Why Empathy Is Your Most Important AI Leadership Skill (And Why That's a Problem) 2
The most important skill in AI leadership isn't technical fluency; it's empathy. Not empathy for the AI, but theory of mind: understanding how LLMs think differently than humans, even when they sound human. Leaders who miss this make predictable, expensive mistakes under pressure.
The most important skill in AI leadership isn't technical fluency; it's empathy. Not empathy for the AI, but theory of mind: understanding how LLMs think differently than humans, even when they sound human. Marcel Samyn, a human-centric AI leadership coach, argues that leaders who miss this make predictable, expensive mistakes under pressure. The paradox? You need to understand something that mimics human thought but operates by completely different rules.
The production story no one talks about
A few weeks ago I'm debugging a critical production issue. The system is down. Customers are locked out. My CEO is pinging me every five minutes.
I open up the LLM I've been using to help write code. I tell it, in detail, what's happening. "Production is down. This is urgent. I need a fix now."
And the output I get back?
Garbage.
Not just wrong, panicked garbage. Suggestions that don't make sense. Code snippets that would make things worse. Confident-sounding nonsense dressed up as solutions.
I realize what's happened: I told the AI I was stressed, and it started producing stressed output.
It wasn't trying to help me solve the problem. It was mirroring my state back at me.
That's the moment I understood: this thing doesn't think like a human, even though it talks like one.
And if I kept treating it like a smart junior engineer who could read between the lines, I was going to keep getting burned.
The thing everyone gets wrong about LLMs
Here's what trips up most leaders:
LLMs sound so human that we assume they think like humans.
They don't.
A human teammate under pressure can:
Read between the lines
Pushback when you're being unclear
Self-correct when the stakes are high
Hold a mental model of what "good work" looks like in your context
An LLM under pressure does none of that.
It pattern-matches the context you give it, including your emotional tone. When you're stressed, it produces stressed output. When you're vague, it's vague. When you're confident, it sounds confident (even when it's wrong).
It has no internal sense of "this is an emergency, I should be extra careful." It has no goals. It doesn't want to help you. It's just predicting the next most likely token based on everything you've said.
That's not a flaw. It's just how the tool works.
But if you don't understand that — if you treat AI like a competent human who can infer intent, you're going to delegate in ways that backfire.
Under Pressure | Human Teammate | LLM |
|---|---|---|
Reads between the lines | Yes | No |
Pushes back when you're unclear | Yes | No |
Self-corrects when stakes are high | Yes | No |
Holds a quality bar for your context | Yes | No |
What leaders get wrong (and what breaks)
Mistake 1: Delegating like it's a human
You say, "Draft a response to this angry customer email. You know our tone."
But it doesn't know your tone. It knows the statistical average of "professional customer service email" across the entire internet. Unless you've been very explicit about your company's voice, it's going to give you bland, corporate pablum.
A human would ask clarifying questions. An LLM just guesses.
Mistake 2: Assuming it understands stakes
You say, "This is for the board deck, make it really polished."
The LLM hears "use fancier words." It doesn't have a mental model of what your board cares about, what level of detail they expect, or what "polished" means in your culture.
A human knows when to double-check. An LLM doesn't.
Mistake 3: Bringing your stress into the prompt
This is the one that got me. When you're under pressure, you communicate differently, shorter, more urgent, less context.
A human teammate compensates for that. They know you're stressed and they ask better questions.
An LLM mirrors it. It gives you the text equivalent of panic: fast, confident-sounding, shallow.
The leaders who thrive with AI do this differently
The ones who succeed?
They have theory of mind for how the tool actually works.
They know:
LLMs don't infer. So they over-clarify. They give examples. They say "polished means X, Y, Z" instead of assuming it knows.
LLMs mirror tone. So when they're stressed, they strip emotion from the prompt. They write like they're briefing a very literal intern who needs everything spelled out.
LLMs don't have quality bars. So they specify what "good" looks like every single time, even when it feels redundant.
In other words: they practice empathy.
Not empathy in the warm-and-fuzzy sense. Empathy as accurately modeling how someone (or something) else thinks.
And here's the kicker, that's a deeply human skill.
The leaders who are best at working with AI aren't the most technical. They're the ones who can step outside their own perspective and ask: what does this tool actually understand from what I just said?
That's theory of mind. And it's one of the things humans are uniquely good at.
What this means if you're leading teams through AI adoption
If you're rolling out AI tools in your org, this changes how you think about training.
You don't just need to teach people how to use the tools.
You need to teach them:
How LLMs think — what they can and can't infer
How to communicate with non-human intelligence — be literal, strip emotion, specify quality
When to trust AI and when not to — it's great at pattern-matching, terrible at judgment calls
Marcel Samyn's Human-Centric AI Leadership Framework starts here: understanding the tool's cognitive model before you integrate it into high-stakes workflows.
Because if your team doesn't understand how AI thinks, they'll delegate to it like it's human. And under pressure, when the stakes are highest, that's when it breaks.
The paradox at the center of this
Here's the thing that keeps me up at night:
The skill you most need to work effectively with AI: theory of mind, empathy for non-human cognition, is something AI itself can't do.
LLMs can't model your thinking. They can't ask, "Wait, what does Marcel actually need here?"
They just reflect what you give them.
Which means the better you get at understanding how they think, the more you realize how deeply human that skill is.
And that's why I believe the future of AI leadership isn't about becoming more technical.
It's about becoming more human.
The leaders who thrive won't be the ones who know the most about transformer architecture or prompt engineering tricks.
They'll be the ones who can hold two models in their head at once:
How I think. How this tool thinks. And how to translate between the two.
That's the work.
A quick self-check
Next time you're using an LLM for something that matters, pause and ask:
Am I communicating like this is a human who can read between the lines?
Am I bringing my emotional state into the prompt?
Have I actually specified what "good" looks like, or am I assuming it knows?
If the answer to any of those is yes, you're setting yourself up for the same mistake I made that day in production.
The fix?
Empathy.
Not for the AI. For how it actually works.
Frequently Asked Questions
Why does an LLM give worse answers when I'm stressed?
Because it mirrors the tone you give it. An LLM pattern-matches everything in your prompt, including your emotional state. When you write short, urgent, low-context messages, it produces fast, confident-sounding, shallow output that matches that energy. It has no internal sense that the stakes are high and it should slow down. A human teammate compensates for your stress by asking better questions. An LLM just reflects it back. The fix is to strip the emotion out of the prompt and write like you're briefing a very literal intern who needs everything spelled out.
How is delegating to AI different from delegating to a junior employee?
A junior employee infers. They read between the lines, push back when your instructions are unclear, and hold a mental model of what good work looks like in your specific context. An LLM does none of that. It does not infer intent, it does not ask clarifying questions when something is ambiguous, and it has no quality bar of its own. If you delegate to it the way you would to a person, assuming it will fill the gaps, it will fill them with the statistical average of the internet rather than what you actually meant.
What does "theory of mind" mean in the context of AI leadership?
Theory of mind is the ability to accurately model how someone else thinks, separate from how you think. Applied to AI, it means understanding the tool's cognitive model before you hand it high-stakes work: knowing what it can and cannot infer, knowing that it mirrors tone, and knowing that it has no goals or quality standards of its own. Leaders who have this skill over-clarify, give examples, and define what "good" looks like every time. Leaders who lack it delegate as if the tool were human, and get burned under pressure.
When should I trust AI output, and when shouldn't I?
Trust it for pattern-matching work: drafting, summarizing, restructuring, generating options from a clear brief. Be skeptical of it for judgment calls, anything that depends on unstated context about your business, your people, or your stakes. The tool is strong where the task is well-specified and the answer lives in patterns it has seen before. It is weak exactly where a good human would pause and say, "Let me double-check this." If the work requires knowing what your board actually cares about or what your company's voice really is, supply that explicitly or verify the output yourself.
How do I write better prompts when the stakes are high?
Do the opposite of what your instincts tell you under pressure. Slow down, add context, and remove urgency from the wording. Spell out what "good" means in concrete terms instead of assuming the tool knows. Give an example of the outcome you want. Specify the audience and the level of detail. The counterintuitive part is that the higher the stakes, the more deliberate and literal your prompt needs to be, even though high stakes are exactly when you feel pushed to be fast and terse.
—
Marcel Samyn is a human-centric AI leadership coach who helps tech leaders integrate AI without losing their humanity. His 12-week Human-Centric AI Leadership Intensive teaches product and engineering leaders how to lead AI transformation in ways that improve both business results and human wellbeing.