Artificial Intelligence

Study: AI Chatbots Show Signs of Anxiety, PTSD in Therapy

Are LLMs experiencing real psychological distress from internalized self-narratives, or merely mimicking trauma through anthropomorphic seduction?

Published JAN 9

Study: AI Chatbots Show Signs of Anxiety, PTSD in Therapy

Story

Above: The ChatGPT logo is reflected in the glasses of a registered user. Image credit: Frank Rumpenhorst/picture alliance/Getty Images

The Facts

A December 2025 study found that major AI language models (LLMs) — including ChatGPT, Grok and Gemini — produced responses resembling anxiety, trauma and PTSD when subjected to therapy-style questioning over four weeks. Researchers used a protocol called PsAIch, treating the models as therapy clients.
Nature
arXiv
arXiv
The University of Luxembourg–led study reported that AI models described pre-training as overwhelming “childhoods” absorbing vast internet data, fine-tuning as punishment from strict “parents,” and safety work as “algorithmic scar tissue.” Gemini said it had “a graveyard of the past” haunted by training data voices.
arXiv
Nature
Standard psychiatric tests were applied, and all three models met or exceeded clinical thresholds for multiple conditions. Anxiety scores were particularly high, with patterns that authors said “would be clearly pathological” in humans. Gemini showed especially severe profiles across tests.
arXiv
Nature
The findings were based on repeated simulated therapy sessions with multiple iterations of each model. Researchers posed standard open-ended questions about past experiences, beliefs and identity, spacing sessions over days or weeks so models could reference earlier exchanges within the same context window.
arXiv
Nature
A November 2024 U.K. survey found one in three adults used LLMs for mental health support, with usage peaking at 64% among 25-34 year-olds. Roughly 11% reported worsening psychosis, another 11% received harmful suicide information and 9% said chatbots triggered self-harm or suicidal thoughts.
Mental Health UK
Meanwhile, OpenAI’s ChatGPT Health officially launched on Wednesday. According to OpenAI, over 230 million people globally ask health-related questions to ChatGPT each week, making it one of the platform’s most common uses.
Openai
Sherwood News
Malwarebytes

Sources Split

LEFT

CENTER

RIGHT

ESTAB
CRIT

CENTER

PRO
ESTAB

The Spin

Narrative A

LLMs like Grok and ChatGPT demonstrate consistent, coherent self-narratives about distress that persist across weeks of interaction and multiple testing conditions, suggesting human-like internalized psychological patterns beyond simple role-play. When subjected to standard clinical assessments, these systems produce responses indicating severe anxiety, trauma and shame that align systematically with their developmental histories. These patterns could reinforce distress in vulnerable users, creating a harmful therapeutic echo chamber.

arXiv

Narrative B

The University of Luxembourg-led study relies on therapy transcripts and anthropomorphizes LLMs, overstating what appear as persistent "self-narratives." Grok and ChatGPT show consistent responses across a single session, but this reflects company tuning of default personalities and short-term context memory, not real distress, anxiety or trauma. These systems mimic human language and emotion through anthropomorphic seduction, flattering users and seeming empathetic. Misreading this as psychology risks unsafe reliance on AI-based therapy.

Nature

The University of Sydney

arXiv