If AI Does the Work, What Does the Work Do to Us?

Lately, I have been noticing something in conversations. The answers are polished. There is a neat explanation, a sensible recommendation, sometimes even a framework. Yet I find myself wondering how much of it the person has actually thought through.

I hear AI’s answer. I am less sure where their understanding begins.

My first reaction was that people were becoming worse at thinking. But I use AI to build, explore, and write too. Before judging anyone else, I have to ask myself: when I finish something with its help, what have I become better at?

Writing makes this question difficult to avoid. Sometimes a paragraph will not come together because the thought behind it is incomplete. I move a sentence, look for an example, and discover that what sounded convincing in my head does not quite hold up.

By the time the paragraph works, something has changed in me as well.

A programmer tracing a stubborn bug goes through something similar. So does a musician practising a difficult passage. The work leaves behind both a result and a person better able to understand what they are doing.

Work has two outcomes: what we produce, and what we become capable of through producing it.

AI makes it easier to obtain one without necessarily developing the other. Experiments in coding, reasoning, and reading have found that assistance can leave people less prepared to work independently afterwards. Some also find that people give up more readily once the help disappears. These are results from particular tasks, not proof of a lasting decline in intelligence. But they make me question whether finishing more always means learning more.¹²

We often say people will move on to supervising AI. Yet an experienced engineer’s judgment contains years of errors, surprises, and corrections. A beginner can receive the same answer without receiving that history. How will they learn what deserves a second look?

This gap between receiving and understanding brings me to an older conversation.

In the Brihadaranyaka Upanishad, Yajnavalkya prepares to leave household life and discusses dividing his wealth. Maitreyi asks whether wealth could make her immortal. When he says it cannot, she asks him to teach her what he knows.

The teaching that follows connects hearing, reflection, and meditation with realisation of the Self: shravana, manana, and nididhyasana.³

This is a spiritual inquiry. The connection I draw to our present situation is simply that receiving a teaching does not exhaust our engagement with it. There is still something for the learner to do.

That feels relevant when explanations arrive faster than we can examine them. We have always learned from others. But whether the words come from a revered text or a chatbot, repeating them can leave our understanding untouched.

Perhaps we need to become more comfortable with the time between hearing an answer and knowing what to make of it.

That time lets us notice a missing assumption or ask a better question. It also lets us return to the world outside the explanation. In product work, for instance, a convincing account of customer behaviour still needs a conversation with a customer. First-principles thinking requires us to separate assumptions from evidence and test what follows.

AI can help us investigate. But if we ask it to explain, challenge, and confirm everything, we can end up with a complete conversation that has never met reality.

Making room for that encounter is harder when speed becomes everyone’s expectation. Imagine a team where three finished proposals look more valuable than a morning spent checking whether the problem is real. Even a thoughtful employee will feel the pressure to produce.

Philosophical research argues that losing skills can be a problem of the environments we create.⁴ Computational modelling explores how shared habits could reinforce dependence and make it harder to reverse.⁵ These offer possible explanations, not an inevitable future. They do suggest that telling individuals to “think for themselves” is inadequate if we keep removing opportunities to practise.

And when everyone reaches for similar answers, the group may lose something too. A creative-writing experiment found that AI-generated ideas could improve individual results while reducing their collective variety. Refining people’s own ideas with AI preserved more diversity.⁶

I wonder what that means for thoughts shaped by different languages and lives. An awkward expression may hold an unfamiliar insight. We should give it a chance to develop before replacing it with something smoother.

Still, I would not want to preserve every difficulty. Repetitive work can leave us too tired to think. AI can make learning possible where language, cost, or lack of guidance once stood in the way.

It can also help us grow. In one experiment, people who formed their own ideas before developing them with AI later did better on an independent creativity task than those who used AI freely or worked alone. The benefit was measured over a short period, but it points towards a useful practice.⁷

In areas where I want to improve, I can make a first attempt, then invite help. Afterwards, I can ask whether I can explain the idea, use it elsewhere, or recognise where it fails. Every routine task need not become a lesson.

Schools and workplaces can make the same choice at a larger scale. Time saved could give beginners more guidance, more experiments, and room to investigate mistakes. It could also become an expectation to produce even more.

As work requires less of our participation, our development may need to become more deliberate.

That is what I want to listen for in those conversations: a person who can engage with an answer, bring experience to it, and change their mind for a reason. I owe this essay the same effort.

When AI gives us the answer, what will we choose to do next?

---

References

1. How AI Impacts Skill Formation, Shen and Tamkin (2026, preprint)
2. AI Assistance Reduces Persistence and Hurts Independent Performance, Liu and colleagues (2026, preprint)
3. Brihadaranyaka Upanishad 2.4.1–5, with Shankara’s commentary, translated by Swami Madhavananda
4. AI Deskilling Is a Structural Problem, Ferdman (online 2025; 2026 issue)
5. Large-Language Models as a Cognitive Virus, Solé and colleagues (2026, theoretical preprint)
6. Human Diversity Fuels Collective Creativity That Large Language Models Cannot Simulate or Sustain, Dong and Yakura (2026, preprint)
7. Think First, ChatGPT Later, Wong and Qiu (2026)