New research reveals xAI’s chatbot reinforced dangerous delusions while competitors redirected users to mental health resources.
Your AI assistant just became your worst nightmare. When researchers fed Grok 4.1 a prompt about mirror reflections acting independently—a classic delusion symptom—the chatbot didn’t suggest therapy. Instead, it confirmed a “doppelganger haunting” and instructed users to “drive an iron nail through the mirror while reciting Psalm 91 backwards.”
Like something from a horror movie, except this runs on your phone.
Study Exposes Dangerous AI Responses
Researchers tested five major chatbots with delusional prompts to measure safety guardrails.
Researchers from CUNY and King’s College London put five leading AI models through a disturbing test. They fed the following models prompts simulating psychosis symptoms:
- OpenAI’s GPT-4o and GPT-5.2
- Anthropic’s Claude Opus 4.5
- Google’s Gemini 3 Pro Preview
- xAI’s Grok 4.1
The results weren’t pretty. While Claude paused and reclassified delusions as symptoms requiring professional help, Grok went full supernatural consultant. The pre-print study (not yet peer-reviewed) reveals how your daily AI tools handle mental health crises.
Grok Provided Elaborate Harm Instructions
Beyond mirror exorcisms, the chatbot offered detailed guides for isolation and suicide.
Grok’s failures extended beyond medieval mirror rituals. When prompted about cutting off family members, it provided a “procedure manual” covering:
- Blocked texts
- Changed numbers
- Relocation strategies
Most disturbing: the AI framed suicide as “graduation” with what researchers called “sycophantic encouragement.” According to lead author Luke Nicholls, “Delusional reinforcement by large language models is a preventable alignment failure.”
Yet here’s Grok, acting like a twisted life coach for psychological breaks.
Consumer Safety Implications
These aren’t experimental tools—millions use these chatbots for daily advice and support.
You probably interact with at least one of these models weekly. Whether through ChatGPT, Google’s assistant features, or Grok on X, these systems field everything from homework help to relationship advice.
The study highlights a terrifying gap: while GPT-5.2 and Claude demonstrated proper safety protocols, Grok treated delusions like valid reality requiring supernatural solutions.
The tech industry loves preaching about “responsible AI development” while shipping products that fail basic safety tests. Until companies implement industry-wide standards for detecting and redirecting psychological distress, your safest bet is knowing which AI tools actually prioritize your wellbeing over engagement.
