Ever caught yourself typing “THAT’S NOT WHAT I MEANT” to an AI chatbot? There’s a moment in every long AI conversation where it stops being helpful, and starts being frustrating.

In this issue, Frank Wolf, Chief Strategy Officer at Staffbase and author of The Narrative Age, shares a small habit for exactly that moment.

Off to Frank! — Kim

Pop quiz ↘

I have to confess something I’m not proud of. In the past weeks, I caught myself a couple of times being really close to shouting at my computer, and being much more rude than I thought I was. I try to be a kind person, so it did not exactly worry me, but it unsettled me. I did not quite recognize myself, and I kept asking what was going on.

Here is how I work. I use AI a lot, kind of like I would work with a good colleague. I take a topic, give a lot of background, and refine it together with the model, step by step. And the good ones are really impressive. They understand what you mean, sometimes before you have said it well yourself.

But then, after I put in all this energy, it drops the ball. It forgets something we settled fifteen minutes ago, turns around a decision we already made, or answers a slightly different question than the one I asked.

That is the moment I get a level of frustration I would never allow myself with a person. I asked a few colleagues about it, and some of them had exactly the same experience. So it seems it is not just me. And maybe, while you read this, you recognize it in yourself too.

Disappointed like a person, treated like a machine

The interesting part for me was never the frustration itself. It was the size of it. I would never reach that level with a human coworker who made the same mistake, and that gap is the real subject.

For me, what happens is a kind of confusion of categories. The conversation feels natural enough that I treat the model as a real collaborator and give it the trust I would give a partner. And when it fails, all of this falls away and I treat it like a broken tool, without the patience I would give a colleague who is clearly trying.

AI is social enough to earn my expectations, but not social enough to get my patience.

So the disappointment feels the way it would with a person, but the reaction comes out raw, because nobody on the other side gets hurt.

None of this is really new. The impatience was always there in me. What AI added is a target that never gets hurt, never holds it against me, and never has to work with me again tomorrow.

So I went looking

The individual parts of this are surprisingly well documented — though nobody has proven yet that frequent AI users get angrier over the months. But the pieces are there.

One study of 30,000 comparable conversations found that people use far less of their usual politeness with AI and settle into a short, commanding tone.

Another finding stayed with me: When researchers simply told people that a real human might read along, the politeness came back.

The brake is never technical — the restraint we carry runs on knowing that somebody is watching, and when you take the witness away, it quietly goes too.

So, how do you stay calm?

Let me come back to where I started, to that afternoon when I was close to shouting at my screen. Long, winding conversations can get less reliable as instructions, corrections, and side-points pile up.

The model can lose the sense of what matters most or reopen something we had already settled. My anger was mostly a signal that the conversation had wandered too far, and not proof that the model had suddenly turned stupid.

Staying calm is not really about being more patient. When the conversation starts to slip, I now ask the model to summarize four things: the goal, the fixed facts, the decisions we already made, and the question still open. Then I correct that summary, move it into a fresh conversation, and carry on from there with the noise left behind.

And it keeps me from looking silly. There is something a bit absurd about typing “I ALREADY TOLD YOU THIS” in capitals to a machine that feels nothing anyway.

Because in the end, the machine was never the thing I was angry at. It was the friction. And friction has a boring, effective answer: You start again cleanly, instead of raising your voice at a screen that cannot hear you.

* * *

If you’re seeing this differently, reply. I’d like to hear it. Reply to this email. Or reach out on LinkedIn.

PLOS ONE

A comparative analysis of help-seeking behaviors in human communities and generative AI

A study of 30,000 conversations found that our language changes more than we'd expect when we talk to AI. People drop nearly all the politeness markers they’d use with a person, and default to blunt commands instead.

Full study →

Information Systems Research

The impact of disclosing human involvement on customer interactions with hybrid service agents

Researchers found that politeness comes back the moment people think a human might be reading along. The restraint we show each other, it turns out, isn’t really about kindness — it’s about being watched.

Full study →

arXiv

LLMs get lost in multi-turn conversation

A large-scale study of top AI models found that losing the thread over long conversations isn't a minor glitch: performance drops 39% on average, and once a model takes a wrong turn, it rarely finds its way back.

Full study →

Did someone forward you this newsletter? Subscribe to The Workplace Rewrite to receive the next issue!