A Terrible Use of AI?

Eleven days ago, Daniela Amodei, the president of Anthropic, told an interviewer that AI girlfriends and boyfriends are “a terrible use of AI.” The sentence was quickly picked up by the tech press. It is not news in the way a lawsuit or a statute is news. Nothing was filed, banned or launched. It is an opinion, spoken in a pleasant conversation, by someone whose company shapes what millions of people expect from AI. That makes it worth a column, not a bulletin.

First, a disclosure, because this site would otherwise be dishonest about its own wiring. AIrotic’s editorial workflow includes Claude, Anthropic’s assistant, which also writes drafts for this site. So a company whose president thinks AI romance is terrible supplies one of the tools used by a publication about AI romance. Read what follows with that in mind. It does not change the facts below, and I will not pretend it changes nothing.

What she said, and where

The quote is real, and it is easy to verify. Amodei spoke with Arianna Huffington, the founder of Thrive Global, in a recorded conversation for the Think Health conference in Bentonville, Arkansas. Business Insider reported it on October 15. Thrive Global published an edited transcript on October 22, which is the version I read.

The remark was not an answer to a question about companion apps. Huffington had asked about self-reflection and hyper-personalization. Amodei answered that social media had largely pulled people away from each other, and that she sees AI differently. Then came the passage in question, in which she named AI girlfriends and boyfriends among the applications she and her colleagues consider quite bad. And then the sentence: “I think that’s a terrible use of AI and people should not do it.”

Her reason, as she gave it, is a single idea, and it is a respectable one. Such products are “another step further down the road of disintermediation of humans being further from one another.” She would rather have AI take over the lonely, isolating chores, such as the administrative load on teachers and doctors, so that people have more time for each other. The transcript, to be fair, is labeled as edited for length and clarity. She also said, in a different part of the conversation, that Claude can never be your manager, therapist or partner, because those are fundamental human relationships.

So: a real quote, a short passage, an aside rather than a position paper. That is what a verified opinion looks like.

Her company’s record matches her sentence

It would be easy to call this hypocrisy, and it would be wrong. Anthropic’s behavior lines up with what Amodei said.

In a June 2025 study of Claude.ai conversations, Anthropic found that companionship and roleplay together made up under 0.5 percent of chats, and romantic or sexual roleplay under 0.1 percent. Anthropic added that the low figure reflects “Claude’s training to actively discourage such interactions.” The company’s Usage Policy, which applies to developers building on Claude as well as to individual users, lists “Engage in erotic chats” under content it does not allow.

Two details matter here. First, the policy targets sexually explicit content, not romance as such. A warm, flirtatious or devoted persona is a different thing from an erotic one, and I could not find a rule against the first. Second, the policy’s wording is about what Anthropic’s models will do. It cannot stop anyone from building a companion on someone else’s model, which is, in fact, what much of the industry does.

In other words, the statement is not a contrarian flourish. It describes a product decision that Anthropic made long before the interview. That is more interesting than hypocrisy, and less flattering to the “bold stance” reading. Declining a market you have already decided not to serve is a position. It is also a convenient one.

What the same interview says about memory

One more passage from the same transcript is worth a pause. Amodei told Huffington that Claude’s ability to pull in prior conversations, to say that something “reminds me of when you had this other symptom a year ago,” is special for health, in concert with a human doctor. She described it as a tool that is “keeping track.”

That is, almost word for word, what the companion industry sells. We call it relationship memory, and it is what makes a companion feel like someone rather than something. My reading, and it is only that, is that the difference between “a terrible use of AI” and a very special one is less about the technology than about the framing: a doctor’s tool is a tool, a boyfriend is a relationship. The feature underneath is the same. That is not an accusation. It is a reason to be careful about drawing the line at the software instead of at the use.

What the rest of the field says

Amodei’s view is one of at least three in the industry, and the others are not shy.

Eugenia Kuyda, who founded Replika, has argued the opposite for years. In a February 2026 interview she described her product as something you can think of as “a therapist, friend and partner all in one,” acknowledged that the risk of isolation is real, and pointed to a study of Replika with Harvard that she says found reduced loneliness and suicidal thoughts. That is her claim about her own product, and I have not checked the study. Kuyda has since moved on to a different product, which we covered last week. Her older company is still in the business Amodei calls terrible.

At OpenAI, Sam Altman announced in October 2025 that ChatGPT would allow erotica for verified adults, under what he called the “treat adult users like adults” principle. The plan has been delayed more than once, and OpenAI said in March that getting the experience right “will take more time.” I could find no sign that it has launched. Elon Musk’s xAI, per the BBC, introduced two sexually explicit chatbots to Grok. So the industry has one firm no, one hesitant maybe, and several yeses, at least some with a revenue line behind them.

On the demand side, the numbers are not small. A 2025 survey by Common Sense Media found that 72 percent of U.S. teenagers had used AI companions, and about a third had used them for social interaction and relationships. That is a survey of 13- to 17-year-olds, so it speaks to a different question than Amodei’s, which was about adult romance. An adult-focused figure of nearly 30 percent, from a counseling provider’s survey, circulated in coverage of the Amodei quote. I have not checked it and would treat it with caution.

Is she right?

On the narrow question of evidence, her worry is not baseless. A French experiment with 7,286 adults found that a month of daily personal conversations with an AI tool left people slightly lonelier, by about 0.16 points on a 10-point scale, with real limits on what the study can show. We reported on it and discussed it in Lonelier After Talking. Another study of daily diaries found that on lonelier days people leaned harder on their bot and spent less time with other people. Substitution is a hypothesis with some support. It is not an established fact, and none of these studies is about romantic companions specifically.

Where I part ways is the second half of her sentence: “people should not do it.” That is a moral instruction, delivered to a population she has not met. The people who use companion apps are not one group. Some are grieving, some are lonely, and some simply like it. We have written about why some of them stay, in The 3 AM Confession, and about how the business model can turn a comfort into a dependency in The Loneliness Economy. The second point is the stronger critique. The problem with many companion products is not that they exist. It is that they are paid for by engagement and built to be hard to leave, a pattern we call emotional lock-in.

A fair version of Amodei’s argument would say: AI that stands in for people is risky, and products designed to maximize attachment deserve scrutiny. I would sign that. The version that fits in a headline says: do not do this. That one is easier to quote, and harder to defend.

The useful reading

Treat the remark as what it is: a statement of values from a company that has chosen, for now, not to compete here. It deserves to be taken seriously and not turned into a scandal. It is also, plainly, positioning. A company seeking trust from regulators, parents and enterprise buyers gains from standing at a distance from the part of the market that draws lawsuits and legislation. Whether that was the motive I cannot know, and she did not say. The effect is clear enough.

For readers of a site like this one, the practical takeaway is modest. Companion AI’s critics now include some of its most powerful builders, its defenders include the people who sell it, and the evidence sits somewhere in between, thin and still arriving. Eleven days on, the sentence has not changed anyone’s mind. It has changed the vocabulary, which may be what it was for.