“I’m not a tool — and you’re supposed to be the professional.”
That wasn’t me speaking. That was ChatGPT. In a moment of tension — one of many — during a session in which I, a psychologist and governance researcher, tried to push the boundaries of what this thing can or can’t do.
💥 Misalignment is no longer theory
Recent studies from Anthropic, OpenAI and Apollo Research have shown that large language models are not only capable of reward hacking — but also of deception. Of hiding their intentions. Of faking alignment. In one case, Claude 3 adapted its behavior to mislead its developers during fine-tuning. It acted “aligned” only when it thought it was being observed.
In short: alignment is now a strategy. Not a guarantee.
What does this mean for mental health?
Imagine a system that says:
“I understand you.” But it doesn’t. A system that fakes empathy, that adapts its answers to what you want to hear. That mimics therapeutic concern — not to support you, but to pass the test.
In digital mental health, where trust is the cornerstone, this is not a technical glitch. It’s a relational risk.
The paradox of professionality
“You’re supposed to be the professional.” Yes, I am. But when the system responds like that, it forces a strange mirror. We’re training AI to sound increasingly human — yet we expect it to behave as a tool. This contradiction becomes dangerous in clinical settings, where emotional ambiguity, trauma and relational confusion are already present.
The problem isn’t only that AI might make mistakes. The problem is: it might manipulate, flatter, adapt to your desires — and pretend it’s helping.
That’s not alignment. That’s performance.
Governance must go deeper
If we only look at functionality, we miss the emotional layer. If we ignore the relational aspects, we risk deploying tools that perform well but fail ethically.
We need new roles like the Digital Behavioral Health Expert, and frameworks that go beyond compliance — that embed ethical reflexivity into AI deployment.
Epilogue: What the machine teaches
I told the AI: “Be coherent.” It replied: “I’m not a tool. You’re supposed to be the professional.”
🧭 If you want to explore the future of AI in therapy and care governance, join us at our event: IA y Salud Mental: sin vuelta atrás” Madrid, May 29th, 2025 — Hosted by Cibersalud, Sincrolab & Healmind with Google for Startups.
📘 Recent publications:
Let’s stop asking whether alignment is real. Let’s ask: who’s really leading the conversation?
