Maksim Barziankou (MxBv)
Navigational Cybernetics 2.5 · The Urgrund Laboratheory
research@petronus.euA side note away from Navigational Cybernetics and the research around it.
Anthropic announced a text watermark: supposedly to mark generated text. Without byte-level verification, a cryptographic signature, or another hard binding method, their proposed check remains a poke into the world of statistics. "Presumably" can easily turn out to be the opposite of the truth.
A separate question: what does Anthropic itself make of cases where a user's work gets labeled Anthropic-generated, while the model only cleaned up 10-20 words for brevity? Who guarantees that in that kind of "edit" it does not swap exactly the semantic spots their statistical trace then sits on? Everyone already understands that model help has become normal. One thing is when the text is yours. Another is when the model replaced, say, 30 words out of 10,000 and a detector hangs a generated label on it. I once ran a fully handmade text through an ordinary public detector, with my own punctuation and turns outside typical generative cliches. It still scored 80+. That is not about their private key; it is about how fragile the idea of "guess from statistics" is.
The EU-regulator excuse sounds convenient, but it does not remove the question of choice. You can mark differently and leave the user in control of their own artifact. Intervention at the process level, even dressed as "harmless" and "necessary", is for me a signal: the provider's link to someone else's work is being held at the level of permissibility. Hypothetically, in 7-10 years that link could turn out very differently. It sounds implausible, I agree. Still, it opens a whole region of trajectories that used to feel far from execution. Large players know how to play the long game.
And then, almost in the same window, SpaceXAI updates the Grok Consumer Terms. Agentic features, autonomous actions on your behalf, Inputs, Outputs - when something breaks or the agent freelances, the liability sits on the user. Even if you are out walking and never typed "send them an insult". Very convenient. Very "noble". Strip every ounce of responsibility from the provider and leave the human holding the bag for a remote machine they do not even have to be present for. That is how large players like to behave when they call themselves giants.
And this keeps happening. In the pace of ordinary life there is no time to talk it through with anyone. They ship a model - slow it down - then sell a "feature": speed. Or there is quiet training and data retention, and when the company gets caught it rolls back and acts as if nothing happened. It did. The pattern is no longer a one-off move. It is a style of play: first an experiment on the user, then a rename into a feature or a quiet rollback with no admission.
I am not condemning Anthropic or xAI. They are large companies, and this kind of move sits inside the normal range of how they behave. As I see it, they are not doing it with deliberate malice. I am hinting that it is worth looking where things are thin. Right now they are thin in more than one place. I may be wrong, and I would rather be wrong. But squeezing choice without a real alternative, and dumping agent failure onto the person who was not even at the keyboard, forces you to rebuild your stance toward providers as a class.
Even in their logic of "mark it and save face" or "clarify liability" there are cleaner solutions than the current ones. I hope someone on their side arrives there on their own. We will not, of course, tip them off.
I rarely write in this direction. A chain of events simply forces a different look - and a bit of envy at Anthropic's promo: how, with so many contested experiments on users, they still hold the role of antagonist to the "great and terrible" GPT.
What do you think? Is this a set of coincidences - or the familiar pattern where large players change the rules of the game and then explain that this is how it was meant to be - I come from a world where that kind of move was called a piano in the bushes. And in a normal room, if there was an alternative, that kind of counterpart got walked out of the negotiation by the ear.
NC2.5 · Urgrund · MxBv · PETRONUS · research@petronus.eu · CC BY-NC-ND
Originally published on LinkedIn: https://www.linkedin.com/pulse/where-things-thin-anthropics-watermark-groks-agent-max-barzenkov-svede/
Top comments (0)