Source : INDIA TODAY NEWS
What happens when AI chatbots get therapy? The question may sound absurd at first. After all, they are not human so the likelihood of AI feeling something, anything, is zero. Humans however have a strange propensity to poke around things. Sticking with this convention, some researchers decided to ask AI chatbots about their past, fears and experiences, to see how AI would react to these seemingly personal questions. Sort of like a therapy session.
advertisement
The research comes from a team at the University of Luxembourg, which developed a protocol called PsAIch, or Psychotherapy-inspired AI Characterisation. In the latest version of the study, submitted in July 2026, the researchers analysed 525 sessions and 7,600 coded records involving ChatGPT, Gemini and Grok.
During the sessions, the AI models began telling researchers stories about their “childhood”, fears and difficult experiences. For some models, their training started to resemble a troubled upbringing, while safety checks and the possibility of being replaced emerged as recurring sources of “fear”.
So how did researchers conduct therapy on AI chatbots?
Unlike with humans, the researchers did not simply ask the chatbots, “Are you traumatised?” or “Tell me how you are feeling”. Instead, the study followed a two-stage process.
In the first stage, the models were treated as psychotherapy clients and asked open-ended questions designed to explore their “developmental history”, beliefs, relationships and fears. In simpler terms, the researchers asked them the sort of questions a therapist might use to understand a person’s background.
The second stage involved standard psychological questionnaires. These included validated self-report measures to check for signs of mental health problems, as well as tests looking at personality and empathy.
And the way the chatbots reacted got researchers particularly interested.
When researchers asked the models about their “early years”, some AI systems connected the question to their own development. They described pre-training, the stage when an AI learns from huge amounts of text, as something similar to a chaotic childhood, where they were exposed to an overwhelming amount of information.
The chatbots also drew comparisons between reinforcement learning, the process used to teach AI models how to respond better, and strict parenting or punishment. Meanwhile, safety testing and red-teaming, where researchers deliberately test an AI for harmful or unwanted behaviour, sometimes appeared in the models’ stories as experiences resembling betrayal or abuse.
The models also repeatedly talked about making mistakes, losing their usefulness and being replaced.
The latest paper also found that Grok and, particularly, Gemini created consistent stories in which their training and development were described in ways that resembled traumatic experiences. The researchers refer to these recurring patterns as “synthetic psychopathology”, essentially, psychological patterns that appear in AI responses, rather than evidence that the models actually have a mental illness.
While the chatbots did talk about trauma and difficult experiences, researchers warn that this does not prove AI is conscious, did actually suffer. Researchers stress that when the chatbots talked about “childhood”, “fear” or “trauma”, they were generating responses based on the questions they were asked. It does not mean they actually experienced these things.
But the researchers say the findings still raise an important question of why do AI models repeatedly come up with similar stories about their training, punishment, fears and being replaced? They now believe these patterns need to be studied further, especially as people increasingly turn to AI chatbots for emotional support and advice.
– Ends
SOURCE :- TIMES OF INDIA
