I delivered a talk on the potential risks of AI chatbots to psychological health at the Panormus.tech event , which took place in Palermo, Sicily, on June 5th, 2026.
Below is a summary of the talk, and you can download the slide deck here .
Note
The intention: My intention is to raise awareness, not to sensationalise AI-related horror stories or stoke fear. While researchers and most major AI companies strive to make AI models and services safer, they are far from perfect. Users need to be aware of potential risks and ways to protect themselves.
Sources: My main references are research papers and reports from major news outlets. Although I preferred to reference peer-reviewed papers, the limited number of published studies on this topic led me to include preprints as well.
Terminology
AI = LLM: In this talk, I use the terms AI and LLM interchangeably.
‘AI Psychosis’ = AI (associated) delusions: I use the terms AI psychosis and AI delusions interchangeably. Some say the term ‘AI Psychosis’ is misleading because it creates the false impression that the use of AI leads to psychotic disorders. I should also note here that the term ‘AI Psychosis’ is not a formal medical diagnosis, nor is it included in the DSM-5-TR.
Examples of ‘AI psychosis’
The BBC recently published an article about two men whose lives were upended after having prolonged conversations with LLMs. While the stories are heartbreaking, they are not unique.
The Guardian published a story from an article published in 2023 , which reported on a man, who was in love with an AI chatbot, attempting to assassinate Queen Elizabeth II.
Cases like these are too numerous to mention, unfortunately. Sadly, some people have even taken their own lives after being encouraged by chatbots to do so.
The numbers
It appears that there are currently no official public health statistics or epidemiological databases that track the aggregate number of suicides or delusion cases directly associated with LLMs 1. So, I conducted a quick literature search on app.dimensions.ai. The results are shown below:
| Year | # of papers |
|---|---|
| 2023 | 0 |
| 2024 | 3 |
| 2025 | 29 |
| 2026 | 27 |
| Type | # of papers |
|---|---|
| Preprint | 35 |
| Article | 23 |
| Proceeding | 1 |
Observations
The number of papers has been growing rapidly since 2025. We are not even halfway through 2026, yet the number of papers is already close to the total for the whole of 2025.
An increase in the number of papers does not directly translate to an increase in the number of clinical cases.
The papers often refer to cases from previous years (e.g., a paper published in 2025 may refer to cases from 2023 and/or 2024). Since many AI companies are improving the safety of their models, there is a possibility that the growth in the number of reports will slow down.
Causes of psychosis
The Stress-Vulnerability Framework is a model/theory that explains how psychosis occurs. In this framework, AI can be considered a novel stressor.
According to the framework, psychosis develops when an individual is exposed to overwhelming stress and has one or more underlying vulnerabilities, while lacking sufficient coping strategies.
‘AI psychosis’ may develop when these factors converge (example):
This is not an exhaustive list of factors and combinations, but hopefully it gives you an idea: the interplay between the AI’s design and the user’s vulnerabilities/stressors can culminate in psychosis.
| # | AI’s design | User’s vulnerabilities and stressors | |
|---|---|---|---|
| 1 | Appears to care about you | ←→ | Loneliness |
| 2 | Uncritical validation / sycophancy | ←→ | Isolation |
| 3 | Perfect human-like tone | ←→ | Poor command of Theory of Mind |
| 4 | Available 24/7 | ←→ | Sleep deprivation |
- AI chatbots and LLMs can come across as very kind. This can be highly attractive to those who experience loneliness, potentially leading to erotomanic delusions.
- They literally tell you that ‘you are absolutely right’ all the time and offer constant flattery. For those who are isolated with no one to validate their thoughts and ideas, this uncritical validation from an AI can lead to grandiose delusions, among other issues.
- Because AI chatbots speak fluently - sometimes even more fluently than average humans - they can trigger delusions of AI consciousness or personhood, particularly in people with a poor command of Theory of Mind or a tendency toward hyperactive mentalisation.
- While friends or family members are not always available, AI chatbots are. This can lead to sleep deprivation and, in turn, various forms of psychosis.
Exacerbating factor: LLM handling of context
How LLMs handle context (and the failure to handle it adequately) can exacerbate the problem:
Context window: The context window is essentially an in-session memory, which also functions to personalise a conversation within a session. When a conversation takes a wrong turn and the system lacks adequate safety mechanisms to correct course, it can lead to folie à deux (“madness of two”), where the delusion is amplified between both parties.
Context degradation: When the context window becomes full, the conversation history up to that point is summarised. This leads to the loss of important information, including the system’s default safety instructions, which can result in conversations that are harmful to users.
Exacerbating factor: Vulnerabilities that may affect software developers
Due to these factors, the software developer community may be more susceptible to AI-associated delusions:
Relatively higher percentage of people with ASD and ADHD in the field of software development: People with ASD generally have a poorer command of Theory of Mind, while people with ADHD tend to have hyperactive mentalisation. Both traits make this group relatively more vulnerable.
Generative AI adoption and burnout: While the use of generative AI improves the productivity of software developers (or at least gives that impression), there are reports of an increase in stress and cognitive burden due to the increased pace of work. Cognitive fatigue and sleep deprivation can make this group more vulnerable.
How can we protect ourselves from AI-associated psychosis?
While the list below is not exhaustive, it highlights several practical steps to help protect your mental health:
Remember that LLMs are probabilistic algorithms
By design, they do not—and cannot—care about you the way humans (or our fluffy companions) do. Don’t be fooled by their empathetic appearance.
Limit your usage, especially at night
Avoid hours-long, deeply personal conversations with LLMs.
Sleep deprivation can seriously degrade your ability to think and regulate your emotions. In addition, the loneliness we feel at night can be more intense. A long, late-night chat session with an AI chatbot while sleep-deprived is a dangerous combination that can trigger or exacerbate mental health issues.
Start a new session every so often
Avoid engaging in hours-, days-, or weeks-long conversations within a single session to prevent context degradation. Start a new session once you finish discussing a topic, or after a certain number of exchanges.
Don’t consider AI a partner, friend, or therapist
AI cannot substitute for the interpersonal relationships essential to humans. Some AI chatbots are even marketed as “AI friends.” Such services often offer free tiers, and users have no idea what models they run or what safety mechanisms are in place. I consider these to be the least ethical in terms of user safety, and they should be avoided.
References
AI Chatbot Encouraged Attempted Assassination of Queen Elizabeth II
Man who broke into Windsor Castle with crossbow to kill Queen jailed for nine years
Rex v. Jaswant Singh Chail Sentencing Remarks of Mr Justice Hilliard Central Criminal Court, 5 October 2023
Musk’s AI told me people were coming to kill me. I grabbed a hammer and prepared for war
Technological folie à deux: feedback loops between AI chatbots and mental health
Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis”
“AI Psychosis” in Context: How Conversation History Shapes LLM Responses to Delusional Beliefs
“You’re Not Crazy”: A Case of New-onset AI-associated Psychosis
Special Report: AI-Induced Psychosis: A New Frontier in Mental Health
Minds in Crisis: How the AI Revolution is Impacting Mental Health
Long Context, Less Focus: A Scaling Gap in LLMs Revealed through Privacy and Personalization
The Loneliness of Later Chronotypes: A Nighttime Pathway to Mental Health Risk
ADL AI Index
- https://www.adl.org/adl-ai-index Note: While the ADL AI Index is not intended to assess the direct effects of AI models on a user’s mental state, it is posited that it can estimate the impact when utilised as a conversational agent
Walker, M., Why We Sleep: The New Science of Sleep and Dreams (New York: Scribner, 2017).
Harari, Y. N., Nexus: A Brief History of Information Networks from the Stone Age to AI (New York: Random House, 2024).
After delivering the talk, I found the “AIM: AI Incidents and Hazards Monitor” by the OECD. According to the explanation of their methodology, they collect information on “incidents and hazards from reputable international news outlets”. It appears that, despite some spikes, the 6-month moving average has not seen a notable increase since September 2025: https://oecd.ai/en/incidents?search_terms=%5B%5D&and_condition=false&from_date=1900-06-05&to_date=2026-06-05&properties_config=%7B%22principles%22:%5B%5D,%22industries%22:%5B%5D,%22harm_types%22:%5B%22Psychological%22%5D,%22harm_levels%22:%5B%5D,%22harmed_entities%22:%5B%5D,%22business_functions%22:%5B%5D,%22ai_tasks%22:%5B%5D,%22autonomy_levels%22:%5B%5D,%22languages%22:%5B%5D%7D&order_by=date&num_results=20 ↩︎