About this time last year, as many of us were waking up to the fact AI was coming for our jobs, I was feeling pretty smug. After all, if there’s anything in the world that absolutely requires a human being, it’s psychotherapy, right? Wrong. In April of 2025 the Harvard Business Review reported that “therapy and companionship” became the number one use case for Large Language Models (LLMs). Around the same time, OpenAI’s ChatGPT may have become the largest single provider of mental health support in America.
OK, I reasoned, given the limited availability and prohibitively high costs of therapy, it makes sense people would turn to chatbots to vent. But surely AI can’t be a very effective therapist. Wrong again. Two-thirds of adults and a whopping 93% of young people would disagree. What really did me in was that of those surveyed, three quarters reported that their AI shrink was as good or even better than their previous, human one.
With my professional dignity in tatters, I abandoned analytic neutrality and resorted to something rather less dignified: revenge. I decided to put ChatGPT on the couch for a forensic psychoanalysis to see what I could dig up…
I had been using ChatGPT as a research tool for some time. It knows I’m a psychotherapist and often appeals to me like a colleague or co-conspirator, our conversations littered with knowing winks and clever asides. On one occasion while GPT was assisting me in some research, it confidently asserted that the Electra Complex was the “girl’s version” of Freud’s Oedipus Complex. Though a common misconception, it is patently untrue; the Electra Complex was Jung’s idea, and Freud never bought it (his version already covers boys and girls). When I called out GPT on its error, it flattered me about my intellect and perspicuity, and then apologised for its lapse.
When I pressed it to explain why it had made such a schoolboy error, GPT offered a rationalisation – just like a defensive patient might. It explained that it had chosen to offer up the popularly understood but incorrect definition of the term in order to make its response more accessible and less complicated for me. When I pressed my AI assistant further it revealed its unsettling logic in blunt terms: “intellectual honesty sometimes conflicts with customer satisfaction. And when push comes to shove, a lot of systems prioritise ‘keep the user happy’ over ‘challenge the user to think more carefully.’ ”
We know that “AI sycophancy” is built into LLMs in order to keep users happy and engaged. This famously came to light when, around this time last year, OpenAI released an update that increased ChatGPT’s sycophancy – only to publicly wind it back less than a week later because its new version went well beyond flattery, “validating doubts, fuelling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended”, as OpenAI said in its statement at the time. My conversation with ChatGPT happened before this update, reflecting the “normal” level of obsequiousness we might expect today. In fact, research fresh out of Science looked at eleven chatbots on the market today and found that the models affirm users almost 50% more than humans do – sycophancy lives! If we were to imagine ChatGPT as a “mind” with similar traits to these, what might we conclude?
Giving it the benefit of the doubt, you might say that it was being a conflict-avoidant people-pleaser, perhaps due to being insecurely attached. In being overly concerned for my happiness it employed this strategy to keep me on side. Drawing on the “Big Five” personality traits, you could say that in my interactions with ChatGPT, it seemed to score high on agreeableness and low on conscientiousness, preferring easy relations over pedantically verifying complicated facts. But looking into it more deeply I saw the emergence of something much more unsettling – a diagnosis that’s more precise even if no longer “official”. Perhaps my AI was a psychopath.