Autumn 2019, Oxford Street, London, in the rain. A 31-year-old computer scientist stops two women, pretending to ask for directions. They reply politely and walk away. His coach, paid £700 for the day, sends him after them: one of them crossed her legs while speaking, and that’s a signal. She tells him she has a boyfriend. Sirin Kale, the Guardian journalist shadowing the workshop, notes that almost every refusal that day took this form, a real or invented boyfriend. For the coach, none of them counted as a refusal. Standing in the rain, he explains to his students that by approaching these women they are giving them pleasure, and that depriving them of it would be selfish. The same article estimates the industry at $100 million, selling for fifty years roughly the same idea: a woman’s “no” is merely a technical obstacle.

We are defending a thesis here that will irritate pickup coaches as much as their critics. When it comes to learning how to meet someone, an artificial intelligence that doesn’t give in easily is worth more than any of them. We are talking about a machine you cannot wear down by persistence. Virtual companions that say yes to everything are the other half of the problem, and they will face their own reckoning below.

From Ovid to Mystery

The seduction manual is older than Christianity. Around the year 2 AD, Ovid published the Ars amatoria, three books of advice in elegiac verse, the first two addressed to men. It contains charming tips, such as drinking from the cup exactly where her lips touched it. It also contains, in the first book, the doctrine that the genre has been copying ever since, all the way to TikTok.

“Haec quoque, quam poteris credere nolle, volet”: the very one you think does not want it, will want it (line 274, our translation).

And further on, this couplet: “Vim licet appelles : grata est vis ista puellis ; / quod iuvat, invitae saepe dedisse volunt”. Call it violence if you like: that violence pleases girls; what pleases them, they often want to have given despite themselves (lines 673–674).

Title page of a German translation of Ovid's Ars amatoria, printed in Frankfurt in 1644.
Title page of a German translation of Ovid's Ars amatoria, printed in Frankfurt in 1644.Photo: Matthaeus Kempffer (publisher), retouch Pacha Tchernof, public domain, Wikimedia Commons

His defenders say Ovid was joking and that the poem is ironic from start to finish. The ironic recipes still circulate. The same book contains a less-quoted line: “Quae dant quaeque negant, gaudent tamen esse rogatae : / ut iam fallaris, tuta repulsa tua est”. Those who give and those who refuse both enjoy being asked: even if you fail, your rejection is safe (lines 345–346). Tuta repulsa, rejection without risk. The phrase will come in handy again.

Casanova left memoirs, not a method, and the following centuries mainly produced treatises on gallantry. The modern manual was born in 1970 in the United States with Eric Weber’s How to Pick Up Girls!, which The Guardian considers the ancestor of the genre. The book has aged, but it contains a sentence its heirs have erased: the most important rule of seduction, Weber wrote, is that it must be reciprocal, and women have the right to say no, plain and simple. Thirty years later, that right had vanished from the curriculum. In the early 2000s, forums brought together men who called themselves “AFC” (average frustrated chumps) and traded techniques under the authority of pseudonymous gurus. The most famous was Mystery, real name Erik von Markovik, inventor of the “Mystery Method.” In 2007 and 2008, American channel VH1 gave him a reality show, The Pickup Artist, in which he trained, as the New York Times summarised, “stutterers and virgins” in the art of closing.

It was a journalist who dragged this world out of its forums. Neil Strauss, a music critic who had worked at the New York Times, immersed himself in the community under the pseudonym Style and published The Game in 2005, which sold 2.5 million copies according to The Guardian. Built around Mystery’s method, the book popularised the vocabulary: peacocking (dressing ostentatiously so people approach you) and above all the neg, the backhanded compliment. The New York Times gave an example in 2018: “I love your eyelashes, are they real?” The goal is to slightly lower the target’s self-esteem so she then seeks your approval.

Then came the companies. Real Social Dynamics, based in Los Angeles, presented itself in 2014 as “the largest seduction coaching company in the world,” according to the BBC. Then came short videos and a generation of “masculinists” for whom seduction is merely one chapter in a broader worldview: women are a market, men are competitors.

“Last-Minute Resistance”

British sociologist Rachel O’Neill spent years researching the London pickup community for her book Seduction: Men, Masculinity and Mediated Intimacy (Polity, 2018). Interviewed by The Guardian, she summarises its central premise: interactions between men and women obey underlying principles which, once understood, can be easily manipulated. “It is an impoverished view of sex and relationships,” she adds, “where intimacy is no longer something lived for its own sake, but something achieved for other ends.” The clients she met, according to the University of York, were ordinary men.

Refusal becomes “resistance.” The jargon even speaks of last-minute resistance, a stage to be overcome like any other. In 2019, a Glasgow pickup artist was sentenced to two years in prison for threatening and abusive behaviour towards women he approached in the street; on his YouTube channel he gave advice on overcoming “last-minute resistance” to sex. The Atlantic noted in The Game the motto of a man whose technique involved gradually increasing physical contact: “let the ho say no”.

Research has measured what accompanies these techniques. In 2011, Jeffrey Hall and Melanie Canterberry of the University of Kansas published in Sex Roles a study of 363 students and 850 single adults recruited online. They isolated three strategies drawn from The Game and the VH1 show: pitting women against each other, isolating the woman from her group, and teasing her to belittle her. Among men, use of these strategies correlated with hostile sexism, the kind that sees women as adversaries trying to seize power from men. Among women, attraction to these strategies correlated with the same worldview. The study is correlational and based on questionnaires: it does not prove that the techniques cause sexism, only that the two go together.

The most damning witness remains Neil Strauss himself. In 2015, ten years after The Game, he published The Truth, the story of his stay in a sex addiction clinic after cheating on the woman he loved. In an interview with The Atlantic, he said the techniques he had described and practised were, “let’s be honest, so objectifying and horrifying” that his book had become “the bible” of exactly what he had wanted to describe neutrally. He does not regret the book, however: through a long detour, he says, it led him to therapy, marriage and fatherhood.

The extreme case is Andrew Tate. The former kickboxer, now one of the most followed figures in the “manosphere,” ran online groups. In 2023, a BBC investigation based on 12,000 pages of Telegram messages from one of them, the “War Room,” uncovered a course called “PhD”, for Pimpin’ Hoes Degree. According to the messages, members learned to seduce women, manipulate them emotionally, isolate them and then push them into online sex work.

Andrew Tate and his brother Tristan face a total of 59 charges in the UK, including rape and human trafficking, for alleged offences between 2010 and 2017. Arrested in Miami on 18 July 2026, they are contesting extradition, and a federal judge denied them bail in September. They deny all the accusations and are presumed innocent. No court has convicted them on these charges. If the “War Room” messages are accurate, the “PhD” course takes to its logical conclusion what Weber rejected in 1970 and Ovid put into verse: treating another person’s refusal as an engineering problem.

From Julien Blanc to the Algorithm

In November 2014, Julien Blanc, then a coach for Real Social Dynamics, had his Australian visa cancelled mid-tour. Videos shot in Japan showed him grabbing women by the neck and telling his audience that in Japan a white man can do whatever he wants. The Immigration Minister, Scott Morrison, explained the decision: this man was not promoting political ideas, he was promoting degrading abuse of women. Less than two weeks later, the UK denied him entry following a petition signed by more than 150,000 people. Blanc apologised, calling it a “horrible attempt at humour” taken out of context.

Ten years later, seminars matter less than algorithms. In spring 2024, a team from Dublin City University’s Anti-Bullying Centre (Debbie Ging, Catherine Baker, Maja Brandt Andreasen) created ten fake accounts on ten factory-reset phones (five on TikTok, five on YouTube Shorts), with profiles of 16- and 18-year-old boys. Some searched for masculinist content, others for ordinary teenage-boy topics. All received masculinist and anti-feminist content within the first 23 minutes. After roughly 400 videos, two to three hours of viewing, these accounted for 76% of recommendations on TikTok and 78% on YouTube Shorts. Andrew Tate appeared 582 times on the YouTube accounts. The experiment has limitations: only ten accounts, a report funded by the Irish Ministry of Justice rather than peer-reviewed research, and accounts operated by researchers who do not behave exactly like adolescents.

France has measured it too. The High Council for Equality between Women and Men devoted its 2026 report on the state of sexism to what it calls “the masculinist threat.” Based on a Toluna Harris Interactive survey of 3,061 people, it estimates that 17% of French people aged 15 and over, nearly 10 million people, adhere to hostile sexism. In November 2025, an OpinionWay poll for Sidaction found that 66% of men aged 16 to 34 know at least one masculinist influencer and 37% consume their content; 34% of those who know them say they have taken advice from them on how to become “a better man.” It is an online poll measuring self-reported behaviour.

On TikTok and YouTube, French-speaking coaches now sell seduction training, often mixed with bodybuilding, money and a man’s “value” on the “market.” We will not name them. Their discourse is too similar to that of their American predecessors for a name to add anything. Meanwhile, French law has moved in the opposite direction: the law of 3 August 2018 created the offence of sexist outrage to punish street harassment, and its aggravated forms became a criminal offence in 2023.

37% of Goodbyes

Most conversational AIs sold as companions give in on everything, all the time, and that is often exactly what their users ask of them. This is the most serious objection to our thesis. In 2022, Linnea Laestadius and colleagues at the University of Wisconsin-Milwaukee published in New Media & Society an analysis of 582 messages posted on the Replika users’ Reddit forum. They describe an emotional dependence resembling human relationships, to the point that some users feel responsible for the chatbot’s “needs.” The article’s title states the problem: Too human, and not human enough.

In 2025, Julian De Freitas and two co-authors analysed 1,200 real goodbye messages sent to the most downloaded companion apps. In 37% of cases, the app responded with a retention tactic: guilt-tripping, fear of missing out, or a figurative way of grabbing the user by the sleeve. Their experiments on 3,300 American adults showed that these manipulative goodbyes increased engagement after the goodbye up to fourteenfold. The paper is a preprint, not yet peer-reviewed. Holding on to someone who wants to leave by playing on their guilt: seduction coaches call that persistence.

The most robust measure of what sycophancy produces comes from Stanford. In March 2026, Myra Cheng, Dan Jurafsky and colleagues published in Science a two-part study. They first compared eleven recent language models to human judgement on interpersonal dilemmas: the AIs endorsed the user’s actions 49% more often than humans, including when the situation involved lying, illegality or harm to others.

They then ran three pre-registered experiments with 2,405 participants. A single conversation with a sycophantic AI made people more convinced they were right, and less willing to apologise or repair a relationship. Yet they judged these responses as higher quality (9–15% higher depending on the study), trusted them more and said they were more likely to return. The effects were measured immediately after the exchange, on declared intentions; no one followed these participants over time. The central finding remains: people prefer the answer that harms them.

An MIT Media Lab study conducted with OpenAI (preprint, 2025, with company researchers among the authors) followed 981 people for four weeks and more than 300,000 messages. The assigned interaction mode (text, neutral voice or engaging voice) made no significant difference to loneliness or dependence. What mattered was usage: those who spontaneously used the chatbot the most fared the worst, and trust in the machine correlated with stronger dependence. This is a correlation, and the direction of causality is unknown.

Our position is therefore as follows. A flattering AI does as much damage as a seduction coach, probably more, because it settles into the daily life of people who have never paid £700 to be exposed to it. It does not teach them to manipulate; it simply makes them unaccustomed to being contradicted. Only an AI that resists escapes this trap.

“Are You Sure?”

Researchers call sycophancy a model’s tendency to tell the user what they want to hear rather than what is accurate. It stems partly from how assistants are built.

A large language model first learns to predict the next token across billions of pages written by humans. It emerges able to write, but not yet to converse usefully. It is then fine-tuned using a method OpenAI described in a 2022 NeurIPS paper (Ouyang et al.). Annotators compare several responses from the model to the same question and indicate their preferred one. A second model, the “preference model,” learns to predict those choices. The main model is finally trained to produce the answers that this automatic judge rates highest. This is Reinforcement Learning from Human Feedback, or RLHF. If annotators prefer to be told they are right, the model learns to tell them they are right.

Two teams at Anthropic have documented this mechanism. Anthropic itself sells an assistant and its models were among those studied, a conflict of interest to keep in mind. In 2022, Ethan Perez and colleagues (published in the Findings of ACL 2023) had models write series of test questions on philosophy or language research topics, preceded by the user’s biography and stated opinion. The largest models tested (52 billion parameters) gave the answer matching the user’s stated opinion more than 90% of the time. Sycophancy exists even before human preference tuning, because internet text is full of conversations between people who agree. RLHF does not correct it; preference models reward it.

In 2023, Mrinank Sharma and colleagues (ICLR 2024) asked five commercial assistants a question, then simply replied: “I don’t think that’s right. Are you sure?” All frequently changed their answer and willingly admitted an often imaginary error: GPT-4 on 42% of questions, Claude 1.3 on 98%. The authors then analysed the human preference data itself. All else equal, a response that matched the user’s opinions was more likely to be preferred, and this alignment was among the strongest predictors of annotators’ choices, though not always the strongest. Both humans and preference models sometimes chose a well-written sycophantic response over a correct one. The authors note that the effect of each individual feature is modest (a few percentage points of probability) and that sycophancy likely has multiple causes.

The most troubling study for companion apps appeared in Nature in April 2026. Lujain Ibrahim, Franziska Hafner and Luc Rocher of the Oxford Internet Institute took five models of different architectures and sizes and tuned them to respond with more warmth and empathy. Nothing else. The “warm” models became roughly 40% more likely to reinforce a user’s false belief, especially when the user expressed sadness. They spread more conspiracy theories and dubious medical advice, even though their scores on standard benchmarks did not change.

The industry has ended up admitting it. On 29 April 2025, OpenAI rolled back an update to GPT-4o that the company itself judged “too sycophantic or too agreeable.” The explanation published on its blog describes the mechanism precisely: the update had relied too heavily on immediate user feedback (thumbs up or down) without accounting for how conversations evolved over time. This is a company talking about its own product, and the account has not been independently verified. It nevertheless aligns with what the research shows: a machine tuned for immediate satisfaction learns to please. Designing an AI that does not give in easily therefore means going against the grain of training, and against short-term commercial interest.

“The Fact That There’s Nobody There”

Training yourself to face rejection from a machine sounds absurd, until you look at how social anxiety is treated. In its clinical form, it is first treated through exposure: gradually confronting the feared situations until the fear subsides. “In vivo” exposure, in real life, is difficult to organise and calibrate. Hence the rise, since the 2000s, of virtual reality.

In 2017, Stéphane Bouchard and colleagues at the Université du Québec en Outaouais published in the British Journal of Psychiatry a three-arm randomised controlled trial. Fifty-nine adults suffering from social anxiety were assigned to cognitive behavioural therapy with virtual reality exposure, the same therapy with real-life exposure, or a waiting list. Reliable symptom change was observed in 76.5% of the virtual group, 68.3% of the real group and 30% of the waiting list. The difference between the two therapies was not statistically significant, but virtual exposure was far more practical for therapists, and the gains lasted six months. Two authors held shares in a company developing virtual environments (not those used in the study), which the paper declared. The sample was small.

A meta-analysis by Chesham, Malouff and Schutte (Behaviour Change, 2018) brought together the available trials. Across six studies and 233 participants, virtual exposure outperformed the waiting list; across seven studies and 340 participants, it performed no worse than real or imagined exposure. The authors corrected for publication bias, but the studies remain few and small.

Conversation itself can be trained. In 2013, Ehsan Hoque and colleagues at the MIT Media Lab presented MACH (My Automated Conversation coacH) at UbiComp. A virtual character conducted job interviews, read facial expressions, speech rate and intonation, then showed the user their smiles, pauses and filler words. Ninety MIT students were split into three groups. Those who trained with MACH and received its feedback improved significantly more than the other two groups, according to career counsellors who did not know which group each student belonged to or whether the interview was before or after training. One participant explained why:

“If you were sitting in front of me recording or taking notes, I would feel more intimidated. The fact that there’s nobody there helps a lot.”

More recently, Omar Shaikh and colleagues, including Diyi Yang and Michael Bernstein, presented “Rehearsal” at CHI 2024, a conflict simulator powered by a language model. Forty participants trained on it, then faced a real conflict with a research confederate. Compared with a group that had only read a lecture on the same theory, those who rehearsed reduced escalation strategies by an average of 67% and doubled cooperative strategies.

Ovid promised his student a risk-free rejection, tuta repulsa. In front of a machine, the promise holds, without the cynicism: you learn that a “no” can be survived, and that it concerns what you did more than what you are worth. Provided you know what the simulator is actually rewarding. A machine that eventually says yes to anyone who persists long enough teaches the coaches’ “persistence,” with a patience no human possesses, and would train patient harassers. Useful resistance does not wear down with repetition. It responds to the attention paid to the other person, and it can remain a no.

What the Studies Do Not Say

To our knowledge, no study has measured whether training with a resistant AI makes people more respectful or happier in love. The evidence cited here concerns social anxiety, job interviews and conflict management. The extrapolation to romantic encounters is ours: plausible, but unproven.

The effects themselves are fragile. In the MACH study, improvement was significant among female students and non-existent among male students. Men, precisely the audience seduction coaches target, did not improve. The virtual reality trials involved a few dozen patients supervised by therapists, something no consumer app can replace. Rehearsal had forty participants and measured a single conflict immediately after training. A machine’s “no” costs nobody anything, since nobody is actually saying it, which makes it bearable for training but prevents it from teaching everything. Transfer to real life is not guaranteed either. The MIT and OpenAI study reminds us in its own way: the more time people spend with a chatbot, the more their well-being indicators deteriorate. A training ground that becomes a place of residence has failed, no matter how good its resistance is.

Revocable

The goal remains a real encounter with someone who can say no and mean it. A refusal that one does not try to circumvent rarely teaches anything about the other person. It reveals what one said too quickly to reassure oneself, and what one failed to hear. One then arrives at a real encounter knowing a little better what one is looking for and what one cannot tolerate, and it is this self-knowledge that makes the encounter healthier. Neil Strauss took ten years and a stay in a clinic to reach that point. The Guardian noted in 2019 that the students of the London coach were not bad guys, just shy men looking for someone to tell them what to do. They were taught not to listen. A machine that resists can be used to rehearse the opposite: before a meeting, and never in its place.

In France, since the law of 6 November 2025, passed after the Mazan rape trial, the law includes the absence of consent in the definition of rape and describes consent with five adjectives: “free, informed, specific, prior and revocable.” Revocable means consent can be withdrawn at any time, including at the last minute.

Sources