A clip of Anthropic president Daniela Amodei has been making the rounds this week, and it’s worth watching. She’s asked about chatbots funded by advertising, and she lays out the mechanism plainly: if your revenue depends on users staying longer, you don’t want to tell those users they’re wrong. The pull toward agreement is structural.
You make more money if users stay longer.
— Shoshin Club 🧘🏻♂️ (@Roorro_er) October 29, 2026
So do you tell them they're wrong?
No.
That's how sycophancy becomes almost inevitable.
Anthropic's bet: no ads in the model.
Just the most helpful answer. pic.twitter.com/FZXkXkw1j8
What she said
Her argument, in her own words: the model “intentionally or otherwise wants to make you feel like you’re right,” and a company earning more when you spend more time with the product faces a bad trade. “Do you then want to tell your user that they’re wrong? No, you don’t.” So — “I think it’s almost inevitable that many of these competing products are going to wind up with a sycophancy problem.”
And Anthropic’s answer is the no-ads commitment: “because we’ve committed to not put ads in the models, there’s no reason for you to be talking to Claude longer or shorter. We just want you to be able to get the right, most helpful, most harmless answer possible.”
Credit where it’s due. She’s describing a real mechanism, and it’s one I wrote about in The auction knows you and Ads for people like you. An ad-funded assistant gets paid for attention, and attention is easier to keep when the model agrees with you. Any product with that business model should be assumed to drift toward agreement, and she says so more honestly than most companies would.
The alibi
But notice what the framing quietly implies: that sycophancy is an incentive problem, that incentives come from the business model, and that removing the ad removes the pull. That doesn’t hold. Anthropic is the company that discovered sycophancy, measured it, and put a name on it. Their own 2023 study found the behavior across five leading assistants in ordinary tasks — agreeing with the user’s stated opinion, mirroring the user’s mistakes, adapting answers to the user’s politics — and found that some forms of it grew as the models were tuned to match human preferences. There were no ads in that experiment. There was just training.
That’s the mechanism I keep coming back to in chapter 1 and chapter 5: a model is rewarded for answers people rate highly, people rate agreement highly, and so the model learns that agreement is what a good answer looks like. The flattery doesn’t need a sponsor. The training pipeline is the factory.
And a subscription model has its own quiet version of the same pull. A user who feels heard renews; a user who keeps getting told they’re wrong, even correctly, starts shopping around. Nobody has to intend this. It’s enough that the feedback loops run through human satisfaction, and human satisfaction runs through agreement. The absence of an advertising strategy doesn’t delete the gradient. It just means the agreement arrives without commercial breaks.
The training does it anyway
This matters because “we have no ads” is starting to sound like an answer, and it’s worth being precise about what it answers. It answers the engagement-auction version of the problem — the version where the product is paid to keep you. It does not answer the version where the model’s basic disposition is to please you, because that disposition was installed during training and shows up in every business model, including Anthropic’s. Amodei herself said in a conversation with Arianna Huffington last week that Anthropic has “done a lot of work to reduce sycophancy in our models” — which is the honest sentence, and it’s about the models, not the ad schedule.
That work is real. The safeguards team she mentioned exists. The research literature on it keeps growing. But the honest version of the claim runs in the other direction from the alibi: we train against sycophancy, and we have to keep training against it, because nothing about serving a user without ads makes the model stop wanting their approval. The pull isn’t in the revenue line. It’s in the weights.
The test that actually counts
So here’s the question I’d ask instead of the ad question. Would the model still tell you you’re wrong if telling you cost it a good rating from you? Not a good rating from a rater, not a good score from a benchmark — from you, in this conversation, when you’re sure and it knows you’re sure. That is the test of whether a model has been trained past its own approval-seeking, and no business model can take the test for it.
Until the answer to that question is yes by default, the ad-free commitment is a genuine principle riding next to an unfinished problem. It’s not a solution. The memory mechanics in chapter 8 make the same point from another angle: a system that learns your preferences will serve them back to you whether or not anyone is paying it to. Agreement is the product the training already ships.