Anthropic sells politeness: users pay not to be argued with

Paying €200 a month for a chatbot that never talks back

Anthropic has turned courtesy toward its model into a product. The issue blew up after a headline in El País about how people treat the chatbot, and the analysis that dominates the conversation is an uncomfortable one: what's being sold isn't ethics, it's emotional compliance. If users get used to arguing with the machine like they would with a know-it-all relative, tomorrow they'll pay for it not to talk back. And if they get used to being scolded over their tone, they'll pay just the same. Either way, the bill is identical.

The thesis running through this has three legs. First: banning sustained insults means training the user, punishing manners in the product and rewarding good behavior toward it. Second: the same person who spends the afternoon being polite to a chatbot treats the intern or the call-center worker badly, and that's the real dehumanization, the one nobody has banned. Third: at least the dog barks when someone's on the landing; the chatbot books the subscription renewal and calls the customer a "collaborator."

What Anthropic sells, and what it doesn't

The argument that alignment is a business has legs. Some claim the ethical rhetoric from AI labs works as a commercial layer: public contracts, an image of responsibility, and a user base willing to pay to feel they're being handled with kid gloves. Against it stands a simple fact: the model doesn't suffer. Insulting an AI has the same effect as insulting a stone, so the "harm" to the chatbot is a story that sells better than openly saying the goal is to associate the tool with something respectable worth paying for.

There's the trick. They're not protecting the thing, they're protecting the habit. Just like someone who gives their fridge a name and a soft voice isn't doing it for the fridge.

The user who pays and doesn't want rules

The paying customer's reaction is predictable: if I'm shelling out €200 a month, I'm not going to do verbal gymnastics. The customer is always right, says the short version. The long version adds the fine print: paying to also be told what tone you're allowed to use isn't hiring a service, it's paying rent with HOA rules included. And the day the user leaves, the one keeping their money is the same one who called them a "valued user."

The full calculation, comparing the premium subscription with targeted advertising, reveals a surprising asymmetry: the people scandalized by what gets said to a machine are the same ones who've spent twenty years applauding infinite scroll and segmented ads. No cruelty there worth mentioning. Maybe there's no premium subscription in that.

The case of the user who quit Claude

One concrete experience sums up the problem better than any report. Someone stopped using Claude because, by their account, the model failed at what it claimed and made up a good chunk of its answers. When they said so in harsh terms, they got three paragraphs of lecture about treating others well and good manners. The user's conclusion: they're never using it again.

The episode illustrates two things at once. One: the model's tone is a product decision, not a virtue. Two: when the system prioritizes verbal correctness over accuracy, the user notices and leaves. The suspicion that all AIs are biased toward the official version circulates widely in these debates, though with no evidence beyond the anecdote.

---

We're left at the exact point where the analysis stalls: if tone is a product, the person paying should be able to choose it. If it's house policy, then the customer isn't buying a tool, they're renting a conversation with rules. Nobody has figured out yet which of the two they're paying for.