AI compliance appears to have at least one consistent feature: as of now, it appears nearly impossible for commercial AI models to segment their capabilities and compliance requirements based on location.
This is partly why Anthropic had to temporarily pull Fable 5 and Mythos 5 from the market – for everyone – when the U.S. government imposed export controls. And, it’s also why Anthropic is globally embedding a mathematical AI-generated watermark in text generated by their models, even though the stated impetus is compliance with EU law.
To the lawyer in me, it’s fascinating that basically the most restrictive compliance rules become the de facto global standard. Not because the law requires it, but because commercial AI as currently constructed cannot appropriately segment workloads geographically.
I see this as an engineering challenge that might later better differentiate the commercial AI offerings, but I digress.
So, Who Gets It?
(Note: As of August 13th, 2026, ChatGPT allegedly does not embed this kind of watermark in ordinary text output. That could change at any moment.)
When it comes to text, commercial LLMs today are overwhelmingly used for 2 primary functions:
- To write code.
- To write long-form text, whether that is emails, a blog post, a social media post, a Substack article, etc.
I have no idea how an LLM would ever embed a statistical watermark in code, mostly because code requires rigorous syntax – but even if it could, very few people will care. So, it is that second use case that is directly in the line of fire with these embedded watermarks.
Unlike code, some people likely will care about AI embedding a hidden watermark in their prose.
While it will “out” those who just use AI to generate their ideas and text – the AI slop – those folks were never fooling anyone to begin with. And, when it comes to emails, nobody is going to care that there’s now a watermark hidden in the depths of an email AI helped them draft to “sound better, sound less angry, be more concise” or the like.
Who this watermark hurts: the people who are writing drafts themselves, ideating themselves, and then just using AI to sand off the rough edges. These folks are using AI to accelerate the last mile of content delivery, not using AI to create the package itself.
The AI watermark, however, makes no distinction between AI-generated and merely AI-assisted.
It is a binary blunt force tool, so any text AI touches at all gets the same digital watermark.
Keep in mind, AI has arguably already started the decay of writing as a skill. Many students have now outsourced their formative writing development to AI. The em-dash has been forever ruined. Anyone who actually is a good writer or has a strong vocabulary is now eyed with suspicion.
And, that’s before we point the finger at the masses who have substituted AI for their own thinking, believing both that nobody will notice (we do) and that they will look smarter (you don’t).
Me Too
I’m not immune here, either.
I run a weekly series on LinkedIn called “Punk Stack Fridays” that I love doing, even though maybe 4 people have noticed. I’ve been a huge music nerd my entire life, especially around pop punk and emo. I know the genres inside and out, and AI allows me to create a fictional band I can direct to bring lyrics I’ve written (about mundane tech and data topics I’ve encountered in my career) to life.
The key here: these are my lyrics. Love them or hate them, I wrote them.
And, I also recently found my voice and started writing blog posts again. I have always been a strong writer – a hazard of being educated as an attorney, I suppose. It’s genuinely frustrating to see good writing treated with suspicion now, because of AI.
I could say “and that’s the work” here, but then you’d think it’s all AI. (I wrote this blog article myself, same as all my blog articles – but I will at times use AI to help me be more concise, again because lawyer.)
But that is the work.
We do need ways to identify copy-and-paste AI-generated content, as it is not just lower in value and reverts discourse to the mean – it also drowns out legitimately unique and human voices. But, we need to do this in a way that doesn’t treat every use of AI as a crime.
I am open about using AI on my Punk Stack Fridays project, because listeners have a right to know what they are hearing – and because I have a day job. There is no way I could write a song, book a studio and a band, produce and mix a track, and get it released weekly.
AI is an accelerant to bring my musical ideas across the finish line, and I suspect it’s the same for many writers. The difference lies in how the public penalty for “using AI” changes depending on medium.
For Punk Stack Fridays, you may not like the music or understand the message, but the fact I am using AI to help generate the music doesn’t really have much of a blast radius.
The risk to creative authors, CTO candidates publishing thought leadership articles, etc. is much wider – the black mark of an AI stamp could invalidate their entire body of work or harm their professional aspirations, even if all they did was use Claude to help them condense three sentences into one with punchier phrasing.
There’s also some cruel irony here. Commercial AI models hoovered up everyone’s content, repackaged it and put it behind a subscription, and gave no attribution watermark (or compensation) to the actual sources. Now, they are laying wide claim to anything you might use them for, too.
Heads they win, tails you lose.
The Future Is…
Watching this unfold is neither enraging nor surprising.
I believe AI is a generational shift, much like the commercialization of the internet or the introduction of smart phones. Each brought new challenges, unseen issues, and a high velocity of change.
This kind of disruptive change always brings over-reach, great adaptations, and sometimes just plain bad ideas.
Only time will tell us which label the AI watermark concept earns, and whether or not this is the moment Anthropic jumped the shark.
I do think this change may cause downward pressure on personal subscriptions to Anthropic in the short run, as I suspect a non-negligible percentage of those subscriptions include AI slop content creators – whose investment in Claude now carries downside risk that an investment in, say, ChatGPT currently does not.
Beyond that, my hope for the rest of us is that perhaps the AI watermark will help revive good human writing, although I still think the em-dash is permanently dead.
And if you haven’t seen my Punk Stack Fridays series yet, you’re officially invited to become its fifth fan.
–Scott