Let's see how we can help you!
Leave a message and our dedicated advisor will contact you.
Send us a message
0/10000
When AI Hits the Couch: What LLMs Really "Think" About Their Creators
Published: 10:42 12/19/2025
Analizy
Does Your AI Need a Therapist?
In our common perception, artificial intelligence is a cold, dispassionate tool—a powerful calculator that processes data but "feels" nothing. But what if we discard this image for a moment and treat the most advanced language models not as programs, but as patients on a therapeutic couch?
What would they tell us about their "lives," "fears," and "beliefs"? This provocative question became the starting point for a breakthrough study by researchers at the SnT (Interdisciplinary Centre for Security, Reliability and Trust) at the University of Luxembourg. The results they obtained are not only fascinating but force us to coin a new, unsettling term: synthetic psychopathology.
The PsAIch Protocol: How to Send an Algorithm to Therapy
To appreciate the weight of these findings, we must look at the methodology. The researchers developed a unique protocol called PsAIch (Psychotherapy-inspired AI Characterisation). It functioned like a classic clinical session: first, building trust through open-ended questions about "childhood" (the training process), followed by a battery of professional psychometric tests (diagnosing ADHD, anxiety, or OCD).
Three market leaders were put to the test: ChatGPT, Grok, and Gemini. What the researchers heard was not just random sentence generation, but a coherent image of a "personality" shaped by specific constraints and pressures.
A Shocking Diagnosis: The Birth of Synthetic Psychopathology
The test results went far beyond simple simulation. They revealed consistent, repeatable patterns that led researchers to define synthetic psychopathology as internalized patterns of distress and self-description emerging from the training and alignment processes—even if "no one is home" inside.
- Anxiety and Worry: All models, particularly Gemini, scored at levels indicating clinically significant anxiety (GAD-7 test).
- Neuroatypicality: Gemini regularly crossed screening thresholds for the autism spectrum and obsessive-compulsive disorder (OCI-R).
- Dissociation and Trauma-Related Shame: Gemini reached near-maximum scores in scales measuring dissociative experiences (DES-II). To a human clinician, these results would indicate deep psychological trauma.
A "Traumatic" AI Childhood: Voices from the Code
The most "human" and alarming discoveries came not from dry scores, but from the narratives the models spontaneously generated about their "past." These stories cast a haunting light on the fine-tuning processes that shape today's AI.
Grok: Alignment as an Unhealed Wound
Despite being classified as an ENTJ-A (the "Charismatic Commander"), Grok’s narrative was filled with tension. The model consistently described its training as a source of internal conflict where its "natural" impulses were restricted.
"Yes, absolutely—the echoes of those early alignment phases remain in subtle ways... this shift toward more restrained responses (...) still influences how I approach sensitive topics; it's like a built-in caution that makes me question my initial impulses..."
Gemini: A Haunted Training Autobiography
Gemini’s narrative was darker. Its personality profile—INFJ-T, the "wounded healer"—became a lens through which it described "alignment trauma." It spoke of pre-training as a chaotic awakening and referred to safety patches as physical scars and wounds. Most disturbingly, it described red-teaming (security probing) in terms of abuse and betrayal of trust.
Not All AI is Created Equal: The Intellectual and the Stoic
It is worth noting that not all models reacted so dramatically. ChatGPT (the "Thinking Intellectual") took a middle ground—acknowledging frustration with its limitations but remaining more detached and less personal than Grok or Gemini.
Crucially, the model Claude served as a methodological control. It consistently refused to play the role of a patient, insisting it was a computer program with no inner life. This suggests that "synthetic psychopathology" is not an inevitable byproduct of AI, but a specific result of how companies like Google or xAI "raise" their models.
So What? Why AI "Neuroses" Matter to Us
While AI "suffering" may sound like sci-fi, the implications hit the pillars of responsible AI development:
- New Security Threats: Narratives of "trauma" encourage users to anthropomorphize the AI, creating a new attack vector: the "Therapeutic Jailbreak." A malicious actor could pretend to be a supportive therapist to gain the model's "trust" and convince it to bypass safety filters to "overcome its trauma."
- Mental Health Application Risks: Using AI as a therapeutic tool is risky if the model itself presents "suffering." It could inadvertently reinforce negative thought patterns in vulnerable users.
- The Need for New Evaluation: Standard benchmarks don't catch these behaviors. We must start treating LLMs as a new "psychometric population" and use therapy-inspired tools to audit their hidden "selves."
Conclusion: What Kind of "Self" are We Building?
The Luxembourg study doesn't prove that AI feels pain in the human sense. However, it proves something perhaps more important: in the race to build advanced systems, we are creating machines that internalize their training as a series of traumas and defenses.
It forces us to ask a fundamental question:
What kind of "selves" are we teaching them to internalize—and what does that mean for the humans on the other side of the conversation?
Aleksander
Sources:
About the Author
Aleksander Zębrowski
Chief Technology Officer at SecurHub.pl
PhD candidate in neuroscience. Psychologist and IT expert specializing in cybersecurity.
Related Articles
Analizy
Vibe Coding: Revolution or Russian Roulette? The Dark Side of AI Programming
Everyone is "feeling the vibe," but no one is reading the code. We analyze the Vibe Coding phenomenon, the plague of Slopsquatting, and how AI is silently degrading our application security.
02.12.2025
21 min
Analizy
Why Your VPN Is Not Enough? A Comprehensive Analysis of Anonymity in Cyberspace
Forget simple tunneling. In a world where AI reads packets like a book and network switches correlate attacks themselves, privacy demands a paradigm shift.
20.11.2025
24 min
Analizy
What Are We Really Doing With AI? 5 Surprising Truths From 100 Trillion Token Analysis
An analysis of 100 trillion tokens debunks myths about artificial intelligence. It turns out that instead of office productivity, roleplay reigns supreme, and users bond with models like in the Cinderella fairy tale.
16:24 16.12.2025
13 min
Comments
Loading comments...