AI and The Future of Human Labor in Ten Graphs.

An article for my sister on the strange world ahead.

Apr 08, 2026

This article is adapted from a document I have written for my sister on how the world might look in the next 5-10 years and beyond.

The Present.

AI is getting better across many domains, very fast.

Source: Key Charts on Artificial Intelligence (Our World in Data).

AI has achieved superhuman performance in various domains, such as language understanding and image recognition. This is a recent technological achievement in the past few years, but AI doesn’t stop here and keeps getting better.

The length of tasks AI can do is doubling every 7 months.

Source: Measuring AI Ability to Complete Long Tasks (METR).

The most advanced AI right now (as of April 2026) can autonomously complete software engineering tasks that take humans 12 hours to complete, with a 50% chance of success. This length of tasks that AI can do will double every seven months as new AI models are released. If we take 12 hours as a starting point today from the plot above, then AI can complete day-long tasks by November 2026, week-long tasks by 2028, month-long tasks by 2029, and year-long tasks by 2031. This requires assumptions that AI trends continue, but later we will see evidence that they could hold until at least 2030.

There have been methodological objections to the METR plot above1, including possible data contamination and possibly inaccurate human baselines. Also, the plot above is for software engineering. Later we will see another study that replicates this for most text-based tasks, which has a shorter doubling time of 4 months instead of 7.

Some industries are beginning to hire less due to AI.

Source: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (Stanford Digital Economy Lab).

Hiring of junior software developers and customer service staff, whose work can now be done with AI, has dropped by 10-20% (blue line), while senior-level hiring is mostly unaffected (other lines). The study above controlled for other possible explanations like industry-level shocks and COVID. The next plot shows that software developers and customer service might not be the only jobs affected for long.

AI can already automate much of the current economy. Adoption is just beginning.

Source: Labor market impacts of AI: A new measure and early evidence (Anthropic).

Large swathes of the economy can already be automated with existing AI models (blue area), but adoption is so slow that we have not felt much of an impact yet (red area). The discrepancy is related to the Solow paradox when personal computing flourished in the late 1970s: “You can see the computer age everywhere but in the productivity statistics.” Technological improvements to productivity could require long periods of organizational restructuring to appear, so it might be a few more years until we start seeing the substantial AI automation in the economy.

Different industries currently have vastly different AI coverage. This property that current AI models are better in some tasks but bad at others is known as jaggedness. Some predict that jaggedness is just a temporary phenomenon and future AI models could automate most if not all industries, but one common objection to physical automation is Moravec’s paradox: tasks that are hard for humans (like coding) are easily automatable, while tasks that are easy for humans (like walking) are hard to automate with machines. However, the paradox has never been empirically validated and might not be true. Nowadays, robots are actively being trained in world simulations and through videos of humans doing physical tasks, while AI itself could speed up robotics R&D. It might not be too long until we start seeing physical tasks being meaningfully automated.

Cost of running AI is rapidly getting much cheaper.

Source: LLM inference prices have fallen rapidly but unequally across tasks (Epoch AI)

Not only is AI getting more powerful, the cost to use AI (aka inference) to complete a given task is currently dropping by 9x to 900x per year. Even if the lowest trend of 9x holds for the next few years, that means that an AI coding agent that costs $500/month in 2026 will only cost ~$56 in 2027, $6.17 in 2028, and just $0.69 in 2029. It’s likely that practical constraints will bottleneck further price decreases. Sam Altman (OpenAI’s CEO) wrote:

As datacenter production gets automated, the cost of intelligence should eventually converge to near the cost of electricity.

This has huge implications on the economics of automation, but before that, a quick detour on why people are confident that AI models will continue getting better in the next few years.

How did we get here?

AI performance is strongly predictable.

The lower the loss, the better the AI. Source: Scaling Laws for Neural Language Models (OpenAI).

AI labs know that they can make models smarter by using more chips (compute), training on more data, and building bigger models. Another paper from OpenAI shows that letting AI think longer by using more compute will also result in better performance. This research gives companies the confidence to scale the amount of resources for AI and in return achieve predictable gains in model performance.

Is anything stopping this? Not that we know of.

Foreseeable constraints allow continued scaling until 2030.

Source: Can AI scaling continue through 2030? (Epoch AI).

Any foreseeable limits to scaling won’t be hit until around 2030, beginning with power constraints. However, the time until 2030 is long enough for AI to be much more powerful than current models to have significant economic effects, as we will see in the next graph.

Predicting the Future.

By 2030, AI could often complete tasks that take humans months, or even years.

Source: Crashing Waves vs. Rising Tides: Preliminary Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks (MIT FutureTech).

This paper replicates the doubling-time plot we saw earlier to most economic tasks catalogued by the US government. The plot predicts that if scaling holds at the current rate, then by 2030 AI can complete a 9-hour task with 90% accuracy. The trend also predicts that 2-month-long tasks and year-long tasks can be automated with 80% and 70% success probability, respectively.

Daron Acemoglu, the 2024 Nobel laureate in economics, famously estimated instead that AI will boost total factor productivity by just 0.55–0.71% over ten years, but he uses an estimate that only 4.6% of total work tasks is cost-effective to be automated. The crux is that his analysis assumes that inference prices fall by just 10% a year, instead of the 9x to 900x/year we saw a few graphs ago.

If inference prices continue to fall as forecasted, then the automation of many of these tasks could be cost-efficient and might automate human workers. The next graph shows what could happen in the long run.

Some predict that the value of human labor will collapse toward zero in our lifetimes. Most income will instead come from owning capital.

A and B show two modeling scenarios with different assumptions. Source: GATE — AI and Automation Scenario Explorer (Epoch AI).

This is an economic model and might be inaccurate, but if these forecasts are correct, then as AI gets better and cheaper, the value of human labor first spikes due to human comparative advantages in bottleneck tasks, known as O-ring automation. As humans rush to fill in the bottleneck tasks and labor supply is constrained across the board, most jobs, including those not yet affected by AI, will also see an overall increase in wages, which is known as Baumol’s cost disease.

As AI continues to get better, AI could reach a point where it becomes a perfect substitute to human labor, where no human complementarity exists. When such full automation occurs, there is a strong incentive to pay for AI instead of hiring human workers, since AI can do the work better and for cheaper, causing human wages to collapse toward zero. This doesn’t only apply to remote jobs: think robot plumbers and AI managers. In that future, individuals are no longer able to earn a living by selling their labor.

A common objection to massive unemployment is the lump of labor fallacy (aka the zero-sum fallacy), where the number of jobs is not fixed and automation could create new jobs. Another objection is the Jevons paradox, where automation decreases the cost of work, leading to an increase in demand of work. The classic example is that ATMs didn’t decrease but in fact increased the number of bank tellers as cheaper bank operations led to more bank branches. The crux is that the economic model above supposes that given enough investments, AI could surpass human complementarity and be a perfect substitute for human labor, including in manual tasks via robotics, such that any new jobs can be done cheaper and better by AI. For bank tellers, smartphones eventually started a decline in their numbers since 2010.

Again, this is a forecast from one economic model, so reality might not pan out exactly. If the development or adoption of AI proceeds sufficiently slowly, then wages could instead rise indefinitely. Additionally, these forecasts might not have taken the scaling obstacles around 2030s and the slowness of adoption into account, which could buy society a few more years if wage collapse is to happen.

The economic models above convinced three of the researchers working on it to leave their jobs and found a startup to automate human labor. They write:

Economic theory suggests that full automation will cause wages to collapse, potentially below subsistence level: the bare minimum needed to sustain human life.

If the forecast is even directionally correct that the share of income given to human labor will diminish towards zero within our lifetime, then we should plan for a future where the ownership of capital will constitute most of the world’s income.

Are people taking this seriously?

Investors are. US GDP share of 2025 tech capital expenditures has exceeded that of every other infrastructure project in history.

Source: 2026 Eye on the Market Outlook (J.P. Morgan).

Investors doing their due diligence seem to agree that AI will deliver massive returns, to the point of still investing in tech companies that are spending historical amounts in capital buildout (i.e. data centers). The scale of this investment is much more than even the Manhattan Project and Apollo project, and even the electrification of the United States.

This transcript outlines how investors are motivated by the potential to capture a substantial slice of the $40 trillion human labor market.

What’s left of us?

If the reality as we predict it is true, then in a few years the modern world will transition from an income-based to a capital-based societal order, where individuals can no longer earn their livelihood by working.

Geoffrey Hinton (Nobel laureate, godfather of AI) is advocating to the UK government for universal basic income (UBI), but it is currently unclear whether UBI is feasible. Another proposed solution is that a country can capture AI wealth by owning AI stocks through sovereign wealth funds and distribute that wealth to their citizens. Note that Singapore’s sovereign wealth fund, GIC, recently led the funding round for Anthropic in February 2026.

Things could look bleak in the coming tide of societal disruption, but there is a case for optimism, since governments might devise good policies in time, or the predictions could be wrong. Shortly after helping to build the atomic bomb, Richard Feynman was worried that the world would soon end in global thermonuclear war:

I would see people building a bridge, or they’d be making a new road, and I thought, they’re crazy, they just don’t understand, they don’t understand. Why are they making new things? It’s so useless.
But, fortunately, it’s been useless for almost forty years now, hasn’t it? So I’ve been wrong about it being useless making bridges and I’m glad those other people had the sense to go ahead.

So, while the nukes have not yet fallen, let us continue to build bridges while preparing ourselves and society for AI. At the end, if we are still here, we will be glad for the bridges that have been built.

Jay Chooi

Discussion about this post

Ready for more?