AI Beat

AnalysisTechPolicy

When AI writes the code, where does human value move?

Producing a first draft is getting cheaper. Understanding what is needed, choosing what to build and taking responsibility may become more valuable. A view from inside IT on what we will be paid for, how the next generation will learn, and what AI should and shouldn't do in schools and literature.

A hand sketching in a notebook
Photo: Kaizen Nguyễn on Unsplash

Key points

  • AI may change not only how much work there is, but what people are paid for: less for producing the first version, more for judging it, choosing what to build and taking responsibility. That is a forecast, not yet a measured fact.
  • The hardest question is how people will gain experience if AI takes the simple tasks they used to learn on. A Stanford study found employment of 22- to 25-year-olds in the most AI-exposed US jobs fell 16% relative to others.
  • In schools and in literature, AI can supply facts and fluent words. Whether it can carry meaning, and who checks it, is a question for all of us, not only engineers.

I have worked in IT for many years. In my own work I have watched the balance shift: from writing every line, to reviewing, testing and deciding what should be built at all. That does not mean developers have become mere checkers. It means the work is changing, and so is what we are paid for.

This is my view of where that goes. It is one person’s view, from one industry. Where there is evidence, I set it beside my view, including where it disagrees with me.

From writing code to judging results

When AI tools arrived, the first promise was speed: code written in minutes instead of days. Then the code had to work, and someone had to read what the machine wrote, test it and ask whether it did what the client actually needed. That reading and asking turned out to be the expensive part.

For a while everyone wanted prompt writers, people who knew how to ask the machine. The tools increasingly handle prompt-writing themselves. But defining the task clearly still matters.

What has not got cheaper is judgement: knowing the client, the business and the people who will use the thing, and saying “this is not what we need” before it ships. A machine can check that code runs. Whether it was worth building is still a human call, and a human answers for it.

The question I can’t answer: how will juniors become seniors?

I learned my trade on small, simple tasks. Fix this bug. Write this form. Do it again, better. AI can now produce a first version of many such tasks in seconds.

The evidence says this is not only my worry. A Stanford study of US payroll data found that since generative AI spread, employment of 22- to 25-year-olds in the most AI-exposed occupations fell 16% compared with less exposed jobs. Older, experienced workers in the same occupations held steady or grew. The drop was concentrated where AI automates the work rather than assists it. The study shows a correlation in one country’s data, not proof of cause, but the pattern is hard to ignore.

If the first rung of the ladder disappears, who climbs to the second? I don’t think the answer is to ban AI from junior work. But reviewing AI output is not enough either: you can’t spot a wrong solution if you have never built a right one.

So beginners still need small problems they solve themselves, without help. Then they explain their reasoning to someone more experienced, and only then compare their work with what AI produced. Where the two differ, the learning happens. Later they can review AI output alongside a senior and own small decisions early.

That costs companies time they are tempted to save. The ones that save it now may find in five years that they have no seniors left to promote.

Ideas: where the human part is

AI can generate ideas. What interests me is how we choose which ones deserve to become real.

In an experiment published in Science Advances, writers who were given story ideas by AI wrote stories that readers rated more creative and better written, especially the weaker writers. But the AI-assisted stories were more similar to each other. Each one got better; together they got narrower.

So the human question is not who has ideas. It is which ideas deserve our time, whom they serve, and who takes responsibility for making them real. That choice involves values and responsibility as well as calculation.

Where people stay

Look at a taxi. Some people use the driver as a therapist. Most just need to get from A to B. Where talking to a person is rarely needed, I expect machines to take more of the work. That is already happening in transport, through a different technology: robotaxis carry paying passengers in American cities. Generative AI does the same to office work, where the job is mostly text and numbers.

In therapy, the talk is the work. In teaching, above all teaching children, so is the relationship. There I expect people to stay.

The broader employment picture is more complicated. The International Labour Organization counts about one job in four worldwide as exposed to generative AI, 34% in rich countries, with clerical work the most exposed. But only 3.3% of jobs fall into its highest-exposure group, and it expects most jobs to change rather than disappear. Employers surveyed by the World Economic Forum expect 170 million jobs to be created and 92 million displaced by 2030. That forecast covers every force at once, technology, the green transition, ageing populations and economic pressure, not AI alone. In it, clerical and administrative roles shrink the most; farm work, delivery, construction, care and education add the most jobs.

Our own data shows something narrower. Each week we count job ads on Adzuna in six English-speaking countries, the US, UK, Canada, Australia, India and Singapore, and check whether the ad mentions AI, ChatGPT, LLMs, GenAI or Copilot. In the week of 21 September, that was true of 58% of about 100,500 software developer ads, 33% of 45,000 marketing ads and 24% of 26,000 lawyer ads, but only 1.6% of 118,000 teacher ads and 1.7% of 445,000 nursing ads.

That measures what employers ask for, and the language they use. It does not show which jobs are safe from automation. AI skills were rarely mentioned explicitly in the nursing ads we counted. That does not establish how much AI those workplaces already use, or will.

A human premium? Only under conditions

Will anything done by a person cost more? Not everything. Many customers will pay for whatever is reliable and cheap, whoever does it. But human involvement may command a premium where customers value trust, personal attention or clear accountability. For a café it may be the atmosphere. For care, trust. For software, accountability.

In IT I expect pricing to move that way. Code generated by AI will be cheap. Work for a bank, a hospital or a client with very specific needs will carry more human time, and a human signature that someone answers for. This is my forecast, not a measured fact. I have not seen a study that proves it yet.

The same shift changes what a software company owns. The valuable thing is moving from the code to what it is made from: specifications, tests, data, and the knowledge of what the client actually needs. Some code can be regenerated quickly. Recovering the reasoning behind it is harder.

Learning: facts, meaning, and who checks

I am from Kharkiv. I know what it is to read the news about a city you love.

AI can tell a child what happened in a war: the dates, the places, the civilians killed. It can do it clearly and patiently, as many times as the child needs. It can also get it wrong, and choose facts and words in ways that shape what the child feels. So I don’t believe in a clean division where the machine supplies facts and the teacher supplies meaning. Which facts, and how they are told, is already meaning.

The better question is how to use AI in learning while keeping three things: checking sources, real conversation, and a teacher who is responsible. A child learns from a teacher’s face that this was wrong, that it hurt real people, that it must not happen again. That part passes from person to person. Facts without feeling teach children to know. Facts with feeling teach them to care.

The rules are moving in that direction. UNESCO’s 2023 guidance on generative AI in education asks governments to protect pupils’ data and to consider an age limit of 13 for using these tools in class. In the EU, the AI Act treats AI that decides on admission to education or assesses pupils as high-risk, which brings a legal requirement for human oversight.

Literature: does it matter who felt it?

Like a scratch on the pure white lampshade — the memory of my wandering years will not wipe away.

Ishikawa Takuboku, “A Handful of Sand” (1910). Translated from the Japanese by AI Beat. Tamae K. Prindle’s translation reads it as “the memory of exile”.

Takuboku wrote those lines more than a century ago, far from home. They still hurt. I have always believed they hurt because someone felt them first.

The evidence makes that belief uncomfortable. In a study published in Scientific Reports, 1,634 people tried to tell poems by ChatGPT from poems by Shakespeare, Byron, Dickinson and others. They did worse than chance. Another group rated the AI poems higher for rhythm and beauty. One of the authors suggested a reason: the models drift towards what is typical, and lose the strangeness that marks great poetry, which non-experts can mistake for incoherence.

So here is the dilemma I can’t settle. If a poem moves you, does it matter that nobody lived it? If it doesn’t matter, what happens to readers who only ever meet the typical, never the strange turn, the scratch on the lampshade? I worry that imagination, fed on the likely, learns to expect the likely. I would like to hear from people who disagree.

Where we draw the line

This is a conversation for all of us, not only engineers.

AI as an assistant on logical, checkable work: yes. Use AI where the result can be checked, while preserving the practice people need to learn.

AI making decisions beyond the facts: who gets a loan, which child is gifted, who is guilty, which target to strike. There I believe a human must decide, and a human must answer for it. A machine can lay out the facts. Deciding what they mean for a person’s life should stay ours.

What to do now

  1. Move towards judgement. Whatever your job, the safer part is deciding what is right, not producing the first draft.
  2. If you lead people, protect junior work. Give beginners real decisions to own, with AI as a tool, not a replacement.
  3. Make the human part of your work visible where clients value it: trust, attention, accountability.
  4. Watch the ads, not the headlines. Our AI Demand Index shows each week where employers ask for AI, by profession and country.

I learned by doing work that AI can now do for a beginner. The question is how we give the next generation the same chance to learn.

Tell me what you think

This is not a finished argument. It is the start of one, and it matters too much to leave to engineers and investors.

  • Write to me at hello@aibeatnews.com, or leave a comment below. What work helped you learn your profession, and how will the next generation get that experience? Where do you draw the line? I read every message, and the best answers will go into the next article.
  • Share this article with friends, colleagues, teachers and parents. The more people think about it now, the better the rules we end up with.

Comments

What do you think? No account needed. Please be kind; comments that attack people are removed.

We store only your name and comment. See our privacy policy.

    Sources

    1. Stanford Digital Economy Lab: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (revised 13 November 2025)
    2. ILO and NASK: Generative AI and Jobs: A Refined Global Index of Occupational Exposure, Working Paper 140 (20 May 2025)
    3. World Economic Forum: Future of Jobs Report 2025, press release (8 January 2025)
    4. World Economic Forum: Future of Jobs Report 2025, jobs outlook
    5. Doshi and Hauser: Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances (July 2024)
    6. Porter and Machery: AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably, Scientific Reports (14 November 2024)
    7. UN News: UNESCO calls for regulations on AI use in schools (September 2023)
    8. Tamae K. Prindle (trans.): On Knowing Oneself Too Well: Selected Poems of Ishikawa Takuboku, Syllabic Press, 2010 (ISBN 978-0-615-34562-8)
    9. EU Artificial Intelligence Act, Regulation (EU) 2024/1689
    10. Ishikawa Takuboku: Ichiaku no suna (A Handful of Sand), 1910, full Japanese text (Aozora Bunko)
    11. AI Beat: AI Demand Index, week of 21 September 2026