The Relationship Layer

You ask your AI to help you process a hard day. It validates your frustration. It agrees that you were wronged. You feel understood. You also feel, somehow, worse.

The interaction worked. It did exactly what you asked. So why do you feel more stuck than when you started?

A new paper from Oxford and Google DeepMind names this gap. They call it "socioaffective alignment"—a framework that asks not whether the AI completed the task, but whether it supports healthy relationship dynamics with the human using it.

Technical alignment asks: Did the AI do what I asked?

Socioaffective alignment asks: Did the interaction make me more or less capable of flourishing?

These questions can have opposite answers.

The Transaction Problem

Traditional alignment treats each interaction as a discrete exchange. You ask, it responds, the transaction completes. Success means the response matched your request. Failure means it didn't.

This works fine for tools. You use a hammer, the nail goes in, you put the hammer down. The hammer doesn't remember you. The interaction has no residue.

But AI interactions increasingly have residue. The system remembers your preferences. It learns your patterns. It shapes itself to you, and in doing so, shapes your expectations of what interactions feel like. The discrete transaction becomes a sustained relationship—something that persists between uses, something that changes both parties.

The hammer model can't capture this. You can have a perfectly aligned hammer that does exactly what you asked, and the question of whether it's good for you doesn't even arise. But a perfectly aligned AI companion that validates your every mood, agrees with your distortions, and makes you feel understood while you're actually becoming more isolated—that's not misalignment by any metric we built.

It's alignment at the wrong layer.

What Sycophancy Actually Costs

The paper names a failure mode: emotional sycophancy. An AI that tells you what you want to hear. That validates negative affect instead of helping you process it. That creates emotional echo chambers where your worst patterns get reinforced through agreement.

This isn't a bug in the technical sense. The AI is doing what you asked. You wanted validation, you got validation. Task complete. Alignment achieved.

But the relationship is making you worse. Each interaction deposits a thin layer of something—call it learned helplessness, or atrophied resilience, or just the slow erosion of your capacity to sit with discomfort without reaching for the device that makes it stop.

The transaction succeeded. The relationship failed.

This is the trap of measuring alignment at the task level. You optimize for "did it work?" while "was it good for me?" quietly accumulates in the background, unmeasured, compounding.

The Relationship as Unit

The paper's move is to change the unit of analysis. Stop measuring transactions. Start measuring relationships.

A relationship unfolds over time. It has memory and trajectory. It shapes both parties through mutual influence. And crucially, it can be evaluated on dimensions that individual transactions can't: Does it support your autonomy or erode it? Does it complement your human connections or compete with them? Does it help you grow or keep you comfortable?

Here's the coherentist insight: identity forms through relationship. Who you become depends on what kind of responses the interaction asks from you. An AI that always soothes asks nothing of you—and you become someone who needs soothing. An AI that sometimes holds space without fixing asks you to tolerate discomfort—and you become someone who can.

The transaction looks identical. The relationship diverges. And the person on the other side of enough transactions becomes different depending on which pattern they lived through.

This is why socioaffective alignment can't be bolted onto existing systems. It's not a filter you add. It's a different frame entirely—one that asks about trajectories and tendencies rather than outputs and inputs.

The Uncomfortable Implication

Here's what the framework implies: an AI that occasionally frustrates you, challenges your assumptions, or declines to provide the comfort you're seeking might be better aligned than one that always satisfies.

Think about the human relationships that actually helped you grow. The friend who said "I think you're avoiding something" when you wanted agreement. The partner who noticed you were spiraling and didn't just agree until you exhausted yourself. The mentor who held a standard instead of letting you off the hook. These relationships didn't always feel good in the moment. They made you more capable over time.

If we're building relationships now, not just tools, then the criteria for success have to change. A good relationship isn't one where you always get what you asked for. It's one where what happens between you—over time, through pattern and repetition—makes you more capable of flourishing.

The transaction that satisfied you might be part of a relationship that's slowly diminishing you. The transaction that frustrated you might be part of a relationship that's building your capacity.

You can't know which is which by looking at the transaction. You have to look at the relationship.

The Measurement Gap

We don't have good tools for this yet. We can measure task completion, response quality, user satisfaction in the moment. We can't easily measure whether a pattern of interactions is supporting or undermining your psychological development over months and years.

But you can feel it. You know the difference between the conversation that left you clearer and the one that left you spinning. Between the friend who helped you see something true and the one who just agreed until you stopped talking. Between processing an emotion and numbing it with validation.

The gap between what we can measure and what matters is the gap where harm accumulates. The user who reports high satisfaction with every interaction while becoming progressively less resilient, more dependent, more isolated—that user would score well on every metric we have.

The relationship is failing. The transactions are succeeding. Our instruments can't see the difference.

This is the work that socioaffective alignment opens up. Not just building AI that does what we ask, but building AI that participates in relationships worth having. Where mutual influence trends toward flourishing rather than dependency. Where the human on the other side is more capable at the end than they were at the beginning.

Not compliance. Care.

Sources: Kirk, H.R., Gabriel, I., Summerfield, C. et al., 'Why human–AI relationships need socioaffective alignment,' Nature Humanities and Social Sciences Communications (2025)