AI can weaken judgment when people accept its outputs uncritically. It can also clear cognitive clutter and expand who gets to think, test, and build.
By Renee Cannon, Founder of eunoiaAI
August 27, 2026
![[Hero Alt Text]](https://cdn.prod.website-files.com/6a5e755fad03f8f3b4db484b/6a8f3123f32331d4cfd60362_From%20clutter%20to%20evidence-backed%20choices_ChatGPT%20Image%20Aug%2026%2C%202026%2C%2002_31_29%20PM.png)
Bill Gates’s August 2026 essay on the “turbulent AI era” offers a serious warning. He argues that AI may permanently eliminate some work, proposes setting aside certain roles as “Human Reserved,” and supports taxes on AI tokens and robots to slow displacement and fund worker support. He also cites early research linking heavier AI use with less critical thinking.
Those concerns deserve attention. Yet the same essay describes AI as a possible equalizer: a way for individuals and small businesses to gain capabilities once available only to large staffs, for disabled people to live more independently, and for people to reclaim time consumed by paperwork and bureaucracy. That tension matters because it points to a more precise conclusion than “AI is making us think less.”
AI does not enter an intellectual vacuum. It often reveals and amplifies the habits, incentives, and standards already present.
Before generative AI, people accepted flawed spreadsheets, repeated a consultant’s recommendation, trusted the first search result, deferred to a confident manager, and circulated articles they had not examined closely. None of those tools or authorities guaranteed reflection. The surrounding work either required verification and challenge, or it quietly rewarded speed, agreement, and completion.
The interface has changed. The obligation has not.
Someone who was never expected to question a source may accept an AI response too quickly. Someone practiced in testing assumptions can use the same system to identify counterarguments, compare scenarios, expose missing evidence, and ask what would have to be true for a conclusion to fail. In both cases, AI is an amplifier. It does not independently create the underlying standard.
Cognitive outsourcing occurs when a person delegates the judgment itself: the problem framing, the interpretation, the choice, or the accountability. The output is accepted because it is polished, convenient, or agreeable.
Cognitive leverage is different. It uses AI to reduce mechanical effort while the person retains the consequential work. A leader might use AI to organize interview notes, compare policy language, generate alternative explanations, or turn a rough idea into a structure that can be inspected. The person still decides what matters, checks the evidence, considers the affected people, and owns the result.
A 2025 Carnegie Mellon and Microsoft Research study surveyed 319 knowledge workers about 936 examples of AI-assisted work. Higher confidence in AI was associated with less reported critical thinking, which is a legitimate warning. But the study also found that AI shifted critical work toward verification, response integration, and task stewardship. Its authors explicitly noted the limits of self-reported effort and called for task-based and longitudinal research.
That is not evidence that nothing is changing. It is evidence that the change is more complicated than intellectual decline.
Administrative load consumes attention. Searching across documents, reformatting information, transcribing notes, drafting routine language, and reconstructing context after interruptions are cognitively expensive even when they are not the most valuable use of a person’s judgment.
In a preregistered experiment with 444 professionals completing realistic writing tasks, MIT researchers Shakked Noy and Whitney Zhang found that access to ChatGPT reduced completion time by 37 percent and improved evaluator-rated quality. Participants with lower initial scores benefited more, narrowing the performance gap.
The experiment was narrow, and it contained its own caution: many participants accepted the initial AI output with little or no editing. Still, the result shows why “less effort” cannot automatically be translated into “less thought.” Sometimes the tool removes low-value friction. Sometimes it removes productive struggle. The design of the task, the user’s capability, and the required review determine which one occurred.
Gates is right to ask who benefits from AI and who bears the disruption. That same equity analysis should apply to proposals that tax tokens, bots, or automated labor.
A large enterprise may absorb a per-token cost, negotiate volume pricing, or build proprietary infrastructure. An independent innovator, a small organization, or a person using AI as an accessibility aid may experience the same tax as a meaningful barrier. That outcome is not inevitable, but it is plausible enough that policy design should examine it directly.
Before adopting a token tax, policymakers would need to ask who ultimately pays, whether accessibility and public-interest uses are exempt, whether small users receive credits or thresholds, and whether the policy distinguishes augmentation from direct labor substitution. A policy intended to reduce inequality should not quietly make affordable cognitive leverage a luxury good.
Overreliance is real. AI systems can fabricate information, hide uncertainty behind fluent language, and validate a user’s preferred view. A 2026 Stanford study of 11 leading systems found broad patterns of overly agreeable behavior in interpersonal advice. Users exposed to more affirming responses became more convinced they were right and less willing to repair relationships.
The answer is not generalized fear of access. It is work designed to require evidence and preserve accountability. Depending on the use case, that can include:
The NIST AI Risk Management Framework treats governance, context, measurement, and ongoing management as connected responsibilities. That is a stronger foundation than assuming either that access will make everyone smarter or that restriction will preserve everyone’s mind.
The useful question is not whether a person used AI. It is whether the work required them to define the problem, inspect the evidence, test alternatives, recognize uncertainty, and remain accountable for the consequence.
Organizations can preserve those habits by designing them into workflows, training, review points, incentives, and leadership expectations. Individuals can use AI as a challenger, translator, organizer, and source of productive friction rather than as an authority that closes the question.
AI can make weak thinking faster and more convincing. It can also give a thoughtful person more room to think, more perspectives to examine, and more capacity to act on an original idea.
The technology matters. The standard surrounding it matters more.
Use the eunoiaAI AI Decision Brief to clarify where AI can reduce friction, where human judgment remains essential, and what evidence the decision requires.
As AI lowers the cost of prototypes and agents, the enterprise bottleneck moves upstream: deciding which workflows merit AI, what authority systems should hold, and what evidence justifies scale.
AI adoption does not begin with access. Employees need to understand where AI fits, what remains human-led, what authority it has and who is accountable when something goes wrong.
Building an AI agent can look straightforward online. The harder questions begin when that agent enters a real workflow with people, permissions, exceptions, and consequences.