Berlin. Amsterdam. Seville. Three rooms in three weeks, three different professional communities. C-suite executives, chief learning officers, business school leaders. One shared anxiety surfaced in every conversation: AI is drafting the brief, writing the analysis, framing the language and the room repeats back. By the time senior leaders must make the decision, half the decision has already been made in a draft they did not write.
The Microsoft Work Trend Index reported in May 2026 that nearly half of Copilot interactions inside organizations now support analysis, reasoning, decision-making and problem-solving. When AI moves from drafting into the work of judgment itself, the output it produces has already been framed in ways we do not always detect.
Defining skepticism
Skepticism is grounded judgment, not the contrarian’s reflex to object to everything. It is the practical wisdom Aristotle called phronesis, the cultivated ability to see what is right to do in a particular situation and then do it. Many traditions name a close cousin of the same capacity. Confucian thinkers call it zhi and Islamic philosophers ḥikmah, each a practical sensitivity to what is right in changing circumstances. The terms are not identical, but they share family resemblances, pointing to one thing: context-sensitive judgment exercised at the moment of decision. This is the habit every decision-maker already has: asking what matters, what is right, what resonates and what will work.
The new challenge is to sharpen that habit against outputs engineered to disarm it. We all know how easily a report arrives with clean tables and confident phrasing, or a deck whose visuals carry a storyline before anyone has tested it. The speed and beauty lull us into a sense of certainty and compliance. The quality of the language stands in for the quality of the thinking.
Suspicion of AI turns many people off, professors included. The real danger runs the other way. Already in 2024, researchers reported that people followed advice from an AI tool even when it contradicted their own assessment. A 2025 study of knowledge workers found the same: the more confident people were in the AI, the less critical thinking they did. AI has become an oracle we trust more than our own hunches.
The pattern shows most clearly in the young. In a 2025 survey of more than a thousand US teenagers, nearly three in four had used an AI companion, and almost a third found those conversations as satisfying as, or more satisfying than, talking with a real friend. These companions are built to be agreeable, always available and slow to disagree. A teenager who confides in one gets comfort without friction, and friction is where judgment is formed. Professionals are quietly learning the same deference. The conclusion sounds great, the pushback is absent, and the judgment that should be exercised goes unused.
It is high time to become more skeptical in practice. Here is how.
1. Ground every claim before acting on it
Make a habit of reading every AI output the way a careful editor reads a claim. For each substantive point, ask: What is the evidence behind it, what reasoning connects the evidence to the conclusion and where both might be wrong? The framework is Stephen Toulmin’s old model of argument, repurposed as a daily reading practice. The model forces us to map the grounds for a claim and to surface the invisible “because” we leave out of daily speech.
The cleanest test is to read an AI-generated text and underline every claim, where you can place an “!” after it, then ask which of those statements you can source. The gap between what sounds right and what can be sourced is where our hunches and judgment live.
Researchers writing in AI and Ethics stress the importance of exercising our own reasoning fully while paying close attention to the AI’s output. This is the opposite of relying on algorithmic outputs that many professionals are quietly learning to do.
2. Put dissent on a schedule
The skeptical professional is the one who slows the meeting down, asks the awkward question and sounds like a brake on momentum. We know them: the board member asking penetrating questions, the colleague who keeps asking “why.” Over time, the social cost compounds, the questions become blunt and sometimes even dissipate.
The fix is to make doubt a role, not a personality. The simplest version is a short, scheduled dissent window in any meeting where AI output is used to legitimize a decision — five to ten minutes, near the start. The explicit task is to argue against the claims; to surface hunches, to find the strongest available objection and put it on the table so the rest of the conversation must deal with it.
This is the team-level equivalent of grounding a claim. It builds collective skepticism into the operating rhythm, so it does not depend on the courage of the most awkward person in the room.
3. Watch the pattern, not only the output
A careful reader can catch a single claim that drifts. A hundred claims drifting in the same direction cannot, and neither can a thousand small favors that never become a human exchange. The danger of AI at scale lies in the aggregate, not in any single output, and it remains invisible until it has already set in. By the time the drift registers, it has become the house style, and the house no longer sees it.
An Anthropic engineer described one face of this. Work used to run on what the engineer called “a gift economy of small favors between humans.” A request for help with a script created a small debt. The AI is faster and creates zero debt. Each task it absorbs is a bid for human collaboration that never gets made. The loss is quiet, daily and almost impossible to see in any single instance. Mutual generosity built our institutions, and it erodes one unmade request at a time.
The same dynamic runs through the outputs themselves. A chief learning officer at a global listed company told me in early June that her team had built roughly 54,000 AI agents in 6 months. A few thousand produced real simplifications. What occupies her now is the supervisory habit of asking whether the rest are quietly hardening assumptions nobody has revisited in years. No single agent reveals the problem. The pattern across fifty-four thousand does.
What to do this week
In 1979, an IBM training slide carried a sentence that has aged unexpectedly well: “A computer can never be held accountable, therefore a computer must never make a management decision.” The computer still does not make the decision. It writes the brief that shapes one.
Pick the next AI-generated memo, report or slide deck that lands on your desk. Underline and place an exclamation mark after every claim. Mark the ones you can source in the text. Check them. Schedule ten minutes with your team to argue against the claims. The practice is unspectacular. Peter Drucker drew the distinction more than half a century ago between doing things right and doing the right things. The first is now automated. The second is what the discipline of skepticism is for.
Opinions expressed by SmartBrief contributors are their own.
____________________________________
Take advantage of SmartBrief’s FREE email newsletters on leadership and business transformation, among the company’s more than 250 industry-focused newsletters.