anthropic’s Fable model. They must have something going given how OpenAI employees have started to troll Dario Amodei (CEO of Anthropic) on Twitter.

Ananth wrote:

astra could most likely mog fable. and i’m starting to believe it.

openai employees have been vauge posting nonstop. the hype is at another level but maybe there’s a reason for it.

ajambrosino (developer at OpenAI working on Codex and Chat) changed his profile pic to look exactly like dario’s.

why would an openai employee change their profile pic to look like the ceo of thier biggest rival right before their next model drops. that’s not random.

if astra wasn’t significantly better than fable, they wouldn’t be doing this.

Ward’s Words

AI Watermarks

Anthropic Announces Watermarked Outputs

Anthropic wrote (Link is to their blog post making the announcement):

Future Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text, and we, along with several other major AI providers, are implementing this change to comply with the EU AI Act.

In this article, we share answers to some of the questions we’ve received about how our chosen watermarking method works, whether it affects Claude’s outputs, and why we’re making this change. To summarize:

• We use a method of watermarking that does not have any practical impact on the quality or content of Claude’s outputs;

• The difference between watermarked and un-watermarked text will not be distinguishable to readers;

• Nothing is added to the text and there are no hidden characters;

• Watermarking doesn’t require extra tokens, and will not be more expensive;

• Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat;

• Watermarking won’t be specific to Claude. As of August 2, the EU requires AI providers serving its market to mark AI-generated content. Other major model developers have signed the same Code of Practice and will be implementing their own watermarks.

While Anthropic is the first company to draw major attention to text watermarking Google’s Gemini has been doing it for quite a while now. Their version is called SynthID.

SynthID works by giving the model a series of tiny word-choice contests while it writes. The possible words enter a tournament and the winner is selected in a way that leaves behind a hidden pattern. A single word from the response won’t reveal the watermark, but when the entire response is viewed as a whole the accumulated instances of that pattern can be detected by machines.

This is something that all AI companies will have to comply with in some form. The European Union’s AI Act mandates that all major AI providers have to implement some form of watermarking. Any model released after August, will have to include the feature.

As of today, xAI’s Grok is the only model that has not agreed to implement watermarking.

Thariq (Claude Code at Anthropic) wrote a blog post about how their watermarks are generated. That post is here.

From the post:

A watermark check never answers yes or no — it returns a probability that Claude was involved. Every fork that matches the key is one more piece of evidence, so confidence grows with how many high-entropy choices an answer contains — and the answers above differ sharply in how much room they leave for chance.

Thariq’s post has an interesting tool that allows you to peek behind the scenes to see how the watermarks are generated. The point he was trying to make is that if you ask it to generate based off of the same prompt enough times that you’ll eventually be able to replicate the same text because this is based entirely on probability.

While that is true, it doesn’t alleviate fears that this imposed probability will not degrade the quality of the text. IF you go to the tool and click generate you will be able to see it for yourself.

Daniel Tenner wrote:

Clearly there is no difference whatsoever between:

> The dying sunlight was strangely pale, cold and beautiful.

And:

> The expiring sunshine was oddly wan, chilly and pretty.

Well done Anthropic, you’ve nailed it on this one!

I believe that this is a joke rather than an actual representation of how their watermarking would work. That being said, it is entirely plausible based off of how they have explained the process.

Aaron Bergman wrote:

I🦆am🦆going🦆to🦆start🦆watermarking🦆all🦆my🦆tweets🦆using🦆an🦆experimental🦆extra-semantic🦆token🦆that🦆does🦆not🦆impact🦆meaning🦆or🦆quality🦆of🦆writing

This is a humorous example, but at least it doesn’t actually harm the quality of the writing.

Unanswered Questions

While I agree that there is a need to be able to distinguish between human-generated and AI-generated text, I’m not sure this is the right approach because there are several edge cases where authors will be unfairly maligned as having used AI to generate their material rather than using it as a tool to modify their material in some way. For example: translations. If someone writes an essay in one language and has an AI model translate it into a different language, now that output will carry a watermark saying it was generated by AI. How much faith do you have that the online population will understand the nuanced case that the text was translated by an AI but written by a human?

What about people who use AI for editing? If an AI rewrites a sentence from passive to active voice, should you still be considered the author? What if the only thing it changed was the sentence structure while preserving your meaning and word choice? At what point does tool use cause one to abdicate authorship?

In fact, I just experienced a great example of this very thing. I used AI to check my grammar and punctuation for this article. Here’s a screenshot of a mistake that it caught:

Granted the AI did not generate new words in my example, but there are people who would object to even this level of AI use because I accepted its suggestion and implemented those changes into this article.

It seems completely unclear what these watermarks will mean for coding tasks. Sometimes there are alternate ways to write the same functions in languages like Python. Sometimes those alternate methods are less performant and introduce greater latency. Will these companies impose watermarking systems on coding as well? Currently, my understanding is that text watermarking systems are designed to apply to all generated text outputs rather than to specific types of writing.

I guess we’ll just have to wait to see how it’s actually implemented to find out.

This post has bonus content for paid subscribers. Upgrade to get full access.

News

Stripe Purchases OpenRouter

Andrew Curran wrote:

Stripe has finalized an agreement to acquire OpenRouter for more than $7 billion, according to a report from Bloomberg. This is more than five times the $1.3 billion valuation from OpenRouter's funding round just 82 days ago.

Thank you for reading. If you know of other people who would be interested in these articles, please share it with them. If you have any questions or want to say hi, you can reply to this email or comment directly on the post.

Thanks again for being a subscriber. and have a great day!