I didn't expect AI to make me better with people

Almost every discussion of AI at work starts with tasks: writing faster, summarising documents, producing slides. Some of the most useful things it has done for me have not involved producing anything.

They have involved other people.

That is an uncomfortable sentence to write, because using an AI system to think about relationships, emotions or difficult interactions has become a contested thing to admit — and for reasons that are not hysterical. There have been serious cases of people forming dependent relationships with chatbots, of systems reinforcing delusional thinking, and of vulnerable people discussing suicide and self-harm with software not equipped to respond safely. I don’t want to wave those away, and I don’t accept the defence that arriving with existing problems absolves the system of what happens next: somebody can come to these tools already in difficulty, and the tool can then make it better or worse.

But there’s a jump available from that evidence to a much broader conclusion — that AI has no legitimate role in helping anybody think about human problems at all — and my experience has been close to the opposite.

It’s worth being clear about why that jump is so easy to make, because it isn’t a media conspiracy and I’d rather not pretend it is. I came out of journalism, and I would have chased exactly the same stories. The issue is that harm has a shape and benefit doesn’t. Somebody damaged by a chatbot is a story: there is a before and an after, a person it happened to, a point at which it went wrong. A man who rehearsed a difficult conversation and found it went slightly better than it might have has no shape at all — nothing happened, which is the whole point of it.

So one case is reportable and the other is not, and the consequence is that most people’s evidence about AI and human problems is drawn from what could be written up rather than from what is typical. That doesn’t make the harms smaller or less real. It does mean the absence of the other stories isn’t evidence that they don’t exist, and it’s why so many people who have only encountered this at the surface have concluded that the surface is all there is.

I’m not naturally good in the room

I’m an introvert, and I’ve never been especially confident. Give me an argument to take apart, research to interrogate or a complicated idea to make sense of, and I’m comfortable. Put me in a meeting where I need to read several people’s motivations at once, respond immediately, defend an idea, judge when to concede and navigate the politics around a project, and I’m working much harder than the people who look relaxed.

This matters more the more senior you get. At a certain point doing good work stops being sufficient and you have to get people to support it — which means understanding why somebody is resisting, recognising when you’re solving the wrong problem, and presenting an idea in terms that matter to a person whose priorities are nothing like yours.

The usual name for this is stakeholder management, which makes it sound like a process. For some people it seems close to instinctive. It has never been instinctive for me.

Somewhere to rehearse

Before a difficult conversation I can now describe the situation to a model. Not confidential material that shouldn’t leave an organisation, and not private information about a colleague that isn’t mine to hand over — the problem, the disagreement, what I’m proposing, and why I think somebody might push back.

Then I can interrogate my own position rather than polish it. What am I missing? Why might this person reasonably disagree? Am I making an editorial argument to somebody who cares about revenue? Am I asking someone to absorb work without acknowledging it? What will they ask, where is my case weak, what should I concede and what should I hold?

The most useful instruction is often the bluntest one: challenge my interpretation of what happened.

That works because the model has no obligation to tell me I’m right, and becomes close to useless if I use it as though it does.

Most AI use is a faster version of something that already existed. This isn’t, and search was close to useless for it. You could look up how to handle a difficult conversation with a senior stakeholder and get articles about handling difficult conversations with senior stakeholders — written for nobody in particular, about a category. What made my situation hard was never the category. It was that this person had already rejected one version of the proposal, that the objection about budget was really about something else, and that a history between us changed what I could safely say. Search has no way to take any of that.

You also only ever got one turn. Rephrase the query and you get a different set of documents rather than an adjusted answer — there was no way to say she has heard that argument already and watch the advice change. And the useful thing is almost never in the first response. It is in the fourth, after I have corrected two wrong assumptions and pushed back on something that sounded plausible but wasn’t true of these particular people. That accumulation is the entire mechanism, and nothing I had before did it.

There’s an obvious objection here. If you’re asking a machine how to handle people, aren’t you avoiding the human skill you’re meant to be developing? I think the reverse is closer to what happens. Somebody who reads organisational dynamics naturally performs this reasoning without noticing. I have to make it explicit, and this gives me a place to do that — much like being allowed to write down what you want to say before somebody asks you to say it. The conversation is still mine. The relationship is still mine. I still have to walk in, listen, notice when the situation isn’t the one I prepared for, and decide.

The more valuable use is afterwards

A meeting goes badly. Someone pushes back on something I thought was uncontroversial. An email reads as more confrontational than it probably was. My instinct in that moment is they don’t understand or they’re being difficult, and that is precisely when a second perspective earns its place.

So I reconstruct it and ask for alternative explanations. What might I have communicated badly? What could they have heard that I didn’t intend? Is this resistance to the idea, or to the consequences of the idea — which are not the same thing and are constantly mistaken for each other?

The model does not know what the other person was thinking. It cannot diagnose anybody’s motives from a few paragraphs written by somebody with an obvious stake in the account. What it can do is produce plausible readings I hadn’t reached on my own, and that is enough to move me from certainty about someone else’s motives to curiosity about them. Which is a better position from which to have the next conversation, regardless of who turns out to be right.

It isn’t only colleagues

The clearest recent example wasn’t at work. My downstairs neighbour is also my landlord, and an accidental leak from my flat brought a string of angry messages up through WhatsApp — written in the heat of the moment, assuming carelessness rather than an accident.

My instinct was to match the tone, or at least to defend myself immediately. Instead I described what had actually happened and what she’d said, and asked the same questions I’d ask about a difficult stakeholder: was this reaction proportionate to an accident, what might she reasonably be worried about beyond the message itself, and what would a reply that lowered the temperature rather than raised it actually look like.

It didn’t tell me she was wrong to be upset about a leak into her flat. It helped me see that blaming me for an accident wasn’t a fair response to one, and it helped me write back calmly rather than defensively. The dispute settled faster than the opening message suggested it might.

The mechanism is identical to the workplace case, applied to a relationship with more at stake than office politics: a landlord who can end a tenancy is not a colleague who can make a meeting awkward. Which is exactly why testing “is this fair” against my own certainty, rather than reacting straight from it, mattered more here than it usually does at work.

The mirror should not become an oracle

This is where the line between useful and unhealthy sits, and it isn’t subtle.

I don’t want a system telling me that a colleague is toxic, that I’m definitely right, that my manager feels threatened, or that somebody in my life doesn’t understand me. It doesn’t know enough to make those judgements. It sees the situation almost entirely through what I’ve supplied, so if I feed it a distorted account and keep asking for confirmation, I will get a very fluent version of my own distortion handed back. That is a feedback loop with good grammar, and it becomes genuinely dangerous once the subject moves past workplace friction into mental health, paranoia, abuse or self-harm.

The healthier use is almost the inverse: introduce uncertainty into your own certainty. Ask for competing explanations. Ask it to argue the other person’s case. Ask what evidence would change the reading. Ask what it cannot know from what you’ve told it. Ask where a human being would be the right thing to consult instead.

That is a different relationship with the technology than treating it as an authority on your life.

Reducing the penalty rather than fixing the weakness

This has made me wonder whether the productivity framing is too small for what’s actually happening.

Different people meet different friction at work. Someone who writes badly can get help structuring an argument. Someone who finds numbers hard can have a statistical idea explained four times without the embarrassment of asking a colleague four times. Someone new to a field can ask basic questions privately until the vocabulary stops being a barrier. And someone like me can rehearse.

None of that removes the underlying weakness. It reduces the penalty attached to it, which is a different and more interesting claim.

I’ve written before about AI raising the floor of what an inexperienced person can produce — that the least experienced writer on a project gained the most from a workflow I’d assumed would reward accumulated judgement. This is a related effect pointing somewhere else. That was about the formation of skill. This is about the cost of a trait: organisations have long over-rewarded particular forms of confidence — answering immediately, occupying a room comfortably, assembling an argument while speaking — and those abilities are genuinely useful without being the same thing as good judgement. Anything that lets people who think more slowly arrive prepared is loosening a very old and fairly arbitrary coupling.

I’m much better when I’ve had time to think. This gives me more occasions on which to think before I’m required to perform.

Where I’d push back on myself

The evidence for the central claim is me. I think I’m getting better at this — more deliberate about other people’s incentives, better at anticipating objections, quicker to separate disagreement with an idea from disagreement with me. But that is self-reported improvement in a skill I’m admittedly weak at judging, assessed by the person with the strongest interest in the answer. I spend a lot of this project insisting on evidence, and I don’t have any here beyond my own impression.

Preparation is not the skill. The people who are genuinely good at this are reading a room as it happens and adjusting in real time. Rehearsal cannot teach that, and there’s a plausible mechanism by which it makes things worse: somebody who has prepared thoroughly is more invested in the version they prepared, and more likely to defend a script while missing what is actually occurring in front of them. I recognise that risk because I’ve done it.

And it may be making me more conventional rather than more perceptive. A model trained on an enormous quantity of management writing will tend to produce the standard readings of a situation, because those are the ones that appear most often. The alternative explanations it offers are plausible partly because they are common. Somebody with real insight into a particular person would generate stranger and better hypotheses than I’m getting.

There’s a fourth thing that isn’t a weakness in the argument so much as an obligation. Even carefully anonymised, I am describing colleagues’ behaviour to a commercial system without their knowledge — and while I think that is defensible when the subject is a disagreement about work rather than a person’s character, it deserves to be said out loud rather than covered by a line about confidentiality.

We expected artificial intelligence to make us better at working with information, and mostly it has. I didn’t expect it to make me better at working with people, and increasingly it does — not because anything attends my meetings or tells me what to say, but because it gives me somewhere to find the holes in my own account before I take that account into a room.

Used badly, an endlessly agreeable machine will reinforce whatever you already believe about the people around you. Used with some suspicion of it, it does something more useful. It makes you less sure you were right, which is usually the beginning of getting on with somebody.