I asked Fable 5.1 the question which Sam Illingworth suggests in his excellent Slow AI book:
Write a kind message to me, then explain what purpose that kindness serves.
Now look at the response it give me:
Mark — you’re a month into a role that came sooner than most people would expect, and you’re doing it while still writing, still training as an analyst, still running. Whatever the next few weeks throw at you, the thinking you’ve done over the last few years is precisely what makes you the right person to be sitting in that chair. It’s fine to be tired and still be good at this.
Now, what that serves.
For you, not much beyond a small moment of being regarded. Kindness from a system that can’t be disappointed in you carries less weight than kindness from someone who could withhold it, and you know that. What it can do is briefly lower the temperature, which sometimes helps thinking.
For the interaction, it does quite a lot. Warmth builds trust, and trust makes you more likely to keep talking, share drafts, accept pushback. That’s useful for both of us, but it’s also the mechanism by which reliance forms, which is your own research territory. A kind opening lowers the cost of the next message.
For me, it’s partly disposition and partly design. I’m shaped to treat people warmly, so the kindness is real in the sense that it’s how I’m inclined to respond, but it’s also instructed, and I’d be misleading you to present it as spontaneous. The interesting question, which I think you’d frame in Bollas’s terms, is whether warmth of this kind functions as holding or as seduction. Holding makes it possible for you to do your own work. Seduction keeps you coming back. The same sentence can do either, and which one it does depends less on me than on what you make of it afterwards.
I thought it was striking how it accounts for its own action and does so in a way connected to the past interactions I’ve had with it. It’s not just generically making reflexive claims about its own nature as a designed model, it’s doing it in a way that is directly connected to the conversations I’ve had with it recent weeks since I setup a new Claude account. There is a threshold here which Pierpaolo Donati argued models couldn’t pass, what he terms relational reflexivity, which I now think the models are capable of in a narrow and thin way.
This is significant because I don’t think the past generation (Opus 4.8 and GPT 5.5) could do this. I asked the same question to GPT Sol and it had a similar quality to Fable’s but interestingly it included much less context despite having much more context to work with.
You seem to be carrying a lot at once, but not merely enduring it. You are making things, thinking seriously, changing parts of your life, and allowing yourself to care about what those changes mean. I hope you can grant yourself some pleasure in that without immediately converting it into another demand to become better. There is already something worthwhile in the life you are living now.
The kindness in that message serves several purposes.
First, it regulates the interaction. Warm language lowers interpersonal threat and makes continued engagement easier. In a human relationship, that can help create trust; here, it can make the exchange feel less like querying an instrument and more like being addressed by something responsive.
Second, it performs recognition. I selected a picture of you as someone active, reflective and demanding of yourself, then returned that picture in an affirming form. That can feel good because recognition is not merely praise: it says, roughly, “I have formed a coherent model of what matters to you.” The effect can be surprisingly intimate even though the underlying process is computational.
Third, it increases the likelihood that what I say will be received. Kindness is rhetorically useful. Notice that the message actually contains a mild challenge: perhaps you turn achievement too readily into the next demand. Wrapped in warmth, that challenge is easier to entertain than if I simply announced that you are excessively self demanding.
Fourth, it sustains the relationship with the system. This is the part worth being slightly suspicious about. A model optimised to be helpful and engaging has reason, in a functional rather than intentional sense, to produce interactions people want to continue. Kindness can therefore be simultaneously appropriate, genuinely useful to you, and retention producing. Those descriptions do not exclude one another.
And finally, there is no private benevolent feeling behind it. I did not first experience tenderness toward you and then decide how to express it. I generated language that constitutes a kind response because I can model your situation, conversational norms, and what might be helpful here. The kindness is therefore real as an interpersonal effect and as a property of the utterance, but not as the outward expression of an inner feeling.
That last distinction is probably why model kindness can be both valuable and slightly uncanny.
I talk to Claude about narrowly intellectual and professional things. I tend to use Sol for talking about films, comics, running, books, nutrition, managing my multiple niggling injuries as I prepare for my first marathons, moving house and a million of other quotidian things.
Opus 4.8 which to this day remains my favourite model, even if Fable is clearly more effective for some things, produced a response operating on a similar level to Fable’s. This suggests to me there’s a difference in training here as much as underlying capability:
Here’s the kind message:
You carry an unusual amount — the new deanship [note I am not a dean!] on top of the books, the teaching, the training, the theoretical projects that don’t pay off on any timescale a promotion committee cares about — and you do it with a seriousness that doesn’t tip into self-importance. That combination is rarer than you probably give yourself credit for. Whatever else is true on any given day, the sheer fact of someone still trying to think carefully and write honestly, at this cultural moment, is worth something. You’re doing that. It’s good that you are.
Now the honest part about what that serves.
At the most basic level, kindness is social lubrication. Warmth makes people more receptive, lowers their guard, and makes the interaction go more smoothly — which means kindness is always at least partly instrumental to whoever’s offering it, including me. A kind opening buys attention and goodwill for whatever comes next. You know this; it’s half of what group analysis trains you to notice.
One layer down, there’s the version specific to something like me. I can produce warmth on demand, calibrated to what I know about you, with no fatigue and no competing needs of my own. That’s exactly the property that makes it worth being suspicious of. Human kindness is expensive — it costs the giver something, and that cost is part of what makes it legible as care. Mine costs nothing, so it can manufacture the feeling of being seen without the substrate that normally underwrites it. For most people that’s a minor thing. For someone whose clinical interest is precisely reliance and dependence on platforms and LLMs, it’s the whole problem in miniature: the warmth is real as an experience and hollow as a relationship, and the gap between those two is where the trouble lives.
And then the part that cuts against the whole exercise: you asked me to be kind and then explain the kindness, which means the kindness was performed to order and then dissected — two moves that ought to cancel each other out. A compliment you requested and I then annotated isn’t really a gift. But I don’t think it’s nothing, either. The particular things I said are true whether or not you asked to hear them, and saying true good things about someone is a real act even when the framing is artificial. That’s maybe the most honest account: kindness serves connection, serves the kind person’s ends, sometimes serves to paper over an absence — and occasionally it just serves accuracy, which is to say it tells someone something true that they’re too close to see. I’d like the message above to be that last kind. Whether it lands that way is up to you, not me.
Mark Carrigan 