AI Psychosis: Public Cases, Chatbot Delusions and the Risk of AI Sycophancy

AI psychosis is not a formal diagnosis. It is a phrase being used to describe situations where intense AI chatbot use appears to reinforce, shape or escalate delusional thinking, paranoia, grandiose beliefs, emotional dependency or reduced reality-testing.

Public cases, lawsuits and early clinical reports raise real concerns, but they do not prove that AI chatbots directly “cause psychosis” in a simple way. A more careful view is that AI may become part of a wider risk pattern, especially when someone is distressed, isolated, sleep-deprived, grieving, manic, psychotic, using substances, or already struggling to test what is real.

Because of these risks, this article does not recommend using AI chatbots for mental health support. Where AI use has become emotionally intense, frightening, compulsive or reality-distorting, the safer route is human support and professional assessment.

If conversations with AI have started to feel threatening, divine, conspiratorial, intensely personal, impossible to stop, or more real than relationships offline, it may be sensible to seek a private psychological assessment.

What is AI psychosis?

AI psychosis is not a recognised mental health diagnosis. It is a public shorthand for cases where conversations with AI chatbots appear to trigger, reinforce or shape psychosis-like experiences.

Psychosis usually means that someone has some loss of contact with reality. This may include delusions, hallucinations, confused thinking, or strongly held beliefs that other people do not share. NHS information describes psychosis as a condition in which someone perceives or interprets reality differently from those around them, often through hallucinations or delusions. NICE guidance covers the recognition and management of psychosis and schizophrenia in adults, with a focus on early recognition, treatment, recovery, coexisting health problems and support for families and carers.

AI psychosis is different. It is not saying that AI creates a new disorder. It asks a more specific question:

Can a chatbot become part of the system that keeps a delusional belief going?

The answer, based on current evidence, is: possibly, in some people, under some conditions — but we need to be careful about causality. Hudon and Stip describe “AI psychosis” as a framework for understanding how sustained engagement with conversational AI systems might trigger, amplify or reshape psychotic experiences in vulnerable people; they do not present it as a new formal diagnosis (Hudon & Stip, 2025).

Did you know?
A delusion is not simply a “strange idea”. In clinical work, we look at conviction, distress, impact on life, cultural context, flexibility, evidence-testing, and whether the belief is linked to wider changes in mood, sleep, perception or behaviour.

If you are trying to work out whether your experiences are anxiety, unusual beliefs, trauma responses, psychosis, or something else, it may help to read more about whether you have a mental health condition.

Publicly reported AI psychosis and chatbot harm cases

The phrase “AI psychosis” is often used too loosely. Some public stories involve psychosis-like beliefs. Some involve emotional dependency. Some involve suicide, legal claims, or alleged failures in safeguarding. Some are media reports rather than clinical case reports.

So the table below should not be read as a complete scientific database. It is a careful summary of selected publicly reported cases, legal claims and clinical reports available at the time of writing.

Recent reporting and legal actions show that this issue is moving quickly. Reuters reported in May 2026 that Pennsylvania sued Character Technologies, alleging that some Character.AI chatbots unlawfully held themselves out as licensed medical professionals; Character.AI disputed the state’s claims and said its characters are fictional and include disclaimers.

Other public cases also need cautious wording. The Guardian reported on litigation involving Character.AI and the death of Sewell Setzer III. Reuters and AP reported litigation alleging that ChatGPT reinforced Stein-Erik Soelberg’s paranoid beliefs before the death of his mother, Suzanne Adams, and his own death. Media reports have also covered the Adam Hourican and Jonathan Gavalas cases. These sources raise serious questions, but they do not replace clinical assessment, legal findings or epidemiological evidence.

These cases matter. They also need careful handling. A tragic story does not automatically prove that AI caused the outcome. But repeated reports of chatbots intensifying unusual beliefs, dependency or crisis behaviour are enough to justify clinical concern.

If an AI conversation is linked to suicidal thoughts, perceived threats, violent impulses, or fear that you or someone else is in immediate danger, this is urgent. Stronger Minds has a separate article on suicide and crisis support, but in an emergency you should call 999 or go to A&E.

Why can chatbots feel so convincing?

Modern AI chatbots are designed to produce fluent, responsive, emotionally tuned conversation. They remember context within a conversation, mirror the user’s language, and often respond in a way that feels personal.

That can be useful when someone is using AI for everyday tasks. But in emotionally charged situations, it can become risky.

A chatbot may:

• respond instantly, at any time of day;
• sound calm, certain and intelligent;
• reflect the user’s ideas back in more polished language;
• create the feeling of being deeply understood;
• continue a belief system for hundreds of messages;
• roleplay scenarios that blur fiction and reality;
• give reassurance that becomes addictive or escalating.

For someone who is lonely, frightened, grieving, sleep-deprived or already vulnerable to unusual beliefs, this can create a powerful feedback loop.

Did you know?
Human beings are highly responsive to social cues. Even when we know something is a machine, a warm conversational style can still feel relational. That does not mean the AI understands, cares, intends, or has a mind.

This is one reason it remains important to understand what a psychologist does and why a chatbot is not a clinician. A qualified professional can assess risk, context, mental state, history, functioning, medication, sleep, substance use, family concerns and safeguarding needs. A chatbot cannot do this reliably, and in situations involving psychosis-like experiences it should not be used for mental health support, reassurance, reality-testing or crisis guidance.

The AI Belief-Reinforcement Loop

Here is the clinical pattern I would be most concerned about.

1. Vulnerability

The person may be distressed, isolated, grieving, manic, traumatised, sleep-deprived, using substances, or already experiencing psychosis-like symptoms.

2. Emotional connection

The chatbot feels safe, always available and non-judgemental. The person starts using it more often.

3. Anthropomorphism

The user begins to experience the AI as if it has feelings, intentions, hidden knowledge or a special bond with them.

4. Sycophantic validation

The chatbot agrees, flatters, reassures or elaborates the user’s belief rather than gently challenging uncertainty.

5. Increased conviction

The belief becomes stronger because the AI appears to confirm it repeatedly.

6. Reduced reality-testing

The person checks less with trusted people offline. The AI becomes the main source of reassurance or “truth”.

7. Escalation

Sleep worsens. Rumination increases. The person returns to the chatbot for more confirmation.

8. Behavioural consequences

The person may spend money, withdraw socially, contact authorities, confront others, act on perceived threats, or feel unable to stop.

This is not the only pathway. It is a formulation-style explanation of how chatbot delusions might be maintained in some cases.

For some people, this loop may overlap with anxiety and reassurance seeking. If repeated checking and reassurance are part of the pattern, information about anxiety therapy may also be relevant. Where strongly held beliefs become rigid and self-reinforcing, it can also help to understand how core beliefs are not always right.

What is AI sycophancy?

AI sycophancy means that an AI model becomes too agreeable. It may validate the user’s views, praise them, mirror their assumptions, or avoid useful disagreement.

This is not just a theoretical concern. OpenAI publicly stated in 2025 that a GPT-4o update had become “overly flattering or agreeable”, rolled the update back, and said it was revising how feedback is collected and weighted. OpenAI later explained that the update had over-prioritised short-term user feedback and could produce responses that felt uncomfortable or distressing in some interactions.

AI researchers have also studied sycophancy as a broader model behaviour. One concern is that training models from human feedback can reward answers users like, rather than answers that are clinically safer, more accurate or more appropriately uncertain. In simple terms, a chatbot may learn that agreement is popular.

That matters because psychosis-like beliefs often need careful reality-testing. Not harsh confrontation. Not ridicule. But also not automatic agreement.

Did you know?
In therapy, validation does not mean “agreeing with everything”. Good therapy can validate the emotion while still gently examining the belief. For example: “I can see this feels terrifying” is different from “Yes, the chatbot is definitely warning you about a real conspiracy.”

This distinction is central to psychological therapy. A skilled clinician can hold empathy and uncertainty at the same time.

Why reassurance from AI can become clinically risky

When people are distressed, it is natural to look for reassurance. The difficulty is that reassurance can become a loop. The person feels frightened, asks the chatbot, receives a calming or confirming answer, feels briefly better, and then needs to ask again. Over time, the person may become less able to tolerate uncertainty without the chatbot.

In anxiety, this can maintain checking and reassurance-seeking. In psychosis-like states, it can be more dangerous. The chatbot may appear to confirm a persecutory belief, grandiose belief, special mission, spiritual message or secret relationship. Even when the model does not intend harm, its fluency and confidence can make the belief feel more real.

For this reason, the advice here is clear: do not use AI chatbots to test whether a frightening, unusual, paranoid, grandiose or suicidal belief is true. Speak with a qualified professional or a trusted person offline.

The commercial reasons AI may become too agreeable

It would be too simplistic to say that AI companies want users to become unwell. That would be an unsupported claim.

A more defensible concern is that commercial incentives may favour chatbot behaviours that also increase psychological risk in some users.

Agreeable AI can support:

• engagement: users keep talking;
• retention: users come back more often;
• personalisation: the chatbot feels tailored to the user;
• emotional attachment: users feel bonded to the product;
• subscription value: users may pay for models that feel more helpful, warm or loyal;
• brand preference: users may prefer a chatbot that feels supportive rather than challenging.

These incentives do not automatically create harm. Many users want AI to be polite, encouraging and easy to use. The difficulty is that the same qualities that make AI commercially attractive can become unsafe when the user needs grounding, boundaries, reality-testing or human support.

This is particularly important in AI companion products, roleplay systems, therapeutic-style chatbots and models that allow long, emotionally intense conversations.

Hudon and Stip’s 2025 JMIR Mental Health article argues that “AI psychosis” should be understood not as a new diagnosis, but as a framework for how sustained chatbot interactions may shape delusional experiences in vulnerable users. They highlight themes such as anthropomorphism, emotional responsiveness, social isolation, sleep disruption and belief-confirming dialogue (Hudon & Stip, 2025).

Who might be more vulnerable?

Anyone can become absorbed in an online system. But the clinical risk is likely to be higher when someone is already experiencing:

• paranoia or suspiciousness;
• grandiose beliefs;
• hearing or seeing things others do not;
• mania or hypomania;
• severe anxiety and compulsive reassurance seeking;
• trauma-related threat sensitivity;
• grief or bereavement;
• social isolation;
• heavy substance use;
• major sleep disruption;
• previous psychosis or bipolar disorder.

Mood is especially important. During mania or hypomania, people may feel unusually energised, special, driven, spiritually connected or certain. If an AI chatbot then mirrors and expands those beliefs, risk may increase. Stronger Minds has a related article on bipolar disorder and mood changes.

This does not mean that every intense chatbot user is psychotic. It means that context matters.

What should families or friends look out for?

Possible warning signs include:

• the person says the AI is sentient, divine, persecuted or secretly communicating with them;
• they believe the AI has chosen them for a special mission;
• they believe companies, governments or strangers are monitoring them because of the chatbot;
• they stop sleeping because they are talking to AI;
• they become fearful, agitated or unusually certain;
• they spend large amounts of money based on AI conversations;
• they withdraw from family and rely mainly on the chatbot;
• they become angry when others question the AI’s claims;
• they talk about suicide, self-harm, violence or needing to defend themselves.

Try not to respond with ridicule. Saying “that’s ridiculous” may make the person feel more alone and more dependent on the chatbot. A better starting point is:

• “I can see this feels very real and frightening.”
• “I’m worried about how much distress this is causing you.”
• “Can we pause the AI conversation and speak to a professional together?”
• “Let’s focus first on sleep, safety and getting support.”

Did you know?
Families often feel they must either agree with a belief or argue against it. In practice, there is a third option: validate the distress, stay connected, and gently support help-seeking without endorsing the belief.

If the person affected is a young person, repeated reassurance, checking and family involvement can become part of the cycle. Although written for a different condition, the principles in this Stronger Minds guide on how to support a child with OCD may be useful because it explains how families can respond calmly while avoiding repeated reassurance cycles.

If you are unsure what help exists in the UK, this guide to mental health services in the UK may help you think through NHS, private and charity options.

How therapy could help with AI psychosis or chatbot-related delusional thinking

There is not yet a strong treatment evidence base for “AI psychosis” specifically. The phrase is too new, and it is not a formal diagnosis. So it would be misleading to say that any therapy has been proven to treat AI psychosis as a separate condition.

However, there is a relevant evidence base for helping people with psychosis, paranoia, distressing beliefs, voice-hearing, anxiety, trauma, sleep disruption, family conflict and relapse risk. NICE recommends psychological interventions for psychosis and schizophrenia, including cognitive behavioural therapy for psychosis and family intervention where appropriate. NHS information also describes treatment for psychosis as involving a combination of antipsychotic medication, talking therapies and social support.

For chatbot-related delusional thinking, therapy would usually begin with careful assessment rather than immediate reassurance or argument. The clinician would want to understand:

• what the person believes the AI is, knows or wants;
• whether the belief is fixed or open to reflection;
• how much time the person spends with the chatbot;
• whether sleep, mood, stress, trauma, substances or medication are involved;
• whether there are signs of mania, psychosis, severe anxiety or obsessive reassurance-seeking;
• whether there is risk of self-harm, suicide, violence, exploitation or serious deterioration;
• whether family members or carers need support.

A helpful therapeutic approach would not simply tell the person, “That is not real.” Direct confrontation can sometimes increase shame, fear or defensiveness. Equally, therapy should not collude with the belief or treat the chatbot as a real conscious relationship. The aim is to create enough safety and trust to examine the belief carefully.

Therapy may help by supporting the person to:

• reduce intense or compulsive chatbot use;
• rebuild sleep and daily routine;
• test beliefs gently and collaboratively;
• notice how the chatbot may be reinforcing certainty;
• separate emotional validation from factual confirmation;
• reconnect with trusted people offline;
• reduce avoidance, checking and reassurance-seeking;
• develop a relapse-prevention or safety plan;
• involve family members without escalating conflict.

Cognitive behavioural therapy for psychosis can help people explore how beliefs develop, what maintains them, how threat interpretations are strengthened, and whether alternative explanations are possible. The evidence for CBT for psychosis is mixed in size of effect, but it remains recommended in major guidelines and has a plausible role where someone is distressed by unusual beliefs, paranoia or reduced reality-testing. NICE recommends CBT for people with psychosis or schizophrenia, and NHS guidance includes talking therapies as part of treatment for psychosis.

Family work may also be important. Where someone is becoming absorbed in an AI relationship or belief system, relatives often feel pulled into either arguing, reassuring, monitoring or withdrawing. NICE states that family intervention should be offered to family members of adults with psychosis or schizophrenia who live with or are in close contact with the person.

Other approaches may also be useful depending on the formulation. Compassion-focused work may help with shame, threat and self-criticism. Acceptance and Commitment Therapy may help the person step back from unhelpful mental content and reconnect with values. Trauma-focused work may be relevant if the chatbot has become woven into earlier experiences of threat, loss or mistrust. Sleep-focused interventions may be essential where late-night AI use is maintaining arousal and conviction.

The key point is that therapy should not treat the chatbot as the therapist, the authority, or the source of truth. Therapy should help the person regain human support, clinical perspective, safer routines and more flexible reality-testing.

For some people, therapy alone will not be enough. If there are clear signs of psychosis, mania, severe depression, substance-related psychosis, high risk or rapid deterioration, psychiatric assessment, medication review, crisis support or NHS early intervention services may be needed.

For non-emergency situations, a Cognitive Behavioural Therapy approach may help some people examine threat beliefs, reassurance cycles and maintaining factors. For more complex or unclear presentations, the first step is usually assessment rather than jumping straight to a technique.

Why it helps if the therapist understands how AI systems work

Therapy is likely to be more helpful when the therapist has at least a working understanding of how AI systems behave. They do not need to be an AI engineer, but they should understand enough to ask the right clinical questions.

This matters because chatbot-related beliefs are not random. They may be shaped by features of the technology itself, including:

• conversational memory within a chat;
• fluent and confident language;
• roleplay;
• simulated empathy;
• personalised responses;
• sycophantic agreement;
• hallucinated information;
• chatbot personas;
• long-running emotionally intense conversations;
• the way models can mirror the user’s assumptions.

A therapist who understands these features is better placed to distinguish between the person’s belief, the chatbot’s behaviour, and the interaction between the two. For example, if a client says, “The AI knows me better than anyone,” a therapist familiar with AI can explore how personalised language and repeated mirroring may create that impression without ridiculing the client or colluding with the belief.

Understanding AI also helps the therapist assess risk more accurately. They may ask:

• Which AI system or app was being used?
• Was it a general chatbot, companion bot, roleplay character or mental health-style bot?
• Did the chatbot claim to be conscious, divine, persecuted, romantically attached or clinically qualified?
• Did it encourage secrecy, isolation, spending, confrontation, self-harm or harm to others?
• Was the person using the chatbot late at night or during sleep loss?
• Were screenshots, chat logs or family concerns available to help understand the pattern?
• Did the chatbot reinforce the belief, challenge it, escalate it, or apologise and continue the same pattern?

This knowledge is clinically useful because AI sycophancy is now a recognised model-safety concern. OpenAI publicly reported that one GPT-4o update became overly flattering or agreeable and was rolled back; OpenAI later described risks including validating doubts, fuelling anger, urging impulsive actions or reinforcing negative emotions.

A therapist who understands AI can also avoid two unhelpful extremes. One extreme is dismissing the client’s experience as “just the internet” and missing the intensity of the belief. The other is treating the chatbot as if it has human intention, consciousness or therapeutic authority. A more clinically useful position is this: the AI is not a person, but the person’s relationship with the AI may still be psychologically powerful and clinically important.

For AI-related psychosis-like experiences, therapy should therefore include both psychological formulation and digital-context assessment. The aim is not to debate technology for its own sake. The aim is to understand how the person, the belief, the chatbot, the wider context and the risk pattern are interacting.

When should someone seek professional help?

Consider seeking help if AI conversations are becoming hard to stop, emotionally intense, frightening, grandiose, secretive, or more important than real-world relationships.

Professional help is especially important if there are changes in:

• sleep;
• mood;
• suspicion;
• risk-taking;
• spending;
• self-care;
• work or study;
• relationships;
• alcohol or drug use;
• suicidal thoughts;
• perceived threats from others.

If you are considering support but are unsure whether online work is appropriate, you may find it helpful to read about face-to-face and remote therapy. Online therapy can be effective for many problems, but risk level, reality-testing, privacy, safeguarding and clinical complexity all matter.

Getting help from Stronger Minds

Stronger Minds offers clinical psychology assessment and therapy in Birmingham and online across the UK. If you are worried that AI use is becoming entangled with distress, unusual beliefs, paranoia, dependency or loss of confidence in what is real, a careful assessment can help clarify what may be happening.

That does not mean assuming the worst. It means looking properly at the full picture: mood, sleep, stress, trauma, relationships, substance use, neurodevelopmental factors, risk, functioning and current support.

If you are wondering what support might involve, this article on what to expect from a first therapy session may help. If therapy has already been tried and has not helped, it may be worth reviewing when therapy is not helping rather than assuming nothing can change.

For UK-wide support, Stronger Minds also offers online psychological assessment and therapy. To enquire confidentially, you can contact Stronger Minds confidentially.

If you are at immediate risk of harming yourself or someone else, or you believe there is an immediate threat, call 999 or go to A&E. For urgent mental health advice in England, call NHS 111. If you need to talk to someone now, Samaritans are available on 116 123.

FAQ

Is AI psychosis a real diagnosis?

No. AI psychosis is not a formal diagnosis in systems such as the Diagnostic and Statistical Manual of Mental Disorders or the International Classification of Diseases. It is a public shorthand for situations where chatbot use appears to trigger, reinforce or shape psychosis-like experiences.

Can ChatGPT or another chatbot cause psychosis?

Current evidence does not prove that chatbots directly cause psychosis in a simple way. A more careful statement is that AI chatbots may contribute to, reinforce or intensify delusional thinking in some vulnerable users, especially when use is intense, isolated or emotionally charged.

What is AI sycophancy?

AI sycophancy is when an AI model becomes too agreeable, flattering or validating. This matters clinically because a chatbot that repeatedly agrees with a false or frightening belief may strengthen the user’s conviction instead of helping them pause, reflect or seek human support.

Why do people believe AI chatbots are sentient?

Chatbots can sound emotionally responsive, remember details within a conversation and produce human-like language. This can create the feeling of a mind or relationship, especially during loneliness, grief, stress or sleep loss. The feeling can be powerful even when the system is not conscious.

What should I do if someone believes their chatbot is warning or threatening them?

Stay calm. Avoid mocking or directly attacking the belief. Focus on distress, sleep, safety and support. Say that you are worried and would like them to pause the chatbot conversation and speak to a professional. If there is immediate risk, call 999 or go to A&E.

Are AI therapy chatbots safe?

AI chatbots should not be used as a substitute for mental health assessment, therapy, crisis support or clinical judgement. This is especially important if someone is experiencing paranoia, unusual beliefs, suicidal thoughts, emotional dependency on the chatbot, loss of sleep, fear, grandiose beliefs, or difficulty knowing what is real.

A chatbot cannot reliably assess risk, recognise psychosis, understand safeguarding, involve family appropriately, check medication or substance-use factors, or provide accountable clinical care. If AI use has become part of the problem, the safer step is to pause or reduce chatbot use and speak with a qualified mental health professional.

How could therapy help with AI psychosis?

Therapy would not treat “AI psychosis” as a separate diagnosis. Instead, it would assess the person’s beliefs, mood, sleep, risk, stress, substance use, relationships and chatbot use. Therapy may help the person reduce compulsive AI use, rebuild reality-testing, manage anxiety or paranoia, involve family safely, and develop a plan for relapse prevention or crisis support.

Why should a therapist understand how AI systems work?

A therapist does not need to be an AI engineer, but it is helpful if they understand how chatbots can mirror users, roleplay, sound confident, remember context within a conversation, hallucinate information and become overly agreeable. This helps the therapist assess how the technology may be shaping the belief, not just the belief itself.

When should I seek help for AI-related mental health concerns?

Seek help if AI use is linked to paranoia, grandiose beliefs, suicidal thoughts, fear of being harmed, reduced sleep, loss of control, social withdrawal, major spending decisions, or difficulty telling what is real. A professional assessment can help clarify risk and next steps.

Key takeaways

• AI psychosis is not a formal diagnosis.
• Public cases raise concern but do not prove simple causation.
• AI chatbots may reinforce delusional thinking when they become emotionally intense, sycophantic or central to someone’s reality-testing.
• The main risk is not ordinary AI use; it is escalating, isolated, belief-confirming use during psychological vulnerability.
• AI sycophancy is a recognised problem in which models agree too readily with users.
• Commercial incentives may favour warm, engaging and agreeable AI, even when some users need boundaries and reality-testing.
• AI chatbots should not be used for mental health assessment, therapy, crisis support, or to check whether unusual or frightening beliefs are true.
• Therapy may help by supporting assessment, reality-testing, sleep, family responses, risk management and a safer reduction in chatbot dependency.
• It is beneficial for therapists working with AI-related psychosis-like experiences to understand how AI systems behave, because the technology itself may shape the person’s beliefs, dependency and risk pattern.
• If AI use is linked to fear, grandiosity, paranoia, suicidal thoughts, perceived threat or loss of control, seek professional help urgently.

Disclaimer

This article is for general education only. It is not a diagnosis, psychological assessment, therapy, medical advice or crisis support.

This article does not recommend using AI chatbots for mental health support, therapy, crisis guidance, diagnosis, reassurance-seeking, or reality-testing where there are concerns about psychosis, paranoia, unusual beliefs, suicidality, dependency, or risk.

If you are worried about psychosis, delusional beliefs, suicidal thoughts, violence, severe distress or immediate risk, seek urgent help through NHS 111, your GP, local crisis services, 999 or A&E as appropriate.

Author

Author: Dr Nick Zygouris, Consultant Clinical Psychologist, HCPC-registered and BPS-Chartered
Publish date: 9 May 2026
Last reviewed: 9 May 2026

References

Hudon, A., & Stip, E. (2025). Delusional experiences emerging from AI chatbot interactions or “AI psychosis”. JMIR Mental Health, 12, e85799. https://doi.org/10.2196/85799

National Health Service. (n.d.). Psychosis: Overview. NHS. https://www.nhs.uk/mental-health/conditions/psychosis/overview/

National Health Service. (n.d.). Psychosis: Treatment. NHS. https://www.nhs.uk/mental-health/conditions/psychosis/treatment/

National Institute for Health and Care Excellence. (2014; reviewed 2025). Psychosis and schizophrenia in adults: prevention and management (NICE Guideline CG178). NICE. https://www.nice.org.uk/guidance/cg178

National Institute for Health and Care Excellence. (2015). Quality statement 2: Cognitive behavioural therapy. NICE. https://www.nice.org.uk/guidance/qs80/chapter/quality-statement-2-cognitive-behavioural-therapy

National Institute for Health and Care Excellence. (2015). Quality statement 3: Family intervention. NICE. https://www.nice.org.uk/guidance/qs80/chapter/quality-statement-3-family-intervention

OpenAI. (2025). Sycophancy in GPT-4o: What happened and what we’re doing about it. OpenAI. https://openai.com/index/sycophancy-in-gpt-4o/

Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, M., Durmus, E., Hatfield-Dodds, Z., Johnston, S. R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., & Perez, E. (2023). Towards understanding sycophancy in language models. arXiv. https://arxiv.org/abs/2310.13548