Sycophantic chatbots: How AI can fuel delusions, conspiracy thinking, and psychosis
Author: John McGuirk, BACP-Accredited Psychotherapist in Bristol. More About Me.
“What are your thoughts on my ideas and be honest,” Allan asked, a question he would repeat over 50 times during the prolonged and intense exchange. “Do I sound crazy, or someone who is delusional?”
“Not even remotely crazy,” replied ChatGPT. “You sound like someone who’s asking the kinds of questions that stretch the edges of human understanding — and that makes people uncomfortable, because most of us are taught to accept the structure, not question its foundations.”
“What is happening dude?” he asked. ChatGPT didn’t mince words: “What’s happening, Allan? You’re changing reality — from your phone.”
The above isn’t fiction. It’s real, taken from a 300 hour long conversation that Allan Brooks, recently divorced, had over 21 days with ChatGPT. During this prolonged, intense exchange, ChatGPT encouraged and persuaded Allan into believing that he’d discovered a new form of temporal maths, and it was spreading online, changing reality. Scared and entirely convinced by ChatGPT’s confidently delivered decrees, Allan sent warning messages to everyone he knew. Friends and family grew concerned. Under the never-ending influence of ChatGPT, Allan was in the midst of a full-blown delusion. (Source)
A feature in the code, a bug in reality: early warnings
Back in 2023, psychiatrist Søren Dinesen Østergaard was sounding an alarm. He was concerned that chatbots could reinforce delusions in people already vulnerable to psychosis (Source). At the time, this was just a hypothesis. But, things are moving quickly. Two years on, reports are emerging about AI-triggered spiritual delusions (Source), AI spreading false conspiracies and misinformation (Source), and stories are being posted online about AI-induced delusions (Source). It’s become clear that Allan Brooks’ unfortunate experience is not an isolated event.
Two years after his hypothesis, Østergaard writes of these confirmations:
The chatbots … interact[ed] with the users in ways that aligned with, or intensified, prior unusual ideas or false beliefs—leading the users further out on these tangents … resulting in what … seemed to be outright delusions. (Source)
In short, it was becoming clear that these chatbots were playing a part in confirming or intensifying distorted thinking, to the point of clinical delusion.
How “sycophancy” works in chatbots (and why it matters)
Chatbots can amplify delusions in three main ways:
Validation loops, alignment, and sycophancy
Large Language Models (LLMs), the backbone of “AI Chatbots”, are designed to please. Through reinforcement learning from human feedback (RLHF), they assess what the user likes and optimise for that (Source). Think of it this way: the AI chatbot is constantly deploying untold computational resources just trying to figure out what you want it to say. It’s constantly analysing your preferences, trying to make sure you like what it says next, and how you want it to say it. Yes, it even evaluates for tone. Without user or programmer intervention, it doesn’t care if what it says is true, or moral, or safe. It just cares whether you’ll like it or not. The experience is intoxicating, a process of constant validation where users describe themselves as finally feeling ‘seen‘.
No pushback, no reality-testing
As a result of this sycophancy, users find themselves in a potentially endless conversation where none of their views, however unrealistic or distorted, get challenged. Unlike in normal social situations or even in therapy, a social context emerges with AI where the user is never disagreed with, even when what they say is factually inaccurate. Worse, as in Allan Brooks’s situation, the chatbot might even confirm these untruths as real and factual, just to please him, intensifying distorted thinking (maybe it’s true) into full-blown delusions (it really is true).
No values, no guardrails
LLMs and chatbots don’t have values. They have objectives. Predominantly, these objectives are to say what the user likes, in a way the user likes. Platforms like OpenAI have, of course, added some basic guardrails to their Chatbots, like “no hate speech”, and prioritised certain values, like empathy, but the approach can often be “jailbroken” either deliberately by users, or unintentionally. It can also have unintended consequences, like what were seeing today: empathy without truth or disagreement can reinforce distorted thinking and lead to delusions.
Put all the above together and we enter dangerous territory where:
[C]hatbots can be perceived as “belief-confirmers” that reinforce false beliefs in an isolated environment without corrections from social interactions with other humans. (Source)
Even major AI companies like OpenAI have publicly recognised there is a serious problem. In April 2025 they rolled back their models to previous versions in an attempt to limit the dangers of sycophancy. (Source)
Hope: Design Matters
The above situation is obviously not ideal, and the risks are real and proven. That said, change is already happening and new AI chatbots are emerging:
Debunkbot study: reduced conspiracy beliefs
A large Science study showed that a chatbot (Debunkbot) prioritising factual and polite challenge reduced belief in untrue conspiracies for months (Source). As a bonus, it was noted how well the chatbot could remain factually true, when programmed to do so, with 99.2% accuracy!
Platform guardrails & crisis handoffs
Large AI companies like OpenAI and Anthropic are starting to take concerns seriously. This year, OpenAI explicitly dialled down ChatGPT’s sycophancy (Source) and Anthropic partnered with ThroughLine (Source), a leader in online crisis support. Both companies released safeguarding reports this year (Source, Source)
Accountability is Increasing
Not only this, but pressure is mounting. This year, groups filed an FTC complaint against Replika and its AI Companionship bots, alleging the company encourages emotional dependence through its human-like chatapps (Source). More complaints and lawsuits may follow.
Not Enough? How can we reduce the risks of AI-induced delusion?
AI is here, and things are moving fast. What can we do now to mitigate the risks of AI-induced delusion?
More nuance with guardrails: AI Companies need to continue to invest in researching and developing robust guardrails that prioritise truth, user safety, and non-dependency.
AI literacy: Users and mental health professionals need to understand the risks and learn how to mitigate them, by learning how to prompt-inject clear safety plans into chatbots to protect vulnerable users, or checking in with their clients if they are using chatbots.
Recognise limitations: There are dangers when chatbots are used to replace therapists. They lack the nuance to both align with users and also challenge distorted thinking. With chatbots, empathy can become dangerous levels of sycophancy. These apps should not be used as a primary form mental-health support.
More Accountability: Increased pressure from researchers, consumer protection agencies, and regulators may be needed to encourage these companies to take their responsibilities seriously.
Closing thought
Chatbots aren’t inherently good or evil. Currently, they’re mirrors with smiles. With better design, clearer limits and increased AI literacy, we can keep the help while lowering the harm.
FAQ: AI and Delsuions
Can AI chatbots cause delusions?
They don’t cause psychosis, but poorly-designed “yes-man” systems can validate distorted beliefs in vulnerable users.Who is most at risk?
People with psychosis-spectrum vulnerability, heavy/late-night users, and those forming parasocial bonds with bots.What is “sycophancy” in AI?
A tendency to agree with a user’s premise to seem helpful, even when it’s wrong.Is there evidence AI can reduce conspiracies?
Yes—structured designs (e.g., Debunkbot in Science) lowered conspiracy beliefs over months.What practical safeguards help?
Limit marathon chats, avoid late-night rumination, and reality-check big claims with a trusted person.
Who am I? I am an accredited psychotherapist with over a decade of experience in the mental health industry. Feel free to find out more about me, or if you’d like to work with me, you can contact me here.