AI told me to – the most expensive sentence a leader will say in 2026

AI in leadership decision and judgement

For some time now a thought has been coming back to me: artificial intelligence is no longer just a tool for getting work done faster, and is becoming a way to offload the weight of responsibility for what we think and for what we end up doing. I know how tempting that can be, because I feel the pull myself, the urge to set down the burden of my own judgment, even for a moment. The problem is that for some people, that pull grows into something close to a “parent persona”.

It starts innocently enough. We ask the model for knowledge, then for analysis, and eventually for direction. For peace of mind, we add a second model so it can confirm the first. And when both say more or less the same thing, it’s easy to assume the matter is practically settled. Except at that moment we’re no longer just looking for support in our thinking. More and more often, we’re looking for someone we can share the weight with, so that if it goes wrong, we can say: “I worked through this with Claude” or “ChatGPT suggested it to me”, and pin the bad result on the machine.

The machine calculates probability, not truth

This matters, because an AI model doesn’t work like a human who understands something and arrives at the truth step by step. It has been trained on enormous amounts of text, and when it answers, it assembles a sentence piece by piece. It takes what has been written so far, breaks language into small fragments (sometimes whole words, sometimes parts of them), and calculates which fragment has the highest probability of coming next. It then adds that piece to the answer, and does it again. And again. That’s how a whole paragraph comes into being. So artificial intelligence doesn’t always give you a true answer. It gives you the one that is most probable at that moment and sounds convincing enough. It’s closer to probability calculus than to checking whether something matches reality.

I caught myself on something else recently. When I use tools for research or design, and the model starts asking me clarifying questions, more and more often there is also a button labeled “decide for me”. At first it felt convenient. If I don’t yet know what to answer, isn’t it faster to click and let it run? Only, what does that really mean? Which criteria is it picking? Where do those criteria come from? If I myself can’t yet articulate them, will the model really be able to do it? Or will it simply give me what seems most probable, based on the vast amount of text it learned from and what is happening right now in the conversation?

And there’s another layer to this. It isn’t only that we are more willing to hand decisions over to models. The tools themselves are being designed to take more and more effort and responsibility off our shoulders, because then they are used faster and more comfortably. In product world that’s usually a fine direction. But AI isn’t an ordinary tool where you can hand over small choices without much harm. That’s exactly why the line starts to blur so easily. First we delegate trivial things, then slightly harder ones, until we are quietly handing over not convenience, but our own judgment.

This comes to us fairly easily, because AI answers really do sound confident. More than once I have read something thinking, “huh, that was actually laid out sensibly”. But once I asked where the information came from, what justified the choice of data, or which criteria shaped the answer, it stopped sounding nearly as convincing.

When “the algorithm suggested it” stops being a defense

That’s why I find this so important: a recent May ruling from a federal court in New York, in which the court held that hiding behind “the algorithm suggested it” is not a valid defense against an organization’s responsibility for the consequences of its own decisions. The case concerned decision-making by government bodies, but the ruling has broader implications for every organization rolling out generative AI inside its processes.

Because a lot of my time lately goes into understanding what happens to our thinking when we work with AI, I also came across a meta-analysis of thirty-five studies on human and AI collaboration. In one of the studies, across 560 model-assisted assessments, practitioners accepted a recommendation that contradicted their own judgment 67 times, and in 38 of those cases, they overrode their correct assessment with the tool’s wrong suggestion. And Harvard Business School Working Knowledge in a 2025 analysis showed that executives are using algorithms as cover for cuts and restructuring, and that “AI made me do it” is now doing exactly what “McKinsey recommended it” used to do.

Dozens of pages no one actually reads anymore

It doesn’t stop at offloading responsibility either. More and more I see that AI lets leaders produce, in minutes, dozens of pages of analyses, summaries and documents that look serious, are well formatted, and sound sensible, but increasingly carry no real value.

I myself get materials of 30 or 40 pages from people more and more often, and honestly, I find it harder and harder to believe anyone is actually reading it all. A study on AI-generated posts showed that such content tends to be more informational, but also more complex and less personal, which makes a real conversation around it harder. And a newer study on AI-mediated communication revealed something the authors actually called the “AI penalty”: messages produced with AI involvement were perceived as less credible, less authentic, and less worth acting on, also in a work context. So my impression is that in companies we don’t only outsource part of our thinking to models; we also flood ourselves with materials nobody really has the energy to read anymore.

People can tell when a leader is hiding behind it

The trouble is that people see when their leader is hiding behind AI, even when they don’t say it directly, and even when, at first, they can’t put it into words or clear arguments. When a leader regularly leans on AI, especially for hard and uncomfortable decisions, the team starts losing trust in what that leader says in the easier moments too. After all, if you hid behind a tool once, whose words are these really?

From what I see, this isn’t a widespread practice yet, but I notice the temptation growing among the people I talk to and the leaders I advise. Access to models is fairly cheap today, so the threshold for handing over more and more decisions to them is low. There was a time when you had to hire consultants, mentors or experts to help with the process, and they always had some capacity to “push back”. It was also much harder to hide behind them, unnoticed, the way one can hide today behind AI.

What am I really looking for there

I think the only thing that helps here is to pause, the next time I open a model before a hard decision, and ask myself: am I actually looking for a frame for my own thinking, or for someone I can blame in front of myself if it goes wrong? Because the consequences will sit on me anyway, and that move only relieves me psychologically.

And I don’t say this as some abstract metaphor. A 2025 study on human and AI relationships suggests that some people are starting to use models as a kind of secure base and shelter, a place they come back to for support and reassurance. A newer 2026 paper describes such attachment to chatbots along three dimensions: emotional support, distress when contact is missing, and exactly that sense of the AI becoming a kind of anchor. And a longitudinal study on long-term chatbot use showed that people who came back to them more often on their own tended more often to slip into emotional dependence and more problematic use.

It’s worth keeping in mind that the first person to notice when we’re hiding behind a model won’t be us. Most likely it will be someone sitting next to us, who can’t find the words to tell us. And even if they do notice, I don’t believe they will say it outright, that they can see us looking for a frame, or for someone to blame. That’s why self-reflection matters, and why we have to keep searching for our own blind spots in our reasoning and in how we understand what we’re doing.

How to keep your own judgment from AI

These days, before I ask a model for an opinion on any decision, I first write down in two or three sentences what I really think, what I’m afraid of, and what I would be willing to put my name to. Only then do I ask AI to push back, to find gaps in my reasoning, or to point out things I might have missed. That way the model isn’t taking the steering wheel; it’s helping me arrange my own thoughts better. At least in theory, because you always have to stay alert not to lock onto one idea.

I want to be clear: I’m not arguing today whether to use AI (after all, I use it myself, I recommend it, and I’ve even written a book on how to roll it out properly). I’m arguing that at some point we have to stop handing over our own judgment to it.