Social Issues

The value of uncertainty: why AI perfection is very imperfect

Heavy reliance on AI can weaken the everyday practice through which people learn to judge, care, and act, writes Gleb Tsipursky.

A colleague asks a young employee to draft a difficult message to an unhappy customer. The employee opens a chatbot, pastes in the situation, and receives a courteous answer to send out in seconds. The customer gets a prompt response. The employer saves time. Everyone is happy.

But the young employee has missed a small encounter with responsibility. They have not had to imagine how the other person will read the message, or choose which facts need explaining, or decide when an apology becomes a promise the organization cannot keep. Those decisions can still be made after reading an automated draft, but they are easier to overlook when the draft looks polished.

I work with organizations adopting artificial intelligence, and I have become interested in the moments that disappear when a task becomes effortless. We usually count the minutes saved. We rarely measure the opportunities to develop judgment which are lost when tasks become effortless.

This matters because expertise develops through repeated encounters with uncertainty. 

An experienced nurse knows when a routine explanation leaves a patient confused. A manager hears that an employee’s agreement conceals reluctance. A teacher notices the gap between a fluent essay and a student’s understanding. No single example proves the rule. Over time, people build the capacity to recognize exceptions, test assumptions, and accept responsibility for decisions.

Technology has always changed which skills deserve our attention. I am grateful that calculators have relieved people of endless arithmetic and that search engines make information accessible. The right question is which forms of human learning should accompany the new conveniences.

The striking result from an experiment by Shakked Noy and Whitney Zhang, published in Science in 2023, helps explain the appeal of AI assistance. Professionals assigned writing tasks worked substantially faster with ChatGPT and produced higher-quality work on the experiment’s measures. The study demonstrates an immediate productivity benefit. It cannot establish how routinely delegating first drafts affects the judgment people develop over years of professional practice.

Another study points toward a more encouraging possibility. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond examined the introduction of an AI assistant in customer support. Less-experienced workers benefited disproportionately, suggesting that a tool can make useful practices more accessible to people still learning their jobs. The research offers a reason to hope that AI can widen access to expertise. It also gives employers a practical challenge: can they turn a helpful suggestion into something the employee understands and can use independently?

We should make that distinction part of everyday work.

Imagine a supervisor reviewing a customer response drafted with AI. Instead of asking only whether the draft is good enough to send, the supervisor asks the employee what the customer is worried about, which sentence could create a false expectation, and what they would change if the customer replied angrily. The exchange takes two minutes. It turns a convenient draft into an opportunity to practice professional judgment.

The same approach works beyond offices. A volunteer coordinator using AI to plan a rota can check who is being asked to work the least desirable hours. A charity manager using AI to summarize a difficult service-user complaint can read the original account and identify the detail the summary flattened. A teacher can invite students to explain where an automated explanation becomes misleading. In each case, the person using the tool remains involved in the moral and practical consequences of the decision.

This is particularly important for the people entering a profession. Senior employees often possess the experience to spot a plausible error. Junior employees may receive impressive answers from AI before they have learned why one answer is better than another. 

If organizations use AI to eliminate every routine task that once exposed novices to problems, they will need to create deliberate alternatives. 

Otherwise, they may discover that they have built a fast production process with a shrinking supply of people capable of supervising it.

I suggest three simple habits.

First, ask the person using AI to name the decision that remains theirs. This should be concrete. “I am responsible for checking the factual basis of this claim” is better than “I am responsible for quality.” “I must decide whether the customer’s circumstances require an exception” is better than “A human will review the output.” Specific responsibility creates a place for judgment to operate.

Second, reserve occasional tasks for unaided practice. A musician cannot learn to listen by having an instrument play every passage perfectly. A new manager cannot become comfortable with difficult conversations by outsourcing every first attempt at wording. People need opportunities to struggle, explain their reasoning, and receive feedback. The point is not to make work unnecessarily slow. A small amount of intentional practice protects the ability to act when circumstances are unfamiliar or the tool is unavailable.

Third, teach people to question a fluent response. Ask where an answer came from, which assumption carries the argument, whose perspective is missing, and how to tell if the conclusion is wrong. 

Such questions are more valuable than asking employees to memorize a list of clever prompts. They also communicate respect: the worker is trusted to think, disagree, and correct.

These habits raise a deeper question about the kind of workplaces and communities we want. An institution can treat employees as passive recipients of instructions they must execute, even when those instructions arrive in agreeable language. Or it can give them enough understanding and authority to challenge what they are told. Artificial intelligence makes that choice more urgent because it can produce convincing explanations at a scale no human organization could previously manage.

The pressure to accept a plausible answer is real. People have deadlines. Managers reward speed. Students want to finish assignments. Parents are tired. Volunteers have limited hours. It is unreasonable to demand that everyone independently reconstruct every calculation or research finding. 

We need practical boundaries: where speed is welcome, where checking is essential, and where learning must be preserved even when assistance is available.

The most consequential boundary appears when a decision affects someone who cannot easily object. A rejected applicant, a confused patient, a child receiving a judgment about their abilities, or a client whose complaint has been summarized incorrectly may never see the process behind the answer. Someone must still be able to explain it, reconsider it, and take responsibility for changing it.

In my book The Psychology of AI Adoption at Work, I argue that useful adoption depends on the human habits surrounding a technology: trust, training, incentives, and accountability. Those habits are more difficult to purchase than software. They also determine whether a helpful system becomes a source of learning or a reason to stop paying attention.

We will continue handing machines tasks that once consumed human time. Many of those changes will be welcome. The harder work is to decide which encounters with uncertainty we must keep, because they teach us to become people who can judge, care, and act. Saving minutes matters. Preserving the capacity to use those minutes well matters more.

Gleb Tsipursky PhD, is a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).

Leave a Reply

Your email address will not be published. Required fields are marked *