Hi, Reader!
​
If your feed looked like mine last week, you saw some version of "Claude is watermarking its text now," followed by a lot of speculation, a round of "I told you so" from folks who've been wary of AI all along, an equal round of "But the ideas are mine!" from the other direction, and a few people announcing they were switching tools. Keep in mind that social is always going to amplify the extremes on both sides in any situation. That's what gets engagement and keeps eyeballs on their site for them to make money on the ad slots they sell.

I wanted to wait out the hype and do some digging on my own to give you the information you can use to make an informed choice. Like anything with AI, the context matters as much as understanding what it really means for your business. This includes the choices you make in how you want to run things and the level, if any, you decide to incorporate this into your business tool kit.

Why is this happening?

The EU's AI Act's transparency rules went into effect on August 2nd. Around 190 organizations have signed their Code of Practice on Transparency of AI-Generated Content, including Anthropic (Claude), OpenAI (ChatGPT), Google (Gemini), Meta, and Microsoft (Co-Pilot). Every major player is on this list. Some of them have been watermarking since before this went into effect (Gemini since 2024 with SynthID, OpenAI has been doing it with audio and images with text on the way, Anthropic started it this month.) See also Article 50.

Think about the last time you were on social channels or even YouTube - you've seen "AI Slop" and have likely seen the comments that people are calling it out and tired of it. This law is largely in response to this sentiment online and people using the tool to create vast quantities of content, but nothing of actual quality. You've also likely seen things getting flagged as AI on things like Instagram and Pinterest. Pinterest now has settings you can use to have it show you less AI content and I'd expect other social platforms to follow as they see a drop in users because of the "AI Slop" issue.

What is the "watermark" and how does it work?

The watermark that people are scrambling over is one related to text watermarking. This can show up as "hidden" formatting when you copy and paste the code directly, but it's more than that.

Generative AI is merely (and always has been) a very advanced predictive text model, which is how the whole "watermarking" system works to begin with. This watermarking works by nudging the model's word choices in a pattern only the provider can read. The mark lives in the specific words the model picked based on probabilities of the word ordering. Anthropic says so in their own documentation: the watermark "may persist through some editing," and it can disappear once text "has been heavily edited, paraphrased, translated."

The watermark survives exactly where you didn't touch the draft. It's a record of how much of the output you published as-is.

Something else worth flagging is people mashing this type of watermarking with AI "detection" systems. Older AI detectors guess, by measuring how predictable the writing sample under analysis is. That's why things like the Declaration of Independence and the Constitution get flagged as AI-written by those same tools. They've been cited and reproduced so many times that they read as the lowest common denominator of the internet. Watermarking is a different animal from detection. It's embedded at the moment of generation with a key only the provider holds, so it can only appear in text that model actually produced.

What's getting lost in the noise

The regulation (Article 50) has explicit carve-outs for content that has clearly undergone human review and substantive edits to what was previously generated output from the tool. A spell-check pass and rewording a few things here or there to make it sound more like "you" doesn't count.

The watermark behaves the same way. It survives in untouched sections of the model's original wording. Light paraphrasing dilutes it. Complete rewriting removes it. Two separate systems, one written by regulators and one built by engineers, both drawing the same line: did a human actually do noticeable editing work on this or just use the content output our systems came up with?

If you're using AI anywhere in content production, that means full editing passes and rewrites in your own voice, from your own brain, otherwise the "may have been created by AI" still comes into play. Note the phrasing itself hedges on "may" not "was".

Using AI for workflow and operations is a separate conversation, even if tangential to this one.

How does this impact creators that leverage AI?

Creators I've followed for years have started sounding the same, so much so that I've disengaged with their content, email, and even programs. I doubt I'm alone in this behavior.

Most of them believe they've trained AI on their voice. What I keep seeing instead is copywriting tropes and empty words where their perspective used to be. I even did a recent post on this on my LinkedIn that sparked this whole rabbit hole for me. (You're welcome to connect with me there!)

There's a reason why AI has those tells and it comes from the human copywriting and just what a lot of people were taught of what good marketing copy is or isn't. Copywriting formulas and frameworks existed before the internet, AI just brought them into play at scale so those patterns are more easily recognizable now. (I will die on the hill of the oxford comma and em dash usage when warranted or for ease of reading.)

Your readers are already running their own detection systems. The results show up as unsubscribes, scrolling past your posts, or flat-lining open rates. You can see those conversational signals showing up in your own numbers long before you should be concerned about anything flagged with a watermark.

Who actually owns AI-assisted content?

The second layer to this is on the legal side when it comes to IP protection and copyright. Note I'm not a lawyer, I just read their emails and used to work with Disney lawyers to enforce copyrights regarding Star Wars Costuming Group IP they technically owned. One of them is Amy Nesheim, a lawyer for creative entrepreneurs. I'm going to summarize here, but you can read her email on copyrights and AI here if you wish.

US copyright protects fixed expression (a social post, video, book, etc.), actual written words and works. It doesn't protect the idea behind them. So the very common position of "the ideas are mine, I just have the AI do the writing" works backwards from how the law actually operates. The Copyright Office has held that a user of generative AI doesn't have enough control over the output, even with a detailed prompt in a system heavily trained "in your voice", to count as the author.

Amy's recommendation is that it's safer, copyright-wise, to have AI generate 10 concepts and then write the piece yourself, than to write a detailed prompt from your own original idea and publish whatever it gives you back, even with some light editing.

To reiterate - I'm not a lawyer and this isn't legal advice. Amy is and she covers this better than anyone I've found. If you'd like to check out her podcast/blog you can find it here.

Does this matter if you don't use AI?

Another reality you'll deal with if you've published anything online, whether you use AI tools yourself or not: deciding how you want to handle AI training, scraping, and search bots. If you don't set anything up for what you allow on your site, your work is likely getting scraped and used to train AI models without your consent. At the same time, you don't want to disappear from traditional search results or lose out on being cited when someone searches using an AI tool and your work is relevant.

Not all crawlers are the same, and not all of them are "bad." I wrote about how to tell them apart here: How to block AI training bots from harvesting your content without disappearing from search. I also included exactly what you need to get in place if you want that protection for your content.

So where does this leave you?

The headline will fade in a couple of weeks. The question underneath it won't, because IP protection and copyright are going to matter whether or not the EU or others set a deadline and whether or not you make the personal choice to use AI in your business.

What I want you to take away is that we should be having open discussions with our audiences about how we're using AI and what our individual boundaries are with it. I'd rather be transparent in my thought process and approach than pretend I have a blanket position on all of it. I used AI to help me structure this article from a massive brain dump of the patterns and things I'm seeing, but the words are all mine.

If that's the kind of conversation you want to be in, you're in the right inbox. I'd love to hear your thoughts. Hit reply and tell me where you draw your line.

Ashley Clayton Marketing Ecosystem Strategist Because marketing is an ecosystem, not a tactic iterateology.com

​

P.S. There's also measurement aspect to this worth mentioning. AI is turning into a discovery channel in its own right (whether you use it or not), and only about 30% of AI-referred traffic shows up in your analytics and requires deliberate setup work to give you that visibility. Most business owners have no idea how much of their traffic is already coming from it and I'm finding it's one of my most highly engaged channels. Let me know if you'd like to learn more about how to get it set up for your business.