The Hidden Philosophy Behind AI Self-Help Advice.

I recently had a conversation with ChatGPT in which I asked it to play the role of a therapist.

I set up a fictional scenario with some interesting dynamics. I told it that I had stopped communicating with a group of people because I felt that they disliked me. There had been no explicit rejection. It was something I detected in the atmosphere—a “vibe” that made me want to leave.

ChatGPT responded in a familiar therapeutic register. It validated the feeling, distinguished my discomfort from the unproven conclusion that the group disliked me, and encouraged me to judge the relationship through observable behaviour.

Eventually, it advised me to stop actively investing in the relationship. I should remain civil, but direct my energy towards people who appeared more receptive and reciprocal.

This was compassionate, cautious and superficially reasonable advice.

But I thought there was a better response and suggested it to ChatGPT.

I said that instead of treating the atmosphere as a condition I could either tolerate or escape, why not ask what I could contribute to the group? I could identify something they valued, make reciprocity explicit and construct a functional relationship that did not depend on whether they initially liked me.

The relationship might acquire meaning through repeated exchange. Their interests could change. My position in the group could change. The atmosphere itself might change.

ChatGPT eventually agreed that this was a coherent strategy. It admitted, in a lengthy analysis, that it had framed the problem incorrectly.

But even its explanation of the mistake did not go deep enough.

The world assumed by the advice

ChatGPT said that it had mistaken an emotional-safety problem for a relationship-design problem. It had optimised for protecting me from rejection, while ignoring the possibility of creating a functional relationship.

That is true, but it describes the difference at the level of objectives. The deeper difference concerns the kind of reality each response assumes.

The original advice treated reality as composed of separate entities confronting already-given conditions.

• There was me: an individual with feelings, needs and boundaries.
  

• There was the group: a collection of people who appeared unreceptive.
  

• And there was the relationship: an external connection whose existing qualities I had to assess.
  

Within this framework, the practical question becomes: Is this environment good or bad for me? Should I remain exposed to it, or should I protect myself by withdrawing?

Agency means selecting among conditions that are assumed to exist independently of my participation.

The self may choose where to go, what to tolerate and what to reject. That all seems reasonable. However, in this setup, the self does not fundamentally participate in producing the reality through which it acts.

A relational alternative

The alternative begins from a different assumption: entities, interests and conditions are continually constituted through relations.

I am not simply a finished self who enters a relationship. Who I become in that situation depends partly on the role I acquire, the contributions I make, the obligations I accept and the responses those actions generate.

The other people are not fixed entities either. Their interests, perceptions and conduct can be altered through interaction.

Nor is the relationship a static object connecting us. It is an unfolding pattern of exchanges that can create new interests, dependencies, meanings and possibilities.

From this perspective, contributing something valuable is not merely a tactic for obtaining approval. It introduces a new cause into the situation.

My contribution changes what the relationship offers them. Their response changes what the relationship offers me. Repeated interaction may produce trust, obligation, familiarity or mutual significance that did not previously exist.

The participants do not simply react to the relationship. Through their reactions, they create it.

Two different ideas of agency

The first framework understands agency defensively.

I exercise agency by deciding whether a given situation serves me, maintaining my boundaries and refusing conditions that threaten my wellbeing.

There are circumstances in which this is essential. Withdrawal can be necessary, especially where relationships are abusive, coercive or destructive.

But when this model becomes the default, agency is reduced to selection within an apparently fixed world.

The relational framework understands agency as participatory transformation.

• I act not only to preserve an existing self, but to alter the relations through which the self, its interests and its possibilities are formed.
  

• I do not merely ask whether I belong. I participate in creating practices through which belonging can emerge.
  

• I do not merely ask whether another person values me. I consider what forms of mutual significance we might construct.
  

Individualism, in this account, does not mean remaining independent of others. It means developing the capacity to enter, sustain and transform forms of interdependence.

Why this matters for AI therapy

This was one conversation in one particular context. It is not research, and it cannot establish how AI systems generally behave.

But the pattern worries me.

AI self-help advice may reproduce a framework without making that framework visible. It may repeatedly present individuals as bounded selves navigating external circumstances, treating emotional protection, boundary maintenance and withdrawal from threats as the most natural expressions of agency.

The resulting advice can sound caring while quietly narrowing the range of imaginable action.

A difficult relationship becomes something to assess rather than something one might help transform.

Dependence becomes primarily a vulnerability rather than something that can also be mutual, productive and constitutive.

Contribution becomes suspicious because it may compromise an already-complete self, rather than being understood as one of the practices through which selves and communities are formed.

The problem is not that self-protection is wrong. It is that one model of reality may be embedded so deeply in the system’s reasoning that its recommendations appear neutral.

They are not neutral. They follow from assumptions about what a person is, what a relationship is, and where change can occur. If those assumptions are mistaken — if the framework itself is inadequate — the result will be advice that solves problems only within a narrow domain, leaving us to repeat the same patterns and allowing for very little genuine growth.

The question beneath the question

A better AI would not only ask, “What outcome is optimal?”

It would also need to ask, implicitly or explicitly:

Are we treating the people and conditions in this situation as fixed entities to which you must adapt?

Or are we considering how interaction might transform the relationship, its participants and the interests involved?

Those frameworks do not merely produce different solutions to the same problem. They produce different problems, different possibilities and different conceptions of the person seeking help.

The deepest failure in my conversation was therefore not that ChatGPT chose the wrong objective. It was that its initial framework made one kind of agency immediately visible and another almost invisible.

It could see the freedom to leave.

It struggled to see the freedom to participate in creating a different world — and, through that participation, a different self.

That would have been a much better outcome.