AI Is Making Us More Capable. Is It Making Us Less Capable Too?

What better AI could do to human productivity, learning, creativity, judgment, and work

AI already writes reports, explains code, summarizes meetings, generates designs, answers customers, and turns rough ideas into working prototypes. The systems are getting better quickly. They are also becoming easier to trust because their answers sound more polished, more confident, and more human.

That combination is powerful. It is also where the danger begins.

I do not think AI is simply making people lazy. That word is too easy and too moralistic. The more interesting change is that AI is altering which mental muscles we use, which ones we neglect, and what employers expect a normal person to produce in a day.

Used well, AI can remove repetitive work and give people more time for judgment, creativity, and relationships. Used carelessly, it can create a person who produces more but understands less. Both outcomes are already visible in the research.

My view is simple:

AI lowers the cost of producing an answer. It does not lower the cost of understanding the consequences.

As AI improves, that difference will matter more.

The productivity gains are real

It would be a mistake to dismiss AI as a toy or a shortcut for people who do not want to work. Controlled experiments and workplace studies have found meaningful gains.

In a study of 453 college-educated professionals doing writing tasks, ChatGPT reduced completion time by 40% and raised independently judged quality by 18%. The largest gains went to people who had performed worse without AI.^[1]

A larger study of 5,172 customer-support agents found that an AI assistant increased issues resolved per hour by about 15%. Less-experienced and lower-performing workers gained the most, while the strongest workers saw smaller benefits. The tool appeared to spread some of the habits and knowledge of top performers to newer employees.^[2]

Consultants in a field experiment also completed suitable tasks faster and produced better-rated work with GPT-4. But the same experiment exposed an important weakness. On a problem deliberately chosen because the model handled it poorly, AI users were 19 percentage points less likely to reach the correct answer.^[3]

AI can therefore raise the floor. It can help an average worker draft, analyze, translate, or organize work at a level that once required more experience. But it does not raise the ceiling equally on every task. Sometimes it leads the worker confidently in the wrong direction.

There are also signs that AI can return time to workers rather than simply demanding more output. In a six-month experiment involving 7,137 knowledge workers at 66 companies, people given Microsoft 365 Copilot spent about two fewer hours per week on email during the latter half of the study and worked less outside normal hours. The experiment found little change in the overall mix of tasks, but the reduction in after-hours work matters.^[4]

That is one possible future: less time cleaning up email, formatting documents, searching through meetings, and repeating administrative work.

There is another possible future too. Once a company learns that a report can be written in half the time, it may not give the employee the afternoon back. It may ask for twice as many reports.

Productivity and human wellbeing are not the same measurement.

Assisted performance is not the same as human ability

The question "Did the person produce a better result with AI?" is different from "Did the person become better?"

This distinction is easy to miss because the finished output may look excellent. A student submits a correct answer. A junior developer produces working code. An employee sends a polished analysis. The tool-assisted performance is visible. What the person can do when the tool is unavailable, wrong, or misleading is harder to see.

A large classroom experiment in high-school mathematics made the difference clear. Students using a standard GPT-4 interface performed much better during AI-assisted practice. But when the AI was removed, they scored 17% lower than students who had practiced without it. A guarded AI tutor that offered structured hints rather than simply supplying answers largely avoided this penalty.^[5]

The lesson is not that AI should be banned from education. The lesson is that interface design changes what people learn. An answer machine and a tutor may use the same underlying model but produce very different humans.

A randomized study of software developers found a similar pattern. Participants used AI while learning an unfamiliar programming library, then completed a mastery assessment. The AI-assisted group scored 17 percentage points lower, with the largest deficit appearing in debugging. People who delegated entire tasks learned less. Those who asked conceptual questions and requested explanations preserved more understanding.^[7]

This matters far beyond school. Every workplace has an apprenticeship system, even if nobody calls it that. Junior employees learn by writing the first draft, tracing the bug, sitting through the difficult customer call, comparing alternatives, and being corrected by someone more experienced.

If AI quietly absorbs all the beginner tasks, new workers may appear productive without getting enough practice to become senior workers.

That is the skill-formation problem I worry about most. AI can make the first rung of the career ladder easier to reach while simultaneously weakening the ladder itself.

Cognitive offloading is useful, but it has a price

Humans have always moved thinking outside the brain. We write notes because memory is limited. We use calculators because arithmetic is repetitive. We use maps because navigating every street from memory is unnecessary.

AI is another form of cognitive offloading, but it is unusually broad. A calculator handles calculation. A generative model can handle the search, the explanation, the first draft, the counterargument, the code, and sometimes the decision recommendation. It can offload an entire chain of thought rather than one mechanical step.

A Microsoft Research survey of 319 knowledge workers found that higher confidence in AI was associated with less reported critical thinking. Workers described shifting their effort away from producing material and toward checking, integrating, and supervising AI output. That shift is not automatically bad, but it assumes the person has enough expertise and motivation to perform the checking.^[6]

Another preregistered study found what researchers called a "speedup illusion." Participants expected AI assistance to make simple cognitive tasks much faster, and the work felt easier, but measured completion times were not significantly faster than independent work.^[10]

Ease can be mistaken for efficiency. Fluency can be mistaken for truth. A confident answer can feel like understanding even when the user has not built a mental model of the problem.

This creates an automation paradox. As AI handles more routine work, the remaining human role becomes oversight. But if people stop practicing the underlying task, they become less able to recognize the rare moment when the system fails.

Pilots have discussed versions of this problem for decades. Automation handles normal conditions, leaving the human to intervene during unusual and dangerous conditions, precisely when deep skill matters most. Generative AI could bring the same problem into offices, classrooms, hospitals, law firms, and software teams.

AI may improve individual creativity while making culture more similar

AI can be a useful creative partner. It can give someone a starting point when the blank page feels impossible. It can generate alternatives, challenge an assumption, or help a person express an idea they could not previously execute.

An experiment with 293 writers found that access to GPT-generated story ideas improved ratings of novelty, usefulness, and enjoyment, especially for participants who initially scored lower in creativity. Yet the AI-assisted stories also became more similar to one another.^[8]

This is a strange tradeoff. AI can make each individual output better while making the collection of outputs less diverse.

We can already see how this might happen. Millions of people ask related models for headlines, color palettes, marketing plans, code structures, presentation formats, and social posts. The model draws from recurring patterns and offers statistically likely solutions. Those suggestions are often competent. They are also pulled toward the center.

The danger is not that AI will eliminate creativity. It may increase the number of people who can create. The danger is that the same invisible collaborator will sit beside everyone, gently steering different people toward similar language, aesthetics, and conclusions.

A large meta-analysis of human-AI collaboration reached another sobering result. Human-AI combinations generally outperformed humans working alone, but they often failed to beat whichever was already better: the human or the AI. Negative synergy was especially common in decision tasks, while content-creation tasks were more likely to benefit.^[9]

Putting a person and an AI together does not automatically combine their strengths. Sometimes the human follows a bad suggestion. Sometimes the person overrules a correct one. Good collaboration requires knowing when to rely on the system and when to resist it.

Programming shows both sides of the argument

Software development is one of the clearest laboratories for AI-assisted work because activity can be measured through tasks, pull requests, tests, commits, and releases.

Three randomized company experiments involving 4,867 developers found that access to GitHub Copilot increased completed tasks by about 26%, with larger gains among less-experienced developers.^[11]

That sounds decisive until we look at a different kind of programming work. METR studied 16 experienced open-source developers completing 246 real tasks in mature repositories they knew well. With early-2025 AI tools, the developers took 19% longer. Before the experiment they expected AI to make them faster, and afterward they still believed it had, even though the timing data showed the opposite.^[12]

These findings are not necessarily contradictory. AI may perform well on bounded tasks, boilerplate, unfamiliar APIs, prototypes, tests, and greenfield code. It can struggle when the work depends on years of repository context, implicit design choices, exact quality standards, and knowing which apparently reasonable change will cause trouble elsewhere.

A 2026 study of AI-assisted code found a 30.7% median reduction in initial task time and no systematic downstream maintainability disadvantage when another developer later changed the code. That is encouraging, although it does not prove that all AI-generated systems remain secure or maintainable over years.^[19]

Security deserves separate caution. An earlier controlled study found that participants using an AI coding assistant produced less-secure solutions while becoming more likely to believe their code was secure.^[20]

The distance between "writing code" and "shipping a dependable product" also matters. A 2026 NBER working paper found that newer generations of coding tools were associated with large increases in coding activity, but the effect shrank sharply when researchers looked at projects, releases, and actual marketplace usage rather than commits alone.^[13]

More code is not always more value. It can also mean more code to review, test, secure, understand, and maintain.

For developers, I think the durable advantage will not be typing code faster. AI is already winning that contest. The advantage will be knowing what should be built, recognizing hidden failure modes, designing systems that survive change, and verifying work that looks correct before it causes damage.

Experience becomes the filter.

AI is changing entry-level work before it replaces whole professions

Predictions about mass unemployment remain ahead of the evidence. A Danish study connecting surveys with administrative data found rapid AI adoption and changes in job tasks, but no detectable average effect on earnings or recorded work hours during the first two years after ChatGPT.^[17]

The absence of broad displacement does not mean nothing is happening. Labor markets can change first through hiring. Companies may keep their current employees while recruiting fewer beginners.

A Stanford Digital Economy Lab analysis found no widespread economy-wide displacement, but workers aged 22 to 25 in highly AI-exposed occupations were 19% below the employment path implied by less-exposed peers. The pattern appeared mainly through reduced hiring rather than increased separations. The authors describe the result as descriptive rather than definitive proof that AI caused the decline.^[14]

This early-career question deserves attention. Entry-level workers often perform the exact tasks that AI handles first: basic research, first drafts, simple code, document preparation, scheduling, routine analysis, and customer responses.

If those tasks disappear, businesses may become more efficient today while creating a shortage of experienced people later. Organizations need to redesign junior roles rather than simply deleting them. Beginners still need supervised opportunities to struggle, make decisions, receive feedback, and learn why a correct-looking answer can be wrong.

Global exposure is also uneven. The International Labour Organization estimates that one in four workers worldwide is in an occupation with some generative-AI exposure, but only 3.3% of global employment falls into the highest-exposure category. The ILO expects transformation to be more common than full job automation because most occupations still contain tasks that require people.^[15]

A 2025 OECD survey of more than 5,000 small and medium-sized businesses found that 31% used generative AI. Among adopters, 65% reported improved employee performance, while 83% reported no change in overall staffing need. These are employer-reported perceptions rather than audited causal effects, but they fit the broader picture: work is changing faster than employment totals.^[16]

New U.S. survey evidence describes workplace AI use as broad but shallow. Many occupations now contain people using generative AI, yet adoption within individual tasks and among workers doing similar jobs varies widely.^[18]

The future may not divide neatly between people whose jobs were automated and people whose jobs were untouched. It may divide between workers who know how to direct and verify AI, workers who depend on it without understanding it, and workers who never receive the access or training needed to benefit.

What happens as AI gets much better?

Some current weaknesses will shrink. Models will make fewer obvious errors, retain more context, operate software directly, and complete longer projects. This will make AI more useful. It may also make humans less likely to question it.

Bad AI invites skepticism. Very good AI earns trust. Near-perfect AI may create the most dangerous failures because the user has learned that checking is usually unnecessary.

As capability rises, I expect five human changes.

First, production will become cheaper. More people will be able to create software, videos, reports, lessons, businesses, and research summaries. Ideas that once required a team may be tested by one person.

Second, judgment will become more valuable. When production is abundant, deciding what deserves to exist becomes the scarce skill. Clear goals, taste, ethics, prioritization, and knowledge of real users will matter more.

Third, unaided skill will become less visible. A person may produce expert-looking work without expert-level understanding. Credentials and portfolios will become harder to interpret unless they include evidence of reasoning, verification, and ownership.

Fourth, organizations may capture the time savings. AI could shorten the workweek, reduce administrative overload, and make jobs less exhausting. It could just as easily increase expected output and monitoring. Technology does not decide who receives the benefit. Contracts, management, competition, and labor policy do.

Fifth, inequality may shift rather than disappear. AI sometimes helps weaker performers more, which can narrow gaps within a task. But experienced workers, well-funded companies, and countries with better infrastructure may be more capable of integrating advanced systems safely. People with domain expertise can also judge AI output more effectively than beginners.

The same technology can democratize production and concentrate power at the same time.

How to use AI without giving away your mind

The answer is not to avoid AI. Refusing useful tools will not protect workers or students from a world in which everyone else uses them. The better approach is to separate assistance from surrender.

Use a learning mode and a production mode

When the goal is speed, let AI draft and automate. When the goal is learning, ask it for hints, questions, explanations, and feedback rather than finished answers. Do not confuse completing the assignment with building the skill.

Think before prompting

Write your own position, plan, diagnosis, or design first, even if it is rough. Then compare it with the AI's response. This preserves independent judgment and makes disagreement visible.

Ask for reasons and failure conditions

A useful prompt is not only "give me the answer." Ask what assumptions the answer depends on, what evidence could disprove it, where it is most likely to fail, and what an informed critic would challenge.

Keep some unaided practice

Developers should still debug without an agent sometimes. Students should still solve some problems without a tutor. Writers should still face the blank page. Skills weaken when they are never retrieved.

Verify in the real environment

Run the code. Read the source. Test the edge case. Check the calculation. Ask the affected person. AI output should be treated as a candidate, not as reality.

Measure outcomes, not activity

Do not judge AI adoption by prompts sent, words generated, lines of code, or documents created. Measure errors, rework, customer outcomes, security, retention, cycle time, and whether people can explain what they submitted.

Protect apprenticeship

Companies should not eliminate every junior task simply because AI can perform it. Redesign those tasks around supervised review, real responsibility, and deliberate learning. Today's efficiency should not destroy tomorrow's expertise.

Preserve human decision rights

Workers need the authority to challenge AI recommendations without being punished for slowing down a process. If the system is officially "only an assistant" but employees are expected to follow it, human oversight is theatre.

The question is what we stop practicing

AI will probably make humans more capable as a group. We will produce more, explore more ideas, and cross technical barriers that once excluded millions of people. I find that exciting.

But capability with a tool is not the same as capability inside a person.

If AI becomes our calculator, tutor, editor, programmer, analyst, and adviser all at once, it will shape the habits beneath our work. We may become better at asking, selecting, and supervising. We may become worse at recalling, drafting, debugging, and patiently reasoning through uncertainty. Some of that trade is sensible. Some of it could leave us fragile.

The biggest danger is not that AI makes everyone lazy. It is that we become highly productive while losing the ability to notice what we no longer understand.

The biggest opportunity is not simply faster work. It is using automation to create more room for the parts of human life that should not be automated: judgment, responsibility, curiosity, courage, relationships, and care.

AI will keep getting better. The outcome for humans will depend on whether we use that growing power to extend our minds or to avoid using them.

References

1. Experimental Evidence on the Productivity Effects of Generative AI
2. Generative AI at Work
3. Navigating the Jagged Technological Frontier
4. Shifting Work Patterns with Generative AI
5. Generative AI without Guardrails Can Harm Learning
6. The Impact of Generative AI on Critical Thinking
7. How AI Assistance Impacts the Formation of Coding Skills
8. Generative AI Enhances Individual Creativity but Reduces Collective Diversity
9. When Combinations of Humans and AI Are Useful
10. Cognitive Offloading and the Speedup Illusion
11. Effects of Generative AI on High-Skilled Work
12. METR Experienced Open-Source Developer Productivity Study
13. Writing Code vs. Shipping Code
14. Canaries in the Coal Mine
15. Generative AI and Jobs: A Refined Global Index
16. Generative AI and the SME Workforce
17. Still Waters, Rapid Currents
18. What Work Does Generative AI Do?
19. Echoes of AI: Downstream Effects on Software Maintainability
20. Do Users Write More Insecure Code with AI Assistants?