What Are We Protecting?

By Ryan McMillian

Here's the tension: you have a draft. You've written it and captured the main ideas, the bulk of what you're trying to say. You take it to an AI language model and prompt it to help you formalize it, edit it the way a good editor would. It tightens your sentences; helps you formulate your thoughts more clearly. You do a final pass, make it yours again, and publish it.

The watermark is now stamped.

The tag says AI. But it doesn't say what the AI did. The nuance is missing. It doesn't say you wrote the ideas, built the structure, lived the experience that produced the thinking. It doesn't say the tool was the editor and you were the author. So, to whoever reads the tag (most won't read past it), AI made this. The results: your ideas get discredited, and your voice gets attributed to a machine.

This is the concern that is getting overlooked in most of the discourse online. The question isn't whether AI should exist, or whether companies should disclose their tools when used, but whether the mechanism being used to offer disclosure is telling the truth about what is actually happening.

Recently, the EU adopted a law, the AI Act, Article 50; the regulation that companies like Anthropic are responding to; the stated purpose is to address misinformation and manipulation at scale, fraud, impersonation, and consumer deception. These are real concerns, and it makes sense that legislation is being adopted to address them. A deepfake of a public figure saying something they never said. A synthetic voice used to impersonate someone in a position of trust. AI-generated content presenting fiction as fact. The concern is legitimate, and the problem is worth solving.

To be fair, the Code of Practice that companies are signing onto tries to make the distinction you'd want made. It proposes a distinction of categories: fully AI-generated content carries different requirements than AI-assisted content. There's even a carve-out: it argues that the marking isn't necessary when the AI performs only an assistive function, where it doesn't significantly alter the original data input or if it doesn't alter the meaning. The regulation, in other words, knows the difference between an author and an editor. However, Anthropic's implementation didn't draw that line. It marked everything, and I mean everything. Any text Claude touches, any image rendered. And not just in the EU but globally. Beyond what the law strictly required, and without waiting for the distinction in categories to be fully specified. It's interesting to note whether that overarching response came from the complexity of defining the line precisely, or whether it was the legal risk of getting the calculation wrong, or even maybe a brand decision to overshoot rather than optimize; all of which is understandable, of course. Yet none of those reasons are about protecting the person at the keyboard.

As I continued to reflect, a few possible motives behind it all started to formulate in my mind, not the regulation specifically, but the anxiety running underneath it: I wonder if flagging AI use protects against the fear of replacement. I say that carefully. It's a suspicion for sure and not a fact. But I suppose I should state it plainly: when a technology threatens to replace a category of labor, the people whose labor is threatened tend to reach for legitimacy frameworks. I suppose one could argue that if the tool can be branded inauthentic, maybe it loses adoption. Maybe the old ways stay viable a little longer. Maybe we can keep things as they are. Yet it is important to make note that even if true, that fear isn't automatically wrong. You can be afraid of something real and still make the right call in response to it.

The question then is whether the mechanism chosen to respond solves the problem it was designed to solve or solves a different one and declares a victory.

I'll add, though, the person passing off fully AI-generated content as their own, which is the actual target of this regulation, has every incentive to find workarounds. In fact, I've read responses on posts concerning the topic like: Screenshots strip metadata. Heavy editing degrades text watermarks. Open-source removal tools already exist. You see, the mechanism is easily circumvented precisely by the people who would be most motivated to do so. However, for the person who wrote something, brought it to the tool for editing, and published a refined version of their own thinking, they can't un-ring the bell. The watermark travels. The audience sees AI and stops there.

Another double standard I know: we often apply double standards when we don't know what we are actually trying to resolve. It's worth reflecting on a bit longer though. We don't tag the graphic designer whose job was producing typography before AI existed. We didn't demand companies disclose which software rendered an advertisement, which algorithm suggested the layout, which tool processed the image. We assumed the product was the thing worth evaluating, not the production stack behind it. With AI and particularly within creative spaces, we want something different. Whether that's justified depends on whether AI represents a shift significant enough to require different rules. That's an argument worth making seriously… it hasn't been.

As for OATHBORN, we use the tools. Gladly. Deep research, refinement of thought, editing, content creation, image rendering. It allows the work to exist at a scale that would have been impossible otherwise. It allows me to do what I value most: think, read, write, strategize, contemplate, iterate, etc.

The regulation is trying to protect something. The online discourse is trying to protect something. I'm just not sure either has agreed yet on what it is.

Tools of the Trade — Written by Ryan McMillian. Developed in conversation with Claude (Anthropic). Sources: EU AI Act, Article 50 (2024); EU Code of Practice on General Purpose AI (2026).

The OATHBORN Audit Field Manual — 19 pages. Seven rules. Three layers each: the Standard, the Distortion, and the Audit question that forces you to see which one is operating right now. Built for those who carry weight and want to carry it well. Get the Manual — $27.