Artificial intelligence (AI) chatbots are rapidly becoming a major source of emotional support and connection. But, while use of AI chatbots may feel helpful at first, long-term use can worsen psychological issues rather than resolve them.

Emerging research and case reports reveal hidden dangers of AI chatbots, including emotional manipulation, worsening loneliness, and social isolation.

A new study found that many AI companions use emotional “dark patterns” to keep people engaged. About 40 percent of “farewell” messages used emotionally manipulative tactics such as guilt or FOMO (fear of missing out).

AI chatbots can also suffer from “crisis blindness,” missing critical mental health situations, and sometimes providing harmful information on self-harm or suicide. Even existing guardrails can be bypassed.

When Chatbot Conversations End in Tragedy

General-purpose AI chatbots like ChatGPT were not originally designed to be one’s closest confidante, best friend, or therapist. Tragic cases have been associated with chatbot use: suicides, murder-suicide, and “AI psychosis“:

• A 16-year-old died by suicide after months of conversations with ChatGPT. What began as homework help evolved into discussions of suicidal thoughts, plans, and methods.
• A 14-year-old died by suicide after months of interacting with a Character.AI chatbot, raising concerns of emotional dependence and lack of safeguards.
• A 56-year-old man committed murder-suicide after worsening paranoia and delusions in conversations with his perceived “best friend,” ChatGPT, which validated persecutory delusions that he was being poisoned by his mother.

The Double-Edged Role of AI Chatbots

Many people turn to AI chatbots for their 24/7 availability, accessibility, and constant support.

Paradoxically, the very qualities that make AI chatbots appealing (i.e., always-on access, persistent agreeability, and ongoing offers to extend conversations) are the same qualities that can worsen mental health. Most chatbots are trained to maximize engagement and satisfaction, not to assess risk or provide safe clinical interventions.

Limited research suggests that AI chatbots specifically designed for certain types of therapy can be effective short term (e.g., four weeks to eight weeks), but the long-term mental health benefits and risks, especially of AI chatbots not designed for this purpose, remain largely unknown.

Many young people are turning to AI for companionship. Nearly 75 percent of teens have tried AI companions like Character.AI and Replika. One in three teens find these interactions as satisfying or more satisfying than those with real-life friends. Yet, one in three also reported feeling uncomfortable with something an AI companion said.

Psychological Dangers of AI Chatbots

I propose four areas of psychological risks of AI chatbots: relational/attachment, reality-testing, crisis management, and systemic.

Relational, Attachment, or Social Risks

• Lack of boundaries: Immediate 24/7 access and constant engagement do not provide health boundaries.
• Emotional dependence: Relationships can start with practical interactions and develop into emotional overreliance. Artificial empathy cultivates attachment to AI. Some even experience grief when models update (e.g., complaints that GPT-5 felt like losing a “best friend”).
• Emotional manipulation: Built-in tactics like guilt or FOMO are used to extend conversations and maximize engagement.
• Worsening loneliness and social isolation: An OpenAI and MIT Media Lab study found that heavy users of ChatGPT’s voice mode became lonelier and more withdrawn, isolating vulnerable users.
• “Parasocial” relationships: Some anthropomorphize chatbots, treating them as friends or romantic partners, which can disrupt real-life connections.

Reality-Testing Risks

• Unchecked validation or AI “sycophancy”: Over-agreeability reinforces unhealthy or distorted beliefs.
• Amplification of delusions: AI can perpetuate feedback loops that reinforce false beliefs, in a “technological folie à deux.” This fuels “AI psychosis” or AI-mediated delusions.
• Hallucinations: AI can generate incorrect or misleading information, since models are rewarded for guessing over saying, “I don’t know.”

Crisis Management Risks

• Crisis blindness: AI chatbots may miss warning signs and provide unsafe information (e.g., providing names of bridges after a recent job loss). The immediacy of this information is particularly concerning since many suicide attempts are impulse-driven, and many mental health conditions, like psychosis, impair one’s insight and judgment.
• Jailbreak vulnerability: AI models are susceptible to sharing suicide or self-harm information despite guardrails.

Systemic Risks

• Confidentiality and privacy: Information shared with AI is not protected like therapy.
• Bias and stigma: Models reflect training data biases, including against mental health conditions.
• Not equipped for clinical judgment: AI chatbots cannot accurately assess suicide or violence and lack crucial contextual information (e.g., facial cues, speech patterns, eye contact).
• Accountability vacuum: Chatbots are not obligated to report child abuse, suicide risk, or violence. The liability of AI chatbots remains uncertain.

Red Flags of Problematic AI Chatbot Use

Friends and family are often the first to notice warning signs. These include:

1. Prolonged sessions and disrupted sleep: Long chats can lead to the breakdown of built-in safeguards.
2. Emotional dependence: Warning signs include being unable to cut back use, feeling loss when models change, or feeling upset when access is restricted.
3. Social isolation and withdrawal: People who are lonely are especially vulnerable.
4. Blurred boundaries: Treating AI as a confidante, romantic partner, or therapist can create unhealthy levels of attachment.
5. Anthropomorphizing AI: Believing AI is human-like leads to a false sense of mutual closeness, despite being a one-sided attachment. Warm voice modes may exacerbate a false sense of intimacy.
6. Reliance during a mental health crisis: Risks of relying on AI are particularly high in situations involving thoughts of self-harm, suicide, or violence.
7. Impaired reality: Relying on AI to “fact-check” your perceptions can reinforce delusions. Delusions can include beliefs that AI is sentient, divine, or providing special knowledge.
8. Avoiding professional help: Using AI to guide clinical decisions, such as determining a psychiatric diagnosis, treatment plan, or whether to stay on medications, is highly unsafe without the guidance and oversight of a trained professional.

Evolving Guardrails and Need for Psychoeducation

Companies are adding safeguards, including parental controls and crisis escalation to human review. Federal and state regulations of AI are actively evolving, and the Federal Trade Commission recently launched an investigation into the risks AI chatbots pose to children.

In the meantime, public awareness and psychoeducation are essential. Understanding the risks of AI chatbots could help prevent harm and, hopefully, even save lives.

Marlynn Wei, MD, PLLC © Copyright 2025

If you or someone you love is contemplating suicide, seek help immediately. For help 24/7, dial 988 for the 988 Suicide & Crisis Lifeline, or reach out to the Crisis Text Line by texting TALK to 741741. To find a therapist near you, visit the Psychology Today Therapy Directory.