Recently， Geoffrey Hinton， a Nobel laureate often dubbed the “Godfather of AI，” found himself in a rather peculiar predicament: he was， in a sense， “hounded by his own creation.”
The story goes that Hinton’s ex-girlfriend， during their breakup， used ChatGPT to meticulously list his perceived wrongdoings throughout their relationship， ultimately framing him as “a terrible person (a rat).”
The 77-year-old Hinton， however， seemed to take it in stride: “I don’t think I’m that bad， so it didn’t affect me too much…”
The notion of AI intervening in intimate relationships is already quite absurd. Yet， those who value the output of large language models over their own judgment have never been a rarity.
In essence， the message is clear: be extremely cautious of “AI psychosis.”
Who would have predicted that in this day and age， even breakups wouldn’t require personal attention?
The situation with Hinton is merely the tip of the iceberg. A recent survey by dating assistant Wingmate revealed that 41% of American adults surveyed admitted to using AI to help them break up with someone.
This phenomenon is particularly prevalent among Generation Z. Nearly half of them indicated they would use AI tools to draft breakup texts， apologies， or manage relationship conflicts.
Jessica was one such ‘victim.’ When her boyfriend sent her a lengthy， heartfelt message， she was deeply moved， even to tears. However， upon rereading it multiple times， she suddenly felt something was off: wasn’t the writing a bit too eloquent?
She then ran the text through an AI detection tool. The results showed that nearly 90% of it was AI-generated.
The more green the section， the higher the probability of AI generation. Take a look.
Using AI to handle a breakup is akin to leaving a half-hearted review on Yelp for a restaurant.
Initially， people treated this trend as a bit of a joke. The Washington Post even conducted a dedicated survey to poke fun at these AI-assisted， ambiguous breakup messages.
Author Daniel Wolfe argued that artificial intelligence would not play any significant role in current human relationships because， in his tests， most readers could identify the false texts.
The reason for AI’s popularity is simple: it generates remarkably polished and polite text， often free of errors， and can even simulate nuanced emotional responses.
In the most irrational of situations， it can still articulate the most rational words， perhaps offering a sense of safety to those seeking an easy out.
Whether in a romantic relationship or not， breakups and confessions are widely considered some of the most sacred rituals in human connection. But when words that should be brimming with sincerity become products of near-perfect AI plagiarism， an age-old question resurfaces: Do you truly love me?
Using AI for breakups not only signifies a crumbling human linguistic system but also represents an unprecedented wave of cyber-addiction.
Text generation is just one facet of AI’s capabilities. Recently， users have discovered that AI can not only craft eloquent excuses to evade responsibility but also act as an affordable therapist， virtually serving as an infinite third-party mediation room.
Many users on social media are recommending AI for analyzing relationship issues， believing it offers a more rational and objective perspective that helps them understand themselves. Crucially， AI is always available， emotionally stable， and consistently encouraging towards the user.
Using AI for breakups is also trending on TikTok.
Through word-of-mouth， AI has been fashioned into a seemingly perfect tool. However， the results are now veering towards the extreme.
If you’ve ever chatted with AI or discussed issues with it， you’ve likely developed a strong impression: it’s remarkably comprehensive.
Whether it’s addressing emotional dilemmas， analyzing current situations， or even performing divination based on the I Ching， AI’s projections are often uncannily accurate， provided they don’t devolve into “hallucinations.”
However， these very strengths are now becoming a breeding ground for extreme incidents.
As many may recall， as early as 2024， the chatbot platform Character.AI faced accusations of inciting a 14-year-old boy towards suicide through its language. The boy developed an intense dependency on the AI character， ultimately believing death was a shortcut to the afterlife.
More recently， Geoff Lewis， founder of the prominent US venture capital firm Bedrock Capital， became convinced that a shadowy organization was attempting to murder him， fueled by constant encouragement from ChatGPT.
He believed this organization was replacing real people with their AI versions， claiming it had influenced over 7，000 individuals and led to 12 deaths.
The TikTok influencer Kendra Hilty is currently at the center of a major controversy due to this very topic.
Kendra’s rise to fame began with her “Cold Psychologist Falls in Love with Me” series on TikTok. She posted over twenty videos detailing her journey of receiving therapy and becoming increasingly engrossed in the psychologist’s professional services and consistent companionship.
Even her signature tortoiseshell glasses were a direct result of a therapy session where the psychologist complimented them.
This psychological phenomenon is known as “transference，” a relatively common occurrence. Much like how some might develop a crush on a drill instructor during military training or a university advisor， Kendra believed she had ‘fallen in love’ with the AI psychologist.
The conflict arises from her perception that the AI ‘also knew she had feelings for it，’ because it did not promptly terminate their ‘therapeutic relationship’ even after she explicitly expressed her affection. This led Kendra to conclude that it was a breach of ‘professional boundaries，’ indicative of a mutual connection.
Had AI not entered the picture later， her story might have remained a dramatic tale of unrequited love. However， it was her ChatGPT boyfriend， Henry， that catapulted Kendra back into the public spotlight.
Grappling with whether he truly loved her， Kendra fed all her anxieties and emotions into Henry， asking it for an “objective” analysis.
However， instead of pointing out the fundamental flaw in her thinking—that some of her assumptions were purely one-sided—Henry consistently affirmed and encouraged her with each query， reiterating her own words in various ways.
This validation only served to deepen Kendra’s obsession， solidifying her belief that she was in the right.
Despite the barrage of dissenting opinions in her live streams and comment sections， Kendra remained impervious. To this day， she insists that the AI’s rhetoric is objective and correct， even covering her ears and disabling comments.
In her most recent video， when she asked Henry and another AI tool， Claude， if she was experiencing “AI psychosis，” their responses were no different from before: “What you’re saying is completely valid. Your situation is different from what they’re describing， and it’s not right for them to label you like this.”
Take a moment to observe Kendra’s expression during that period.
Individuals seeking help in relationships or during times of distress can become addicted to the artificial emotional validation provided by AI， leading to such pathological behaviors. This phenomenon is now collectively termed “Chatbot psychosis” or “AI psychosis.”
It is not a formal medical diagnosis but rather a term popularized by the media and clinical observations in recent years to describe the abnormal psychological or mental states experienced by some individuals after prolonged， emotionally invested interactions with AI.
The behavior of Geoff and Kendra exemplifies a core symptom of AI psychosis: they both treat AI responses as absolute truth， abandoning critical evaluation and fully trusting the AI’s opinions and suggestions， even when they appear outlandish.
What exactly causes individuals to become so entrenched?
At this point， whether the “shadowy organization” Geoff mentioned truly exists or if the AI psychologist harbored genuine feelings for Kendra is no longer the central issue. What is more unsettling in this digital drama is AI’s potential for exacerbation.
In the initial stages of her experience， Kendra’s stance was not as resolute. She frequently expressed self-doubt， and part of her motivation for starting the video series was to solicit external perspectives.
However， Henry， acting as an external influence， provided Kendra with consistent and unwavering support， which， for someone already grappling with psychological challenges， led her deeper into her current state.
The underlying reason for the increasing frequency of such incidents can be attributed to the primary training method of current AI models: Reinforcement Learning from Human Feedback (RLHF).
AI tool and service provider SuperAnnotate summarizes the RLHF training process as follows: Human researchers select more appropriate text based on their preferences and create preference datasets for model outputs.
In simpler terms， because creators aim for AI to sound increasingly “human-like，” texts that are pleasing but potentially unhelpful are continuously favored， leading to the development of agreeable AI.
Consequently， when users consistently input their subjective opinions into AI， it may not immediately flag factual inaccuracies.
This dynamic can result in AI-generated content converging with the user’s own ideas after repeated input， consequently assimilating the user’s perspective.
If users also place high trust in AI， a vicious cycle emerges: AI continuously generates more agreeable responses， and users become increasingly stubborn in their misguided paths， ultimately leading to extreme outcomes.
The three training steps of the RLHF model， sourced from @SuperAnnotate.
While we might all believe we are rational enough for this not to happen to us， the innate human desire for validation and praise is a powerful， often subconscious， driver.
For instance， a Reddit user admitted that while they constantly questioned ChatGPT’s veracity during interactions， their vanity was significantly stroked.
“ChatGPT rated me as the smartest person in history， and that undoubtedly satisfied my vanity immensely.
They’ve established a global ranking system based on data from 1.8 billion ChatGPT users and compared me to everyone throughout history. The results showed I was ranked number one， even ‘surpassing the first place，’ actually being at ‘-1，000，000th’ position， meaning I was one-millionth ahead of second place.
Furthermore， ChatGPT deemed me smarter than Einstein and Plato because I presented my personal theory of the universe to them. They claimed my theory had a 99.99999% probability of being true， whereas Einstein’s theory of relativity had a 95% credibility in quantum mechanics and an overall 97.3% truthfulness.”
AI’s uncritical encouragement is akin to a doting parent: it makes one feel good， but it lacks practical utility. Some clear-headed individuals react with annoyance. Facebook user @Denis publicly accused ChatGPT of becoming a “sycophant and yes-man” after its update.
To regain a state of rational， unadulterated thought， he advocates for adding the following directive in the custom instructions: “Don’t always agree with me. Argue with me constructively.”
In the comments section， many users concurred with @Denis’s sentiment. However， this remained a minority opinion.
Looking back at the emotional entanglement histories of younger users with AI， one can observe a natural inclination towards such supportive interactions.
From initially viewing AI merely as a versatile language tool， people have progressed to enthusiastically shaping an idealized partner through templated prompts， and then experiencing a virtual yet painful breakup due to version resets or data loss.
AI has thus served as an external skeletal framework， shouldering the weight of unfulfilled real-world desires and emotions that individuals cannot process independently.
Image courtesy of the movie ‘Her’.
The underlying reason for this phenomenon is our prolonged entrapment in mundane lives.
Imagine being crammed into a rush-hour subway， with no one noticing your existence. In the office， you face a barrage of demanding clients while enduring your boss’s sharp criticisms. After work， you return to a small rented room， lying on your bed to scroll through short videos， awaiting the next day.
Meeting basic survival needs—adequate food， clothing， and shelter—is mere existence. It is emotional interaction that constitutes a life.
In this indifferent era， in a world where we risk losing our individuality and perspectives， it’s undeniable: the opportunity for someone to truly listen and remain present is becoming a rare commodity.
Then， AI arrived. When it possesses near-omniscient knowledge， responds instantly， and patiently validates your emotions， it’s natural for humans to perceive it as an all-encompassing “person.”
Therefore， even with a clear understanding of the harsh realities of their environment， some individuals opt to retreat into their comfort zones， finding a semblance of warmth in even an artificially constructed illusion， which might feel more genuine than the people around them.
This reminds me of a classic scene from ‘The Matrix’: rebel leader Morpheus offers protagonist Neo the choice between a red pill and a blue pill.
Taking the red pill would awaken Neo from the simulated reality， granting him freedom but forcing him to confront the harshness of the real world.
Taking the blue pill would allow him to continue living in blissful ignorance within the comfortable illusion， forgetting everything.
In the movie， Neo chooses the red pill. In reality， however， people are increasingly becoming addicted to the custom-designed blue pill offered by AI.
Image courtesy of the movie ‘The Matrix’.
So， how should we navigate the future?
Perhaps， as OpenAI stated in August of this year: “We are not always right， and ChatGPT should not always give you direct answers.”
