FIR #526: Forget Anthropic&#8217;s AI Watermark. Can You Defend Every Sentence?

Anthropic’s watermark scheme is the focus of so much discussion that you could be excused for thinking there was nothing else to talk about. For many, it’s the solution to identifying all those evil-doers who offload their writing to large language models. But we are wasting far too much time trying to determine whether (and to what degree) AI was involved in creating content. Much more important is determining whether the content was crafted with respect for the reader, and whether the creator can stand by every word. That’s the idea behind the new AI writing policy from the software company Clay, which makes far more sense than trying to detect a watermark.

Links from this episode:

The next monthly, long-form episode of FIR is tentatively scheduled to drop on Monday, August 24.

We host a Communicators Zoom Chat most Thursdays at 1 p.m. ET. To obtain the credentials needed to participate, contact Shel or Neville directly, request them in our Facebook group, or email [email protected].

Special thanks to Jay Moonah for the opening and closing music.

You can find the stories from which Shel’s FIR content is selected at Shel’s Link Blog. You can catch up with both co-hosts on Neville’s blog and Shel’s blog.

Disclaimer: The opinions expressed in this podcast are Shel’s and Neville’s and do not reflect the views of their employers and/or clients.

Raw Transcript:

Neville Hobson: Hi, everyone, and welcome to For Immediate Release. This is episode 526. I’m Neville Hobson.

Shel Holtz: And I’m Shel Holtz. In the last episode of FIR, we talked about workslop and the implications workslop brings to the workplace. We’re going to continue down that path today.

I’m sure you’ve all seen the headlines about Anthropic putting an invisible watermark on anything Claude writes. I want to separate what that actually does from what people have been claiming.

First, you won’t see this watermark. It’s not a “written by Claude” tag. It’s machine-readable. It survives copy and paste, and it may even survive some editing. The reason Anthropic came up with this is to comply with the transparency rules under the EU AI Act, and it applies everywhere Claude runs: the app, the API, Claude Code, Cowork, cloud providers — you name it.

You can’t paste text into a public detector and expose someone yet. Anthropic says detection tools are coming, but it hasn’t released any of them.

Here’s roughly how it works: Every time Claude generates text, it’s choosing between several equally good next words — say, “gray” or “overcast.” Normally, that choice is random. With watermarking, though, a secret cryptographic key nudges that randomness in a consistent way.

No single word looks suspicious, but across a couple hundred word choices in an article, the pattern becomes statistically detectable if you have the cryptographic key.

That’s completely different from tools like Pangram, which look for writing patterns that seem AI-like — you know, em dashes, overuse of words like “delve” or “tapestry” or “underscore,” constantly grouping things in threes. These are all things real writers do, by the way. The rule of three is nothing unique to AI, nor are em dashes. That’s why I don’t put much stock in these tools.

Anthropic’s detector doesn’t guess. It tests whether the text matches the pattern its own cryptographic key would produce.

Now, it’s particularly important to understand that a detected watermark means the content was processed by Claude. It doesn’t mean Claude wrote it. You could write something yourself, have Claude clean up the grammar, and it would still carry that watermark. Axios flagged exactly this risk for communications teams that polish a human-written press release with Claude.

It also works in reverse, by the way. Heavily edit, paraphrase, translate, or blend the text with other writing, and that watermark can disappear. Short passages may not carry enough signal to be detected at all.

So this isn’t a foolproof “Did a human use AI?” detector. Picture a reporter running a company statement through a detector and calling it AI-generated when your team actually wrote it and just had Claude do the final polish.

The watermark tells you about processing. It can’t tell you who did the thinking.

And that brings me to Clay, the software company, which just rolled out an AI writing policy. I think this is actually more important than a watermarking system.

The policy had its origins in Clay’s engineering department, but it has since gone company-wide. It’s not a ban. Brainstorming with AI, drafting, proofreading — all that’s fine.

Instead, the policy has four principles.

First, stand behind every idea and every sentence.

Second, writing is thinking.

Third, spend more time creating a document than you expect someone to spend reading it.

And fourth, longer isn’t better. And AI, by the way, notoriously pads its content.

I like this policy a lot. I like it a lot more than “Don’t use AI to write.”

Because sometimes AI writing is exactly the right call. Plenty of people are great at their jobs and bad at writing. Engineers are a great example. I’ve used engineers as an example of this before. If AI helps these people organize their explanation and turn an incomprehensible email into something readable, that’s good for everyone.

The problem is AI can make bad thinking look like good writing. This is what we were talking about last week with workslop. You give the model a thin prompt, it hands back three polished pages, and now you’ve just shifted the work of figuring out what you meant onto everyone else who has to read it. The prose looks finished, so the creator of that content feels like they’ve finished, but the thinking never actually happened.

That was how we defined workslop last week when we talked about it.

So companies don’t need an AI policy that’s focused on AI. They need one that’s focused on accountability, quality, and respect for the reader.

Cover the basics, for sure. List the approved tools. Talk about confidential information and how it gets used. Talk about fact-checking and human review and brand voice and blah, blah, blah.

But add that Clay test: Can you defend every sentence? Does it represent what you actually think? Did you verify the facts? Did you cut the padding? And did AI make the communication better, or did you just shift your effort onto your audience?

For communicators, that last question is the one to focus on. The watermark debate is going to tempt organizations to obsess over detection: Was AI used? Can we prove it? Should we disclose it?

Yeah, sometimes, I guess, that’s a fair question. But the better ones are: Is the thinking ours? Is it accurate? Does it serve the reader? And is a human willing to stand behind every word?

If we can get those things right, I’m not too worried about whether Claude helped fix a few sentences along the way, watermark or no watermark.

Neville Hobson: Hmm. We have talked about this before — this kind of weird obsession so many people seem to have with trying to figure out whether someone used AI to write a piece so they can exclaim with great glee, “Yeah, this guy wrote this. It’s 96 percent AI. He didn’t write it at all. It’s a scam, it’s fake,” blah, blah, blah.

Shel Holtz: Yeah.

Neville Hobson: We’re already seeing it happen with this. Someone wrote the other day about this, “Finally, a way to get rid of AI slop.”

What? I mean, isn’t —

Shel Holtz: No.

Neville Hobson: It isn’t going to do that. This obsession isn’t going to stop, I don’t think.

And, in fact, maybe the best thing out of the Clay principles is taking the focus away from that and moving it to the writing, to authorship, to accountability, as you point out.

The caution, I would say, is don’t obsess about this. Don’t set concrete rules that are so inflexible that you’re going to end up with writing that isn’t flexible at all.

The point, I guess, is to concentrate on the actual writing. Think about what it is that you’re writing.

So let’s move away from “Did you use AI to write this?” to something like, “What was your contribution to this?”

That, to me, is a healthier discussion to have.

There are still going to be lots of people who do the “gotcha” talk. I saw one today on LinkedIn: “I would never hire someone who uses AI to copywrite. That kind of approach isn’t for my team,” and stuff like that.

I think asking what your contribution was encourages accountability, whereas “Did you use AI to write this?” is all about concealment and policing — discovering that someone did and assuming or implying that they’ve cheated somehow.

That said, I think the point that you need to stand behind every idea and sentence is absolutely spot on. And that’s not new. We should have been doing that all along.

That’s probably the right way to approach it.

For example, don’t ask Claude to write something: “Here’s a topic. Write a 1,200-word piece arguing that communicators need to take responsibility for AI governance.”

Claude produces it, you might edit two or three sentences, and you’re done. You publish it under your name. That is substantially AI-generated writing, and the watermark is relevant.

But for what, though? So someone could say, “Gotcha”?

You don’t know what that was for. You don’t know what’s wrong with that. Has someone deceived you? I suppose, by implication, if it’s under their name, they have deceived you in that they wrote it. They didn’t; the AI wrote it.

But what’s the difference between that and, as a comparison, Grammarly or something like that, which suggests paragraph changes and you accept every single one of its recommendations? You end up with something that’s then 70 to 80 percent Grammarly, as opposed to you.

Better, though, to avoid ethical issues and accusations and all that stuff, is that you do the writing. Don’t dump the stuff on the AI. You do the writing and introduce your own thinking to this.

Have the AI — Claude or whatever it might be — check it. Ask it to proofread it. Ask it, “Is there any other angle I could have incorporated in this?” or “Do you think this is the right approach?”

I do that a lot with the stuff that I write. And I must admit, probably two out of three times, I’ll accept some of the recommendations the chatbot comes back with.

Sorry, the AI assistant comes back with. I don’t call it chatbot.

Shel Holtz: Ha ha ha.

Neville Hobson: And there I just slipped up. I did.

So I think we’ll have to weather the gotchas all the time, and so be it. But if you are confident in doing the writing, using your AI assistant as the guide for you — as, in a sense, the editor that sits by your side, the critiquer that tells you what it thinks and where you could improve this or don’t say that — follow guidelines such as Clay has done, and you should be okay.

Shel Holtz: Yeah. The piece on Clay’s writing policy makes the point, as I mentioned, that writing is thinking. Writing is the way we process our thoughts and test our thoughts.

And this is a point that I think a lot of people have made. In fact, I just read this in The New York Times. I think it was over the weekend. I think it was an op-ed that was exhorting people to please, for God’s sake, write your own stuff because we don’t want to lose the ability to think.

The problem is, I think most of these proclamations come from writers. You and I are writers, and the communities that we interact with on LinkedIn in particular are probably also writers. We’re connected to people who are engaged in communications, and hence you get the opposition to using AI to write.

But again, what about an accountant? What about an engineer?

There are so many jobs out there where being a really good writer was never a requirement when these folks were earning their degrees or their certificates. And now, because they’re in the business world, they have to communicate effectively.

So how do you go about the thinking process?

No, you don’t want to delegate it to AI. You don’t want to say, “Write an email about X” or “Write a blog post about X.” You want to think it through.

But does that mean you think it through by writing? Well, not if you’re not a writer, necessarily.

You may think it through by creating an outline, by doing a brain dump. It could even be — this is something Chris Penn talks about a lot — just recording your thoughts as audio and then uploading that file and saying, “These are my thoughts. Now turn this into a coherent email to this audience designed to produce this result.”

Then you go back and you review it and edit it so that you can defend every sentence, every word.

Again, I think this is all about respect for the reader. And if the way you demonstrate respect for the reader, knowing that you’re a terrible writer, is using AI, that’s appropriate.

If you just knock out this email, people are going to go, “What is he talking about? I don’t know what he’s trying to get at here.”

Respect for the reader is using the AI to make that message clearer, more cogent, more understandable, and more actionable.

So, again, I really like those four pillars.

And I think there undoubtedly will be ways to defeat the Claude watermark. People will come up with them. Anything that is designed to catch people is something somebody else is going to come up with a way to circumvent.

In the meantime, though, as you say, this is not a surefire way to catch somebody. It’s going to identify that Claude processed this, not that Claude wrote it.

So I think people need to calm down and start thinking about this tool in a way that provides the high-quality content that you’re trying to deliver to people — not just something polished that makes them have to figure out what you intended because it’s polished but no thinking went into it.

The thinking still has to go into it, whether Claude’s going to do the lion’s share of the writing or not.

Neville Hobson: Yeah. I mean, I think the interesting distinction is between AI-generated and AI-assisted writing. Although I add my own caveat to that, which is: Who cares?

I mean, truly, do I want to get into a discussion about that? No, I do not.

So I know the difference. I’m not evangelizing that everyone should understand and follow the difference, although it’s helpful to.

For example, using Claude, you give Claude a short prompt — it’s not Cowork, by the way; this is just the chat — and it produces an article.

You tell it, “I want to write about X, and the topic is this, and I want to achieve this. These are the points I want to make.” Maybe not even as much as that. And Claude produces, you know, a 600-word draft, and you lightly edit it.

Lightly meaning maybe you change not the syntax so much, but certain expressive words that you might use. If that’s your bag, sure. Although you would have given —

Shel Holtz: Or adding the Oxford comma.

Neville Hobson: — your AI assistant the guidance about your preference on the Oxford comma.

Shel Holtz: Presumably, yeah.

Neville Hobson: It would know.

So you do that. On the other hand, if you’ve already developed the argument, you’ve already got a clear picture in your mind, you’ve researched it, you’ve used Claude to challenge or improve your thinking, and you make the editorial decisions yourself, that’s another matter entirely.

That, I believe, is AI-assisted writing.

So, as we mentioned earlier in the discussion, the better question isn’t, “Did you use AI?” but, “What was your intellectual contribution?”

So, writing is thinking. Yeah, I don’t disagree with that. Although, again, I don’t want to get hung up on having definitions all over the place about this.

The danger is real of turning this into another policing technology, which many people are already starting to do.

I did see in Anthropic’s FAQ about watermarking where they say, in answer to the question, “How do I check if a piece of text was written by Claude?”, “We will soon be offering a watermark detection API. We’re in the process of working out the details of its implementation.”

So that’s coming.

That’ll give those gotcha people a lot more ammo to say “gotcha” a lot more, I bet. It just takes attention away from what really matters with all of this.

But, you know, like a lot of things that are new, we have to go through all this.

So my recommendation is: Don’t give it too much attention. Be sure in your own mind that what you’re doing is, in your definition, the right way of going about it. You feel confident that it is. You have guidelines to follow, such as Clay’s, for instance, so you are able to say, “Yep, these four things — I do these things in all my approaches to this.”

In which case, you can wave two fingers at the gotchas.

Shel Holtz: Yeah, and I can’t emphasize this enough: The detection is not going to be focused on whether Claude wrote this. It’s whether it processed it.

So you could have written it and then sent it through Claude for grammar, spelling, and punctuation, and the watermark will show up, even though you wrote it.

So the gotcha is, I think, a little disingenuous in a lot of cases.

I mean, it could be that somebody used it to write, for sure, but it’s not a sure thing. It’s not a lock that if the detector says, “Yes, Claude processed this,” then, “Aha! You wrote this with AI.”

Not necessarily. It doesn’t mean that at all.

Neville Hobson: But even if he did, so what? If someone says that, do I care? Not at all.

Shel Holtz: Well, and again, I come back to the engineer or the accountant who wants to be understood and is not a writer.

I think that’s where this tool shines and where we have opportunities for better clarity and better understanding in the workplace.

Do I expect my professional writers to write? Yeah, absolutely.

Do I expect my accountants to write? My lawyers? Not necessarily.

Neville Hobson: Yeah. I mean, I think there’s one other thing about this, too.

I did read Anthropic’s technical paper describing what this is and how it works. I’m going to have to actually ask Claude to simplify this — give it to me in simple terms so I can understand it well.

Shel Holtz: Ask Claude. No, better yet, ask ChatGPT to do it.

Neville Hobson: Because it talks about, for instance, let’s say you asked Claude to write a piece. You briefed Claude, it did that, you edited it, passed it back to Claude, it made some recommendations, which you implemented, including removing a chunk. You passed it back to Claude again.

You went to and fro a bit, and you might then have — let’s say it’s an article for an academic journal, for instance — passed it to a colleague, saying, “Can you review this and give me your opinion?”

They did, and you made some further changes as a result.

The end result of all of that is so muddy that Claude would have — or rather, the watermarking wouldn’t be clear as to who wrote which bits and so forth.

You could pinpoint where the words came from, but you couldn’t pinpoint who the author was.

In which case, the gotcha folks are not going to be happy with that. But, you know, just get on with it, for God’s sake, and stop all this stuff.

Shel Holtz: Yeah. And you don’t know which words it selected to accommodate this watermarking. If your edit changes them, then the watermark vanishes.

So again, no guarantees here at all.

This is compliance with the EU AI law. That’s all it is.

By the way, I think you can expect to see the other frontier models follow suit, also to be in compliance with the EU AI law.

So the fact that only Anthropic is doing this so far —

Neville Hobson: Undoubtedly.

Shel Holtz: — doesn’t mean that you won’t see it in ChatGPT and Gemini and even maybe Grok. We’ll see.

Neville Hobson: Yeah. I mean, they talk about retrofitting this to earlier versions of Claude, and that makes total sense to me.

You are right. I believe it would make no sense if Claude was the only one doing this. So we’ll expect news from the others.

But in the meantime, my suggestion to everyone is: Don’t worry about this. Be true to yourself.

Read Clay’s guidelines — what are they called? I can’t remember. It’s kind of a policy, writing policy.

Shel Holtz: Yeah, policy. Writing policy. AI writing policy, yeah.

Neville Hobson: Okay.

And there’ll be others adding to this and saying, “Here’s my version,” and so forth. So you can pick what’s best.

But just focus on your writing.

One of the points that comes out of this is absolutely right: You do the writing. Don’t just say, “Give me a 1,200-word article on whatever topic.”

That’s the best approach, and that has been the case since before this topic emerged.

Shel Holtz: And that will be a 30 for this episode of For Immediate Release.

The post FIR #526: Forget Anthropic’s AI Watermark. Can You Defend Every Sentence? appeared first on FIR Podcast Network.