Watermarking worries

If you use AI to write every sentence, you were already in trouble. But if you say “I never use AI to help write,” I call BS.

GTM leaders need to help their teams find the line.

In my experience, AI is great at writing code but terrible at writing content. Anyone using AI to crank out full articles, web pages, and social posts is not going to see good results long-term. Maybe you get some short-term SEO or AEO benefit, but that will fade quickly.

Two things happened last week on this topic.

1.) Monday. Clay published their AI writing policy. First principle is “You must stand behind every idea and sentence.” Varun’s post had 9,000+ likes and 620 reposts, the highest-engagement GTM post in my panel all month.

The engagement numbers tell me that people are craving this kind of clarity:

2.) Tuesday. Anthropic announced watermarking. This caused a big spike in the topic I track for “Watermarking & Content Provenance.” It’s broader than just “watermarking” but that term drove the spoke. You can see it trending here:

Data note: This is based on 20K+ conversations analyzed each month from my curated panel of GTM voices. Share is engagement-weighted across measured topics. August is 17 days, and share is a rate, so the comparison holds. More on methodology here.

The spike was driven by fear that watermarking could penalize marketers who use AI.

There were two classes of reactions:

Lily Ray summed up both sides perfectly on X:

I would propose a third response for marketing leaders — use this as an opportunity to clarify your AI content policy.

Watermarking itself is not a big issue for a few reasons.

1. It’s just Anthropic responding to European regulators. They signed the EU’s transparency code and applied it broadly. Every other big LLM signed the code and will do the same thing.

2. It mostly shows up in longer text. The why is clear. What’s interesting is how. It’s not metadata and it’s not hidden characters. The mark lives in the actual word choices of the output, so the longer the passage, the more there is to detect. But be clear on what that doesn’t cover. The mark follows whatever words Claude picks, so proofreading counts. So does translating and summarizing. Anthropic is blunt about the limit: their key “can only answer the question ‘What is the likelihood this was partly written by Claude?’”

3. If you’re using AI to generate entire pieces, Google and LLMs already know it. The watermark doesn’t change that. Eli Schwartz said it well on LinkedIn:

Two other details worth knowing (to avoid panic): First, the marks don’t have any identifiers — Anthropic says they aren’t tied to a person or org or a chat. Second, the detection API doesn’t exist yet, so only Anthropic sees it today.

If you set clear principles for your team, and you aren’t mass generating long-form content, then it’s a non-issue.

But the move is part of a bigger arch around authenticity.

As readers, we know what smells like AI. It’s a big turn off when we catch that scent online. That sentiment has made “AI Slop” a big topic in the broader GTM x AI conversation. “Slop” even earned word of the year in 2025 (vibes and maxxing said the judges were bribed). That’s the black line on the chart above. It’s had significant volume for 6 months.

Your audience is increasingly aware and allergic to it. As the cost of production falls, the bar for quality increases. People will ignore things that smell like AI and gravitate toward more authentic, uniquely human content.

We should all be trying to play a quality game, not a volume game.

The best thing GTM leaders can do is establish an AI Writing Policy. A great policy would cover:

1. Principles — What are the team’s guiding beliefs? Great example here from Varun at Clay.

2. Boundaries — Where’s the line, and who owns the output? The test is about who did the thinking, and who stands behind it. Manisha Raisinghani’s framing landed for me: “A watermark can tell you a model touched the words. It can’t tell you who had the idea, who made the decisions, or who stands behind the output.”

3. Best Practices — How can we make the best use of AI in our writing and content production? What is acceptable use, what is slop? How can and should you use AI to help you write better?

Want some ideas for things to include and exclude? Let me be a case study of one.

I’ve been living in Claude Code and Codex for the last 10 months and I publish a lot of content (27 pieces across LinkedIn and newsletters in the last 60 days). AI doesn’t write my sentences. Here’s everything it does to help me write better.

To give a real-time example, here’s the prompt I used for my first draft of this newsletter:

Grade my structure. Fill in my links. Check my claims. Find the best sources from my panel. Nowhere in there did I ask it to write the thing.

Generally, I think of AI helping in 5 categories

I’m sure this is an incomplete list, but these are the places I’m finding value today.

I enjoy writing so even as AI gets better, I still expect to be hands on keyboard for most of the content I produce (and voice on Whispr Flow). I want to own the ideas and sentences, but I’m happy to have AI help me edit, fill in the blanks, and improve my formatting.

Claude is a bad writer alone, but it can help make me a better writer.

Do you have an AI writing policy yet (or an approach that works for you)? If you’re willing to share, I’d love to see it.

I read all replies.