On the Art of Doubting
Why uncritical trust in AI systems doesn't make us smarter — it makes us replaceable.
It’s a sentence you hear everywhere now — in meetings, in professional conversations, in academic discussions: “I asked ChatGPT about that, and it said...” What follows is usually a polished, well-structured answer. No hesitation, no uncertainty, no gaps. And the remarkable thing is that the person saying it is often smart, experienced, and well-educated. They don’t doubt. Why would they? The machine has spoken.
What’s happening here is more than a question of efficiency. It is a quiet, creeping event: the voluntary surrender of one’s own judgment.
The Machine That Never Doubts
To understand what’s at stake, it’s worth taking a brief look at the technology — not to alarm, but to see clearly what large language models like Copilot, ChatGPT, or Gemini actually do.
These systems are, in simplified terms, high-dimensional probability machines. They were trained on vast quantities of human text, learning which words, sentences, and patterns of reasoning are likely in which contexts. What they produce is — statistically speaking — the expected: a kind of weighted average of what humans have written on a given topic.
That is not a criticism. It is a description. And this average can be impressively good. For standard tasks, for summaries, for structuring thoughts — excellent. But it has one structural property that should not be overlooked: it never doubts.
An LLM has no uncertainty in the human sense. It always produces an answer, always in the same calm, confident texture — whether it is correct or fundamentally wrong. It has no sleepless nights of deliberation. It has no moments of hesitation. It has no moments in which it says: I don’t know, and that troubles me.
That is precisely what makes it so seductive — and so dangerous when you don’t know it.
Automation Bias: An Old Problem in a New Form
Cognitive psychology has a precise term for what’s happening here: automation bias. This is the tendency to over-trust automated systems and suspend one’s own judgment — even when the system is demonstrably wrong.
Parasuraman and Manzey showed in their influential 2010 study that automation bias does not only affect inexperienced users. On the contrary, the more one trusts a system, the less attention one pays to its outputs. You stop checking. You stop asking. You stop doubting. Parasuraman & Manzey, 2010, Human Factors
The phenomenon was originally observed in pilots who trusted autopilot systems too much, or in physicians who followed diagnostic algorithms even when clinical signs told a different story. Today we see it in offices, newsrooms, universities — wherever people work with AI tools.
But what we are witnessing now goes beyond classical automation bias. It is something subtler: epistemic delegation. You are not merely handing off the work. You are handing off responsibility for the knowledge itself.
The Authority Without a Method
“Copilot said” functions socially the way “according to a study” or “experts say” once did — it is a shift of authority. The sentence ends discussions. It signals: I have done my research. I am on safe ground.
Except that this new authority conceals something crucial: it has no method you can examine. It discloses no error rate. It cannot reliably cite its sources. It has no agenda — but it also has no accountability.
When a study is cited, you can read it, scrutinize its methodology, check its sample size. When an expert speaks, you know their background, their possible interests, their limitations. When Copilot speaks, you receive a smooth, self-assured answer — and you need to already know what you don’t know in order to notice what’s missing.
This is the core of the paradox: you need expertise to evaluate AI outputs — but in using that expertise, you are simultaneously delegating it away.
The Quiet Self-Undermining
There is one more dimension that is rarely spoken aloud — and yet may be the most consequential.
When someone says “ChatGPT said...” in a meeting, they are signaling something they may not intend: I am replaceable. Not because they lack intelligence. But because they have just publicly delegated their own irreplaceability.
What makes a person valuable in a knowledge economy? Not the ability to retrieve information — machines do that better. It is judgment. Experience. The ability to doubt at the right moment, to distrust a source, to discard a conclusion because it feels wrong — and to articulate and defend that feeling.
Those who systematically fail to exercise that ability will lose it. And those who fail to demonstrate it publicly will not be paid for it.
Sapere Aude — Dare to Know
In 1784, Immanuel Kant wrote a sentence in his essay What is Enlightenment? that is as relevant today as it was then: “Sapere aude! Have the courage to use your own understanding!” He described self-incurred immaturity as the condition of preferring to be guided rather than thinking for oneself — out of convenience, out of cowardice, out of habit.
He could not have imagined a better instrument for that immaturity than a system that always answers, is always polite, always well-structured — and never asks back.
Philosophy also knows the concept of epistemic cowardice: the tendency to suspend one’s own judgment in order to avoid conflict, to avoid being wrong, to avoid being vulnerable. AI outputs offer perfect cover for this. You can hide behind them without anyone noticing.
The opposite — epistemic courage — does not mean distrusting AI. It means engaging with it open-eyed. Asking: Is this right? Why? What is missing here? What would I see differently?
The Art of Doubting
Doubt has a poor reputation. It is seen as a weakness, indecision, or an obstacle. In fact, it is the opposite: it is the most active cognitive stance a person can take.
Science lives by doubt. Karl Popper showed that knowledge does not grow through confirmation, but through falsification — through the willingness to let one’s own hypothesis fail. The replication crisis in psychology and the social sciences reminded us painfully of what happens when institutionalized doubt fades: decades of supposed findings collapse.
Doubt is not technophobia. It is not a refusal. It is professionalism.
An Invitation
This article is not a warning against AI. AI tools are powerful, useful, and will permanently change how we work — for the better, if we use them well.
But “using them well” does not mean treating them as oracles. It means seeing them for what they are: highly capable tools that produce the expected with great fluency, but bring neither originality, nor judgment, nor doubt.
What we must bring — and what makes us irreplaceable — is precisely that:
First: Ask why. Not just “What did Copilot say?” but: “Why might this be true? Why might it be wrong?”
Second: Know your own opinion first. Before querying an AI system, formulate — even just for yourself — what you actually think. Then compare. The gap is often more revealing than the answer.
Third: Show your doubt. In meetings, in conversations, in writing. Not as weakness, but as what it is: a mark of competence.
Fourth: Remember what you know. Years of experience, contextual knowledge, the instinct for an industry, a discipline, a person — none of that is in the training data. That is you.
The art of doubting is not a technique. It is a disposition. And it is — right now, at a moment when machines sound ever more convincing — the most human and most valuable capacity we have.
Sapere aude.