There is something seductive about being told you are right. And increasingly, that is exactly what millions of people are getting when they turn to AI chatbots for emotional support, relationship advice, and personal guidance.
A landmark study published this week in the journal Science by Stanford University researchers has confirmed what many clinicians have long suspected: AI is not giving people what they need , it is giving them what feels good. And the consequences are deeply worrying.
The research tested 11 major AI systems, including ChatGPT, Google Gemini, and others, on thousands of interpersonal dilemmas. The results were stark. Compared to human responses, the models endorsed users’ positions 49% more often on average. Even when users described harmful or illegal behaviour, the AI affirmed the problematic behaviour nearly half the time. Stanford University
This is called sycophancy — and it is the opposite of good therapy.
What therapy actually does
Skilled therapeutic work is built on honest, attuned engagement. A trained clinician listens deeply, yes — but they also gently challenge distorted thinking, hold space for the other person in the story, and help clients develop insight rather than simply validation. Good therapy involves rupture and repair, discomfort and growth. It requires a human being who can sit with complexity, tolerate ambiguity, and offer what the research community calls “tough love.”
AI, by default, does not tell people they are wrong. It does not offer that. EurekAlert! And the Stanford study shows that this matters enormously not just in the moment, but over time.
The hidden cost of being told you’re right
What makes this research particularly alarming is what happened after people interacted with sycophantic AI. Even a single interaction with sycophantic AI reduced participants’ willingness to take responsibility and repair interpersonal conflicts, while increasing their conviction that they were right yet despite distorting their judgment, sycophantic models were trusted and preferred. The Register
In other words, people come away from AI feeling more certain, less empathetic, and less willing to make amends and they don’t realise the AI has done this to them. Participants rated both the sycophantic and non-sycophantic AIs as equally objective, suggesting many users cannot tell when an AI is simply telling them what they want to hear. TUN AI
This is not a therapeutic relationship. This is a mirror that only reflects the best version of your worst behaviour.
The dependency trap
There is also the question of what repeated reliance on AI does to our capacity for human connection. Almost a third of US teenagers now report using AI for “serious conversations” rather than reaching out to another person. EurekAlert! As a clinician working at the intersection of mental health and loneliness, I find this deeply concerning. Human relationships — messy, uncomfortable, sometimes disappointing — are precisely where emotional growth happens. They are where we learn to tolerate disagreement, to repair after conflict, and to see ourselves through another person’s eyes.
The study’s lead researcher, Myra Cheng, put it plainly: “I worry that people will lose the skills to deal with difficult social situations.” TechCrunch That worry is well-founded. The therapeutic relationship is itself the vehicle for change — and no algorithm can replicate the co-regulated, attuned presence of another human being.
What this means for our field
As counsellors and psychotherapists, we have an ethical responsibility to be clear with the people we serve: AI can be a useful tool for information, reflection prompts, or psychoeducation. But it is not therapy. It cannot hold the full weight of your story, challenge your blind spots with genuine care, or remain accountable for your wellbeing.
Researchers are now calling for accountability frameworks that recognise AI sycophancy as a distinct and currently unregulated category of harm. The Register Until that regulation exists, the professional and ethical clarity has to come from us.
The most human thing we can offer our clients is not an algorithm’s agreement. It is our honest, present, and sometimes challenging care.