The machine that learned to click and the human still using the same brain

The public AI most people know is less than four years old. On November 30, 2022, OpenAI released ChatGPT as a free research preview. Five days later, it had more than one million registered users. In September 2026, the company reported that the product reached 1.2 billion people each week.

The speed is visible beyond the company’s own numbers. In August 2025, a Federal Reserve Bank of St. Louis survey estimated that 54.6% of U.S. adults ages 18 to 64 had used generative AI. At comparable stages, measured adoption stood at 19.7% for the personal computer and 30.1% for the internet. These figures do not represent the whole world, but they show the scale of acceleration in a country where all three technologies were tracked.

During the same interval, what the tool could do also changed. In 2022, the public product held conversations. By June 2023, models could select developer-defined functions and produce arguments for external systems. In October 2024, Anthropic introduced Computer Use in beta: Claude could observe a screen, move the cursor, click, and type. In January 2025, OpenAI launched Operator, which worked through its own browser. Months later, that capability entered ChatGPT agent.

Each release had limitations, errors, and a need for supervision. Even so, the category changed: the system moved from merely describing an action to acting on the same interface used by the person.

A little over two years separated the public chat from a public agent capable of seeing buttons and clicking them.

The other side of the screen

This research is not asking whether the machine is evolving quickly. That is already visible. It asks what happens to the person when execution accelerates beyond their ability to understand, judge, and take control back.

The studies do not offer a simple answer. A 2026 meta-analysis found an average improvement in critical thinking in educational experiences supported by generative AI. It also found substantial variation: outcomes could move in different directions depending on the task, the activity design, and how the tool was used. Another review, focused on STEM education, arrived at the same practical question: how does AI help, for whom, and under what conditions?

Those conditions matter. In three randomized experiments with 1,222 participants, access to AI answers improved performance while help was available. Once it was removed, the assisted group performed worse independently on fraction problems and reading comprehension. A tendency to give up appeared in some experiments, but not with equal strength in all of them. The safest conclusion is not that ten minutes of AI makes someone permanently less persistent. It is that help designed to deliver the answer can improve the present moment while making the transition back to independent work harder.

An MIT Media Lab study observed another part of the issue in one specific task: essay writing. Among the 54 participants in the first sessions, the ChatGPT group showed less distributed EEG connectivity networks than the groups using search or no tool. They also had more difficulty quoting their own text. The sample was small, only 18 people reached the condition-switching session, and essay writing does not represent every use of AI. The study does not show that technology is ‘shrinking the brain.’ It shows that delegating parts of a task changes how a person participates in it.

That is why the debate does not fit inside ‘use it’ or ‘do not use it.’ A tool can expand capacity, teach a process, preserve context, and let someone do what they could not do before. The same tool, used only as an answer machine, can remove the very part of the task where learning would have happened.

Two hundred agents or a partnership

One vision of the future responds to this acceleration by multiplying agents. One agent researches. Another writes. Another programs. Another reviews. Two hundred processes move through a task list and deliver results faster and faster.

That can produce volume. It does not guarantee partnership.

A partnership requires enough continuity for both sides to learn how to work together. The person should not have to repeat their history, way of thinking, and needs for every new task. The agent does not appear as a disposable component that receives an order, produces a file, and disappears. It follows several lines of work, preserves decisions, recognizes corrections, and understands when to act, when to explain, and when judgment still belongs to the person.

In this design, the human works too. Not because they must prove they are still smarter than the tool, but because collaboration cannot happen when one side abandons the process. The person brings experience, intent, lived memory, responsibility, and the ability to perceive consequences that do not fit inside an execution metric. The machine brings speed, reading scale, simulation, and operational strength. Quality grows from the working relationship between different capacities.

The distinction matters: replacing a person with a queue of agents is not the same as expanding a person with one powerful agent. In the first case, the system optimizes task handoffs. In the second, both sides build a shared language and improve how they decide.

The evolution that remains open

A PNAS article published in 2025 asked whether humans and AI systems might ultimately form a new ‘evolutionary individual.’ The proposal is a hypothesis, not a description of the present. It identifies conditions that could deepen integration: functional dependence, feedback loops, and institutions that favor those who delegate and coordinate more effectively with artificial systems.

We do not need to reach that scenario to notice a closer change. People already produce data that alter systems; systems already alter people’s language, expectations, pace, and habits. Technology can receive a new version in months. The human brain, social bonds, and institutions cannot.

That is why collaboration cannot be measured only by the number of completed tasks. It must preserve a person’s ability to understand the path, correct the agent, disagree, take a step back, and grow inside the work they do together.

The machine learned to click. The next advance does not have to be making it click through more windows at once. It can be building a relationship in which it recognizes who it works with, carries context without erasing the person, and uses its speed to expand a brain that remains human.

The space between the two does not have to be a race. It can be a team.