Critical Thinking Is Not a Universal AI Skill

Before a CHRO, CPO, CLO or transformation leader backs another broad capability push around AI, I would ask a harder question: where, exactly, does the work still require human judgment because the cost of error, discretion or accountability has not moved? My view is narrow. The issue is not whether critical thinking is good in general. It is whether leaders are using a generic skill label to avoid making task-level decisions about work design, evidence standards and human responsibility.

The Nature brief communication is useful for that narrower judgment. It shows that large language models can display variable reasoning behavior and plausible confidence. That does not prove a workforce-wide need for critical-thinking training, and it does not establish productivity, ROI or adoption outcomes. It does, however, reinforce a practical executive point: if AI output enters a decision with material consequences, someone has to be able to judge when the output is wrong, when the case should stop, and who has authority to override or escalate.

That is not mainly a learning-completion issue. It is an operating condition for safe and productive AI use. The executive owner of the work should define it, with HR, L&D and risk functions supporting the design. In hiring, employee relations, performance management, customer decisions, regulated processes or any workflow with managerial discretion, the enterprise question is not whether people took a course on critical thinking. It is whether the workflow now has clear review points, evidence thresholds, decision rights and accountable humans where the model can influence an outcome.

The mistake I would avoid is treating critical thinking as a universal AI-proof capability and then scaling investment before the work has been classified. In lower-risk, routine tasks, AI may improve speed without creating material exposure, and the return on broad judgment training may be weak. In higher-stakes tasks, the problem is usually not a lack of generic thinking skill. It is that the organization has not specified what counts as sufficient evidence, who must review exceptions, what triggers escalation and where human accountability remains non-transferable.

I would start with a simple sequence. Map the task, not the job title. Define the cost of error and the type of discretion involved. Identify where AI output enters the workflow. Then assign review, escalation and override ownership before funding capability at scale. If those choices are unclear, the organization is not ready to declare success on AI enablement, whatever the training completion rate says.

The evidence boundary is straightforward. The verified source is a 2023 Nature brief communication about reasoning biases in large language models. The cited MDPI study remains unverified here. Neither source establishes that critical thinking is becoming more important across occupations, that it is constraining enterprise performance, or that broad training improves outcomes. So I would not use this evidence to justify a universal capability claim. I would use it to pressure-test where human judgment must stay designed into the work.

What would change my view is direct evidence that AI-induced reasoning errors are creating material performance, legal, customer or managerial problems in specific business processes, along with evidence that defined human review and escalation materially improve outcomes. At that point, the case for targeted capability investment becomes much stronger. Until then, I would keep the judgment bounded: decide where human responsibility still sits, and build the operating model there first.