AI beyond the digital elite

If you spend enough time on LinkedIn or X, you could think that the future of work is already decided: roles are disappearing, everyone should care about AI all the time, and what matters now is whether you are fast enough with the latest model and the latest tool.

I do not think this is a serious view of the world.

It reflects the conditions of a very particular bubble: software, startups, venture, and product: people whose work is already highly digital, highly abstract, and unusually easy to mediate through text, code, APIs, and dashboards. In that world, yes, things are changing fast. But that world is not the world. It is one part of it.

The problem is not only exaggeration. The problem is projection.

The privilege of abstraction

A lot of the people shaping the AI narrative work in environments where the raw material of work is already unusually compatible with current tools: writing, coding, presenting, summarising, querying, routing information, generating drafts, prototyping. That is real. But it is not neutral, and even less so universal.

The people speaking the loudest often sit in a very specific slice of the economy, with a very specific relationship to work, leverage, and tooling. Sam Altman recently said he expects the real impact of AI doing jobs to become "palpable in the next few years." Dario Amodei has suggested AI could write nearly all code within six to twelve months. These are not fringe predictions. They come from the people building the systems. And they are addressed, implicitly, to an audience of people whose work already lives in that register.

From there, it becomes very easy to mistake local conditions for universal reality.

Daron Acemoglu, who received the Nobel in economics in 2024, has been saying something different for some time. He estimates that AI will automate roughly 5% of tasks and add around 1% to global GDP over the next decade, with an annual productivity gain of approximately 0.05%. His sharper point is not the number: it is the direction. "We're using it too much for automation and not enough for providing expertise and information to workers," he has argued. That distinction matters, and it is almost absent from the dominant online conversation.

Most people are not waiting for this story

What really annoys me is the lazy jump from "some people can now build more things more quickly" to "everyone should now organise their relationship to work around AI."

Most people couldn't care less about becoming builders.

That does not make them irrelevant, but I think it does make the current narrative way too narrow.

Most people are trying to do their work well, with less friction, less repetition, less admin, fewer broken systems, and better tools that actually help. They still deserve privacy, dignity, and decent working conditions. They still deserve technology that does not make their work more extractive, more monitored, more fragile, or more absurd. And this is not just philosophy. In Europe, some of the first AI Act rules already in force explicitly prohibit certain uses, such as emotion recognition in the workplace and other abusive AI practices.

China went further last week. Two courts, in Hangzhou and Beijing, ruled that companies cannot fire workers simply to replace them with AI. The legal reasoning is specific: under China's Labor Contract Law, termination requires an unforeseeable change in objective circumstances. The courts found that deploying AI is a deliberate strategic choice, not an act of god. Legal scholars emphasised that the costs of technological transformation should not be borne solely by workers. China's State Council amplified the ruling on April 30, the day before Labor Day. That timing was not accidental. China has no intention of slowing AI adoption; it is moving faster than most. But it is also managing 15% urban youth unemployment and an economy under severe domestic pressure. The ruling is as much about social stability as about labor rights. That combination, aggressive AI development alongside explicit worker protection, is one that Western regulators have not yet found a way to hold simultaneously. I plan to write about this in more detail, and about how it compares to the EU AI Act. But for now, it is enough to note that the country currently leading the AI arms race just said that using AI to cut costs at workers' expense is not a legitimate business strategy. And that is worth sitting with.

There is also a dimension of this conversation that is also almost absent from the dominant discourse: who is actually bearing the cost. The ILO found that female-dominated occupations are almost twice as likely to be exposed to generative AI as male-dominated ones. Around 29% of female-dominated roles face significant disruption, against 16% of male-dominated ones. The reasons are structural. Women are more concentrated in clerical, administrative, and service roles: secretaries, payroll clerks, receptionists, and accounting assistants. These are exactly the roles where current AI tools perform best. For example, among approximately 6 million US workers who would find it hardest to recover from AI-related job loss, 86% are women, concentrated in clerical and administrative roles where tasks are highly automatable.

Sadly, this is not an edge case but the shape of the disruption. And you would not know it from most of the conversations happening on X or LinkedIn, because the people having that conversation are not those workers.

This is where I think a lot of the current discourse misses the point. It assumes desire where there is often indifference, and it assumes relevance where there is often a mismatch.

What real work looks like

Take something much less glamorous than a demo: a medical secretary in a busy clinic.

The problem is not that she cannot generate text fast enough. The problem is that she is switching constantly between systems that do not speak well to each other, dealing with incomplete records, handling exceptions, managing urgency, calming anxious people, protecting sensitive information, and trying not to make a mistake that creates delay or risk for someone else.

We can absolutely imagine AI being useful there.

But the useful version is not that she suddenly becomes some kind of AI-native builder. The useful version is less shiny and more serious: less duplication, better triage, fewer avoidable errors, better retrieval, better handoffs, less administrative drag, and maybe a bit more mental space at the end of the day. That is also much closer to the way European institutions currently describe AI in healthcare: easing administrative burden, supporting decisions, and improving workflows, not turning everyone into a prompt engineer.

That is a very different ambition from the one that dominates online discourse.

A demo is not a working condition

To me, this is where a lot of AI discourse becomes super naive.

In tech, we often confuse what can be generated with what can be done. We confuse a good demo with a meaningful change in working conditions.

Yes, an agent can draft, summarise, classify, route, extract, compare, and send.

But in many real settings, that is not the hard part.

The hard part is exceptions.

The hard part is accountability.

The hard part is trust.

The hard part is regulation.

The hard part is that systems are messy, information is incomplete, people are interrupted, and mistakes do not all have the same cost.

Klarna learned this the hard way. In 2024, it fired 700 customer service workers and replaced them with an AI system. By 2026, it was rehiring human agents after repeat contacts jumped, and customer satisfaction deteriorated on complex interactions. The CEO admitted publicly that the strategy had failed. AI replaces tasks more effectively than it replaces jobs. The work that remains after automation, the judgment, the escalation, the context the model cannot hold, turns out to be more valuable than it looked when the decision to cut was made.

The Yale Budget Lab has been tracking displacement carefully. As of early 2026, there is no significant upward trend in AI-driven unemployment even among workers in highly exposed occupations. The noise is loud. The signal, so far, is more modest.

Significant does not mean universal, and visible does not mean representative.

Where AI could matter more

Meanwhile, we risk underinvesting attention in areas where AI could matter a lot, even if they are less compelling to narrate.

What I find more interesting is not the fantasy that everyone becomes a mini founder with a model. It is the possibility that AI helps us do better work in domains that are slower, harder, more constrained, and more consequential: health, science, materials, climate, education, accessibility, and public services.

The AlphaFold Database now provides open access to more than 200 million protein structure predictions, with over 3.4 million users from 190 countries as of early 2026. DeepMind's GNoME work identified 2.2 million new crystal structures, including 380,000 stable materials candidates. The announcements around these projects have not always been free of hype. But the underlying work is serious, and it points toward a different ambition: AI applied to problems that are slow, constrained, and genuinely consequential, not to convenience or spectacle.

Acemoglu makes a related point. The technology has more potential than current deployment suggests, but the industry keeps pointing it in the wrong direction, toward replacing workers rather than making them more capable. That is a choice, not an inevitability.

A different ambition

Of course, it is good that more people can make things, test things, automate parts of their work, and get leverage they did not have before.

What I am saying is that we should stop mistaking this for the whole story.

Many people simply need better tools, better systems, better interfaces, and less administrative stupidity around their work. And some of those people, specifically the ones in roles that are most exposed, most precarious, and least represented in the rooms where AI strategy gets decided, need us to pay more attention to what is actually happening to them, not to what is happening to us.

That requires more humility than the current conversation allows. We have to look beyond the users who look like us. We have to stop assuming that what feels transformative in a software-native environment will map neatly onto the rest of the economy. And we have to distinguish between what is impressive, what is marketable, and what is actually useful.

I suspect that if we did that more often, we would build different products. We would fund different priorities. We would spend less time projecting the worldview of a narrow digital elite onto everyone else, and more time understanding where intelligence can reduce friction without creating new forms of dependency, surveillance, confusion, or exclusion.

That is a lower bar than the one currently being celebrated. It is also a harder one.