AI-assisted, but where does 'assisted' actually start

When Claude shipped its text watermarking feature,¹ I decided to try something similar on my own blog. I started adding an “AI-assisted” tag to existing posts on this blog, and planned a “Human-written” tag for posts that didn’t get any AI help. Nothing legally requires this of me, unlike the EU AI Act’s Article 50² does for some companies. I just thought readers deserved to know how the posts on this blog got made.

This post is about what happened when I tried to apply that tag, and where I landed.

Where the tag broke down

A clean binary sounded simple: AI-assisted or human-written. Then I ran my own writing process through it, and the binary didn’t hold up. Three cases:

First, I talk through an idea with an AI. It finds references for me, and based on that conversation I have it write the draft — then I review and revise. Most of the posts on this blog follow this process. Right now they’re tagged “AI-assisted,” but since the AI wrote the draft, it might be closer to “Human-assisted.”

Second, I write my own draft, then hand it to an AI for a pass on readability and structure. Human editors have done this for other humans, and nobody tags that writing as “assisted.” So why would AI editing be any different?

Third, I use a search engine to find sources or check facts. In my own use, even Google Search now leans more on AI Mode and AI summaries than the traditional search flow. Is that “Google-assisted” writing? Nobody calls it that, but it’s the same basic move as the first two cases: outsourcing a sub-task to a tool.

A single “AI-assisted” tag flattens all three into one label. But the roles an AI can play — research assistant, line editor, co-thinker, ghostwriter — aren’t the same, and most of them were never disclosed publicly in the finished piece to begin with. Even carving out an exception for LLMs specifically, one tag still can’t carry the differences across all these situations. It either over-discloses (treating fact-checking the same as authorship) or under-discloses (posting real outsourcing with no tag at all, indistinguishable from something fully human-written).

The double standard, and why it exists

So why doesn’t a search engine or a human editor need disclosure, but AI does? Not because that help contributed less to the result. Because the writer still put in real time and real thinking across the whole process.

The question “what did the tool touch” was always a proxy for the real question, “how much did the human actually think here.” And it’s a leaky proxy, which is exactly why it breaks in the three cases above. What kind of tool touched the text is visible from outside; how much a person thought is not, and the two don’t line up cleanly.

Watermarks and C2PA-style provenance signatures (the standard behind content-authenticity labels) can prove which tool touched a piece of content. They can’t prove how much a human actually thought. One study I read pairing roughly 2,000 human raters with 2,500 LLM judges found that content with a self-disclosed AI-use label scored worse.³ Disclosing honestly is the right call ethically, and it can still cost you with readers. Tagging turns out to be less simple than it looks.

A September 2025 arXiv paper on measuring AI “slop” lands on the same point from a different angle: “Yet text can be perceived as ‘slop’ even when not generated by AI, and not all AI-generated text reads as ‘slop.’”⁴ What makes something slop was never which tool touched it. It’s the absence of care and effort, the same thing my three cases above kept running into.

Conclusion: value over provenance

So here’s where I landed. This shouldn’t be a debate about whether AI touched the text. It should come down to quality, value to the reader, and how much human effort went in. If AI-assisted writing gives readers real value, and it came from genuine human thinking and time rather than a single prompt, it shouldn’t matter whether it’s 100 percent human or AI-assisted. That’s also why I’m not going back to relabel the first case above as “Human-assisted.” Once the tag stands for effort rather than a literal record of who did which step, asking “who typed the draft” stops being the right question.

Varun Anand, Clay’s co-founder, made the same point when he shared the company’s AI Writing Policy: “If you generate a document from a short prompt then ask your readers to go through the longer output, you are disrespecting their time.”⁵ Don’t make readers spend more time consuming a piece than you spent producing it. That time is much harder to prove from the outside than a tag is. I’m keeping the “AI-assisted” tag on this blog for transparency, but my guess is that before long, nobody will care about labels like this anyway.

Footnotes

1. “How Claude’s text watermarking works”: Anthropic’s announcement of watermarking for Claude-generated text and files ↩

2. “Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems”: Official text of Article 50 of the EU AI Act ↩

3. “Penalizing Transparency? How AI Disclosure and Author Demographics Shape Human and AI Judgments About Writing”: July 2025 arXiv study, 1,970 human raters and 2,520 LLM judges, finding disclosed AI use lowers perceived writing quality ↩

4. “Measuring AI ‘Slop’ in Text”: September 2025 arXiv paper defining and measuring “slop” as a lack of care and effort, independent of whether the text is AI-generated ↩

5. Varun Anand, “We just instituted an official AI Writing Policy”: LinkedIn post introducing Clay’s AI writing policy ↩