Real competitive advantage with AI lies not in generating information but in knowing when to question it. AI has made “skeptical intelligence” a valuable - and rare - skill.

As AI models grow, so do the theories on how to use them. Yet anyone who has spent enough time working with AI has likely noticed something curious. Give the same model access to the same information and ask ten different professionals to solve the same problem, and the results will often look surprisingly similar. The language may change. The format may change. The recommendations may change slightly. But the thinking tends to converge.

Yet there is always someone who comes back with an idea or output nobody else saw. Did they discover a secret prompt? Probably not. The difference lies in their willingness to challenge the answer given. They ask again and again. They check one more assumption and push beyond the first plausible answer.

Alongside researchers Emi Calicbete , Andrew Kruger , and Priyanka Shrivastava , we hear about the costs of these failures often: documents, reports, and spreadsheets generated with AI being passed off or released while riddled with errors and hallucinations; sources that never existed; quotes made up out of thin air; and balance sheet numbers that did not add up. Whether it is something as seemingly minor as a junior associate failing to fact-check work before sending it to higher-ups, or as consequential as a prestigious consulting firm being forced to refund payment after submitting a report that featured upwards of 20 fabricated sources, these failures are not the result of people lacking intelligence. They reflect our lack of a systematic way to verify AI-generated information.

As AI continues to evolve, the skill of critical thinking is at a premium. At Davos in 2024, IBM CEO Arvind Krishna argued that as generative AI takes over the "lower half of cognitive work," critical thinking becomes far more important across professions. OpenAI CEO Sam Altman has similarly warned that people place too much trust in ChatGPT despite its tendency to hallucinate. Demis Hassabis has made the same point more bluntly, cautioning that using AI lazily can actually make people worse at critical thinking.

The message extends well beyond technology companies. The World Economic Forum has ranked analytical and creative thinking among the most sought-after skills in the labor market, while universities have shifted from restricting AI use to teaching students how to evaluate AI-generated information critically.

The message is remarkably consistent across business, technology, and education. But before accepting that conclusion, we need to answer a much simpler question.

What exactly is critical thinking?

The canonical definition of critical thinking comes from Facione’s 1990 Delphi Report, which defines it as "purposeful, self-regulatory judgment which results in interpretation, analysis, evaluation, and inference." In layman’s terms, critical thinking is the ability to think independently by questioning claims, weighing evidence, and making reasoned judgments. Classical notions of critical thinking emerged in Socratic and Cartesian traditions as methods of forming, testing, and potentially refuting hypotheses based on reasoned judgment and evidence. For decades, this definition served education, research, and professional decision-making remarkably well.

Over time, however, the construct of critical thinking has expanded to encompass personal development, emotional reflection, and dispositional traits such as inquisitiveness, self-confidence, and maturity. This broadening has blurred conceptual boundaries between cognition and affect, complicating measurement and intervention design and making it harder to isolate the specific cognitive skills required to reason well with AI. When a definition has grown to become a catch-all, it loses its teeth. You cannot train what cannot be defined, let alone put it in a workflow. This becomes particularly problematic in the context of generative AI, where the interaction between people and information has fundamentally changed.

The reason critical thinking fails with AI is not technological. It is conceptual. Critical thinking has been asked to mean too much for too long, and now, when professionals need a sharp tool to deploy against a fluent, confident-sounding statistical average, what they reach for is amorphous.

Critical thinking is based on the need for a deliberate, slow, self-regulatory pause. Facione’s "self-regulatory judgment" and Dewey’s notion of reflective thinking assume that people have the time and disposition to step back, examine evidence, and reconsider their conclusions.

AI inverts those conditions. Outputs arrive instantly, in polished language and with remarkable confidence. The cognitive cost of questioning them becomes increasingly high.

A 2025 global study of more than 48,000 people found that 66% relied on AI output without evaluating its accuracy, while 56% reported making AI-related mistakes at work. Uncritical reliance on AI encourages users to offload reasoning and accept fluent but unexamined outputs, undermining deeper forms of inquiry and judgment. AI’s promise as an innovation catalyst is therefore tightly coupled with the risk of "lazy innovation," where speed and fluency are achieved at the expense of genuine novelty and judgment.

Critical thinking often trains us to interrogate claims made by people. We challenge assumptions, expose fallacies, evaluate arguments, and question evidence. AI does not present us with an opponent. It presents us with fluent assertions. The challenge is no longer winning an argument, but determining whether an answer deserves to be trusted.

What is needed instead is a sequence of steps to verify and question AI outputs in our fast-paced world. In our research, we have called this capability “ skeptical intelligence ”: deliberate rational reasoning applied to the accuracy and usefulness of incoming information, stripped of the emotional and self-reflective baggage critical thinking accumulates.

In practice, it runs through eight steps—SKEPTICS, if you need the mnemonic:

  • Solve the right problem. State what you are actually deciding, and why—push past the first version of the question to what is really underneath it.
  • Know what you know. Write down your own preliminary answer before you look at the model's, so your judgment isn't contaminated by its language.
  • Evidence, traced. Trace the sources behind the AI's answer: how current they are, what perspectives are missing, where the answer breaks down, and for whom it would differ.
  • Phrasings, fresh. Rerun the question with the burden of proof reversed, from a hostile skeptic's framing, and as its literal opposite. See if the answer survives.
  • Tear it down. Premortem it: how does this fail badly within a year? What question isn't being asked that should be?
  • Independent opinion. Take it to a different model or a human expert. AI critiquing AI is not independence—shared training data means shared blind spots.
  • Conclusion, yours. Map the evidence against the decision you set out to make at step one. You decide. The model doesn't.
  • Suspicion, still. Notice that this entire process still leans on AI to catch AI's mistakes. That dependency should make you uneasy, not comfortable—especially on anything consequential.

For example, before acting on an AI-generated forecast, a professional would state what decision the forecast is meant to inform, write down their own estimate first, trace the sources behind the model's numbers, rerun the question from a skeptic's framing, premortem the forecast's failure modes, run it past a second model or a colleague, and only then decide—independently—what to use.

Critical thinking describes a broad intellectual capability. Skeptical Intelligence turns the behaviors needed to evaluate AI outputs into eight observable steps. Checklists are what work in fast-paced environments.

Skeptical intelligence is not a rejection of AI, nor is it a return to the slow, self-reflective habits critical thinking once demanded. It is a narrower, sharper discipline built for a world where the answer arrives before the doubt does. The professionals who keep finding what others miss are not smarter or better at prompting. They simply refuse to let fluency stand in for accuracy.

As AI absorbs more of the cognitive load that once separated good work from great work, the eight steps of SKEPTICS are what will separate those who use AI from those AI uses. That gap is where competitive advantage now lives.

Sources & Verification Notes

Deloitte / Australia DEWR — Report on the Targeted Compliance Framework for the Department of Employment and Workplace Relations. A$440,000 contract; corrected version published 26 September 2025 with a disclosure that Azure OpenAI GPT-4o was used; final installment refunded. Errors identified by Christopher Rudge, Sydney Law School / deputy director, Sydney Health Law—approximately 20 fabricated references. Fabricated quotation attributed to Federal Court Justice Jennifer Davies in Deanna Amato v Commonwealth (2019), cited to paragraphs absent from the judgment.

EY Canada — "Points of Attack: Uncovering Cyber Threats and Fraud in Loyalty Systems." GPTZero investigation published 14 May 2026 found 16 of 27 citations fabricated, misattributed, or broken, including references to Forbes, McKinsey, Gartner, TechCrunch, and WIRED. Report pulled the same day.

Arvind Krishna — World Economic Forum, Davos, January 2024. Published at weforum.org, "This is the one skill everybody needs in the age of AI."

Sam Altman — OpenAI official podcast, Episode 1, 18 June 2025.

Demis Hassabis — India AI Impact Summit, 19 February 2026. Full quote: "With AI, if you use it in a lazy way, it will make you worse at critical thinking and so on."

Gillespie, N., Lockey, S., et al. (2025) — "Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025." University of Melbourne & KPMG. 48,000+ respondents across 47 countries; 66% rely on AI output without evaluating accuracy; 56% report AI-related mistakes at work.

Lee, H-P., et al. (2025) — "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers." Microsoft Research & Carnegie Mellon University, CHI 2025. N = 319.

Facione, P. A. (1990) — Critical Thinking: A Statement of Expert Consensus for Purposes of Educational Assessment and Instruction (The Delphi Report), p. 3.

Skeptical Intelligence — construct, Skeptical Quotient scale, and the SKEPTICS protocol (Solve–Know–Evidence–Phrasings–Tear down–Independent–Conclusion–Suspicion): Ladd & Shrivastava (2026); and "Skeptical Intelligence at Work: How Self-Esteem and Gender Shape AI-Powered Innovation."