Leadership in the Age of AI

AI is making it remarkably easy to avoid thinking.

That sounds more dramatic than it needs to be. We have always used tools to reduce cognitive effort: calculators replaced mental arithmetic, search engines reduced the need to remember where information lived, and spreadsheets automated calculations that once had to be done manually. Used well, that is exactly what tools are supposed to do. They remove unnecessary effort so that people can spend more of their attention somewhere more valuable.

Generative AI is different in one important respect. It does not only remove the mechanical parts of knowledge work. Increasingly, it can perform parts of the thinking itself: structuring an argument, developing an interpretation, generating options, challenging an idea, synthesizing research, or recommending what to do next.

That creates an extraordinary productivity opportunity, but it also introduces a leadership problem that organizations are only beginning to understand: what happens to human capability when the easiest way to complete a cognitive task is increasingly to avoid doing the cognitive work yourself?

The early research gives us reason to take the question seriously.

A 2025 study from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers across 936 real-world uses of generative AI. The researchers found that greater confidence in AI was associated with less critical thinking, while greater confidence in one's own ability was associated with more. The finding was not that AI eliminated critical thinking altogether; in many cases it changed where that thinking happened, shifting effort toward verification and oversight. But it also showed how easily confidence in the system can reduce the amount of independent cognitive effort people apply to the task.

Experimental research on overreliance points in a similar direction. A 2024 study published in Computers in Human Behavior found that participants sometimes followed AI advice even when it conflicted with contextual information available to them and with their own initial assessment. In other words, the risk is not confined to AI producing the wrong answer. It includes the human receiving that answer becoming less willing, or less able, to challenge it.

There are also early indications that the effect may extend beyond decision-making into how we engage cognitively with the work itself. An MIT Media Lab study published as a 2025 preprint compared participants writing essays with an LLM, a search engine, or without external tools. In that particular task, the LLM group showed the lowest measured neural connectivity, weaker recall of what they had written, and lower reported ownership of their essays. The study was small—54 participants completed its first three sessions, and only 18 completed the fourth—and it should not be treated as evidence that using AI causes broad cognitive decline. But it raises an important question about what happens when cognitive effort is repeatedly outsourced rather than supported.

For leaders, that distinction matters. The immediate danger is probably not a sudden epidemic of people losing their ability to think. The more plausible organizational risk is the gradual accumulation of cognitive debt.

A junior employee who begins every analysis by asking AI for the answer may become very good at producing polished analyses without developing the underlying ability to construct one. A manager who consistently asks an assistant to identify the implications of a report may become faster at processing information while getting less practice forming an independent interpretation. A team that automatically turns to AI whenever it encounters ambiguity may become highly productive when the system is available, but increasingly uncomfortable working through uncertainty without it.

None of those behaviors looks particularly dangerous in isolation. In fact, many will initially look like productivity improvements. The problem becomes visible later, when someone needs to recognize that the AI is wrong, work through a novel situation for which the model has poor context, challenge a plausible recommendation, or operate without the tool at all.

That last point deserves more attention than it currently receives.

Organizations are building working habits around an unusually powerful technology whose long-term economics are still being established. Today's frontier AI experience is supported by extraordinary investment in chips, data centers, model development, and inference infrastructure. Reuters reported in September 2026 that OpenAI expects cumulative cash burn of roughly $278 billion between 2026 and 2030 as it continues investing heavily in compute and infrastructure. At the same time, improvements in model and inference efficiency are pushing some prices sharply downward; OpenAI, for example, says its current model family was explicitly engineered to reduce the cost of delivering frontier intelligence.

Those two realities can coexist. AI may continue becoming cheaper overall while the most capable models, highest compute allowances, agentic workloads, or enterprise features become more tightly metered or differently priced. Competition may push prices down further. New architectures may dramatically reduce inference costs. Providers may also change usage limits, retire models, alter product tiers, or reserve the most compute-intensive capabilities for customers willing to pay for them.

The important leadership point is not to predict which pricing model will win. It is that organizations should not assume today's combination of capability, availability, and price is permanent.

That turns cognitive dependence into an operational question as well as a human one. If an organization redesigns work on the assumption that every employee will always have inexpensive, near-unlimited access to the world's best reasoning models, it is making a dependency decision whether it recognizes it or not.

Imagine taking away your organization's most capable AI tools for a month. Could analysts still conduct an analysis from first principles? Could managers write a coherent recommendation without asking a model to structure it first? Could employees distinguish a weak argument from a strong one without immediately requesting an AI critique? Could teams continue operating effectively if the model they had designed their workflows around suddenly became significantly more expensive, more restricted, or unavailable?

The objective is not to preserve inefficient work for the sake of keeping people's brains busy. There is little value in making employees manually perform tasks that technology can reliably do better. The leadership challenge is deciding which capabilities are safe to outsource and which capabilities the organization still needs its people to possess.

That may mean deliberately designing moments of cognitive friction back into AI-enabled work: asking employees to form an initial view before consulting AI, requiring reasoning rather than only a final recommendation, periodically completing important tasks without assistance, or using AI to challenge someone's thinking rather than replacing the act of thinking in the first place.

It also means changing what AI literacy looks like. Knowing how to prompt a model will increasingly be the easy part. The more durable skill will be knowing when not to accept its answer, when to think independently, when to verify, and when the convenience of delegation is quietly removing a capability the individual or organization still needs.

The organizations that use AI best will not be those that outsource the greatest amount of thinking. They will be the ones that become deliberate about which thinking is worth keeping.

AI can make cognition cheaper.

Leadership has to make sure it does not make judgment scarce.