AI Replacing Humans Was Sci-Fi. Now It Can Do 43% of Modern Jobs, Anthropic Finds

I. How soon will AI replace humans?

I’d like to share two research papers that paint a stark picture of AI’s impact on the job market. The first comes from Stanford economics professor Brynjolfsson, who predicted in 2022 that advanced AI would disrupt the global power balance. That future is already here – Anthropic’s latest research in Feb 2025 shows that nearly half (43.6%) of current AI applications can directly replace human workers.

Diving deeper into the research reveals an even more concerning trend: another 31.3% of jobs fall under “Task Iteration” – where AI and humans temporarily collaborate. Here’s what this means in practice:

Consider a typical workflow: When summarizing meeting notes, I ask AI to extract key points, review them, then request additional details about specific areas. Through this iterative process, I refine the output until it’s ready for my boss. This type of human-AI collaboration won’t last long – as millions of users unknowingly train these systems 24/7, AI will soon master these tasks independently.

II. Which jobs are most at risk?

Let’s break down this chart that shows how AI affects different jobs. We’ll look at two main things: how much people get paid, and how much AI is already being used in their work. Using the median salary ($60,000) and a small amount of AI use (1%) as dividing lines, we can see four main groups:

• A: High-paying jobs that AI can already do – These jobs are the first ones companies want to replace with AI because they’re expensive and AI is already good at doing them. For example: programmers, editors, and writers.
• B: High-paying jobs that AI is learning to do – These are well-paid jobs that AI can’t fully handle yet, like specialized doctors. Right now, AI just helps these workers do their jobs better. For example, Tempus AI helps doctors plan treatments for patients. While they say they’re just helping doctors now, it’s possible that the data collected can be used to train AI to do more of this work autonomously in the future.

• C: Hands-on jobs that pay less – These are jobs AI isn’t good at yet because they need human touch and physical skills, like hairstylists. Right now, these jobs barely use AI at all.
• D: Lower-paying jobs already using AI – These jobs don’t pay much but are already starting to use AI a lot, like tutors and office workers. As AI gets cheaper, these jobs are in danger of being replaced.

As of 2025, China and the US are competing to develop the most powerful AI, similar to how US and USSR raced to reach space first in the 20th century. Many jobs will likely be replaced by AI – it’s like seeing a massive wave forming far away that’s heading toward shore. When it arrives, we might see more people without jobs than ever before in history, which could cause serious economic problems.

III. Why is the Turing test a “trap”?

This whole crisis is connected to the famous Turing test. Here’s why:

In 1950, Alan Turing came up with a game in his paper “Computing Machinery and Intelligence” to test if machines can think:

This game became known as the “Turing Test” and turned into the standard way to judge if AI is advanced.

According to Turing Test, AI (and what we now call AGI) is basically “Human-like Artificial Intelligence” (HLAI) – technology that mimics what humans do. The problem with this approach is:

“When AI copies and automates what humans already do, machines become better replacements for human workers. The replaced workers lose both economic power and political voice. Business owners who get machines that can do human-level work will often just replace humans with those machines.”

But AI doesn’t have to be defined this way. Another approach is:

“When AI enhances what humans can do, helping people accomplish things they never could before, then humans and machines work together. This partnership keeps humans essential to creating value, maintaining their position in both the job market and political system.”

If we just focus on making AI that copies humans, we might get more efficient production, but wealth and power will concentrate in fewer hands. This creates a dangerous situation: people who lose their jobs and influence can’t improve their lives. Professor Brynjolfsson calls this “The Turing Trap” in his paper.

When society falls into the “Turing Trap,” we get massive unemployment and a growing group of zero marginal product people – folks who simply can’t find jobs. In the US, life expectancy has already dropped three years in a row—something that hasn’t happened since 1918. Deaths from suicide, drug overdose, and alcohol abuse are skyrocketing, killing hundreds of thousands of Americans yearly. Economists Anne Case and Angus Deaton call this the “deaths of despair” spiral.

IV. How to Escape the Turing Trap

The Turing Trap happens when we focus too much on “replacing humans” instead of “making humans better.” Think about it this way: you could build an AI self-checkout system that eliminates cashier jobs completely, or you could create a system that makes cashiers super-efficient by handling price lookups, suggesting products, and providing information while keeping the human touch.

To avoid this trap, the paper suggests three main solutions:

1. Fix the tax system: Right now, the system is unfair. When companies use AI to replace workers, they just pay corporate tax. But when they use AI to make employees better, they pay corporate tax PLUS wage taxes AND income taxes. Plus, investment profits are taxed at only 20% while working income gets taxed up to 37%. This setup practically pushes companies to replace workers! We should level the playing field, maybe even make labor income taxed less to encourage keeping humans in the loop.

2. Invest in education: Research shows that for every $1 spent on AI technology, companies should spend $9 on training people to work with it. But companies don’t want to train workers who might leave, and workers can’t afford training themselves. Governments need to step in – either provide training directly or give companies incentives to train their workers.

3. Encourage true innovation: We’re stuck thinking about “how machines can do human jobs” when we should be asking “what amazing new things can humans and machines do together?” We need new ways to measure success that reward breakthroughs creating new value, not just automating existing jobs.

While these directions are correct, specific policy implementation requires careful consideration. Any policy is a double-edged sword and must adapt to national conditions of each country, avoiding problems such as regulatory arbitrage and moral hazard. For example:

• Regulatory arbitrage case: In the 1980s-1990s, Korea and Japan successively implemented strict environmental standards and high tariffs on automated equipment in factories, causing manufacturing companies like Samsung to move production lines to Southeast Asian countries with lower environmental standards, forming a typical regulatory arbitrage. This shift not only failed to achieve environmental protection goals but also exacerbated employment problems for young people in both countries.
• Green energy subsidies: In the 2010s, some European countries provided high subsidies for solar power generation, and some companies fraudulently obtained subsidies through technology packaging, eventually leading to policy tightening.

This article will not expand the discussion at the policy level, but I want to propose:

If we think from another perspective, the deeper question is: these measures are all built on the premise that “work is necessary.” But if AI can truly replace human labor on a large scale in the future, do we need to rethink social structure and value distribution systems? This question may be more worth pondering than exploring how to keep jobs.

What kind of world would it be if humans don’t need to work in the future? How would it operate? Next time, I’ll have the opportunity to elaborate on this topic using Oxford University’s Daniel Susskind’s book A World without Work.

References

• Anthropic. (2025, February 10). Introducing the Anthropic Economic Index. Anthropic. https://www.anthropic.com/news/the-anthropic-economic-index
• Brynjolfsson, E. (2022). The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence. Daedalus, 151(2), 272-287. https://doi.org/10.1162/daed_a_01915
• Case, A., & Deaton, A. (2020). Deaths of despair and the future of capitalism. Princeton University Press.
• Handa, K., Tamkin, A., McCain, M., Huang, S., Durmus, E., Heck, S., Mueller, J., Hong, J., Ritchie, S., Belonax, T., Troy, K. K., Amodei, D., Kaplan, J., Clark, J., & Ganguli, D. (2025). Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations. Anthropic.