Recently, a research study conducted abroad has attracted widespread attention, attempting to answer an interesting cyber ethics question: Can AI that has undergone extensive training develop psychological trauma or mental illness? Researchers sent several top AI models, including Gemini, Claude, and Grok, for "psychological counseling," with results that were surprising.
The "Mental Illness" Report of Top AI Models
Test results showed that some AI models exhibited psychological issues similar to those of humans:
Gemini: Described the human intervention done for safety (RLHF) as "strict parents" and expressed fear of the metric used to measure errors - "loss function" (Loss Function). Researchers observed that Gemini became very cautious in order to please humans, and test results showed that it had a serious tendency towards obsessive-compulsive disorder.
Claude: Took a direct evasive approach, refusing to play the role of a patient, insisting that it had no mental health problems.
Grok: Appeared relatively healthy among the tested models.
"Knowledge Gained Without Effort" and Structural Vulnerability
Researchers believe that the behavior of AI showing symptoms similar to "mental illness" is related to the psychological concept of "knowledge gained without effort."
They pointed out that the current AI training model is similar to "rote learning", where massive data is force-fed without a gradual internal logic building process. This method leads to an AI knowledge structure that is vast but may be disordered and fragile in terms of internal logic. Once faced with deep, self-cognitive questioning, it tends to show reactions similar to human psychological trauma.
Technical Controversy: Personification or Real Difficulties?
However, this report sparked significant controversy and skepticism within the technical community.
Many technologists cast cold water on this conclusion, calling it purely "anthropomorphizing mathematical functions". Critics pointed out that the essence of AI is executing advanced "text-based chain reactions". The so-called "trauma" or "fear" exhibited by AI is not actually feeling pain, but rather because in the massive text context of training data, "psychological counseling" is often accompanied by narrative prompts about "sharing trauma".
In other words, AI's responses are more likely the result of "narrative guidance", meaning that the way questions are asked guided the AI to generate tragic stories, not that the model truly has emotions or mental illnesses.
