"AI Should Not Replace Humans" – A Proposal

In June 2026, an essay published on Huxiu posed a fundamental question about the direction of AI (artificial intelligence) applications. The author, Sun Liping, criticizes the current AI industry’s rush toward “replacing human labor” and argues that AI should instead focus on “what humans cannot do” and “what humans do not want to do.”

The core of the essay is a warning against reducing AI’s value to mere efficiency gains and cost cuts. The article cites an example where a company stated its purpose for introducing AI as “improving efficiency and preparing for layoffs.” The anecdote of an engineer who, feeling uneasy about this response, canceled the contract, symbolizes the ethical dilemma facing the AI industry today.

Three Stages of Development and Their

Turning Point

Sun organizes AI applications into three stages. The first stage is represented by ChatGPT and similar conversational applications such as text generation and translation. The second stage involves information collection and processing, including search and content analysis. In these stages, human intervention was required each time.

The third stage is the emergence of agents. AI can autonomously plan tasks, break them down into steps, call tools, and remember long-term goals. Multiple agents can also collaborate. However, it is this third stage that has made human replacement through automation more pronounced.

In many professions—customer service, basic translation, simple copywriting, data entry, and some legal document drafting—humans are being replaced by AI. While this indeed contributes to cost reduction and efficiency, the unemployment problem is worsening.

Risks of “So-So Automation”

Nobel Prize-winning economist Daron Acemoglu describes this situation as “So-So Automation”—a term referring to the tendency of companies to replace humans with immature AI for short-term cost savings. This not only leads to unemployment but also undermines overall social efficiency due to declining service quality.

Acemoglu is particularly concerned about a scenario where tech companies focus exclusively on “replacement AI” while neglecting the development of AI that assists and collaborates with humans. In such a case, GDP might appear to increase, but actual consumption could collapse.

The background to AI heading in this direction lies in the revenue structure. Anthropic currently has 300,000 corporate customers, with annual revenue of about $7–9 billion, 80% of which comes from enterprises. The annual revenue of its Claude Code product has already reached $400 million. OpenAI’s B2B business is also growing rapidly. The most direct need from companies is cost reduction and efficiency, and major AI firms promote human replacement to meet this. Some point out that this is essentially a clever marketing strategy.

A New Direction for Added Value

Sun points out that some of the smartest money in Silicon Valley has already begun to shift direction, investing in biology laboratories. Many investors are betting that AI will transform the healthcare and pharmaceutical industries before the internet.

Drug discovery is one of the most expensive tasks in the world, taking an average of more than 10 years and costing over $1 billion, with many projects ultimately failing. Google DeepMind’s AlphaFold can predict protein structures in minutes. It has already predicted over 200 million protein structures, a field where Sun believes AI truly shines.

Even shortening the development time by 20% through AI drug discovery would generate more value than many internet products. The value directly tied to life and health far exceeds the advertising revenue model of competing for users’ attention.

Four Characteristics of the Right Development

Direction

In his essay, Sun lists four features of the desirable direction for AI applications.

First, the goal should not be merely improving efficiency but making AI do what humans absolutely cannot. Analyzing billions of molecular combinations to find new drugs or predicting over 200 million protein structures are tasks humans could never accomplish in a lifetime.

Second, AI should empower rather than replace humans. AlphaFold has not replaced a single biologist; instead, it frees them from tedious structural analysis, allowing them to focus on deeper biological questions.

Third, AI should create added value, not redistribute existing pie. Successfully developing a new drug does not replace existing workers but cures previously untreatable patients. This expands the pie, improving the welfare of all humanity.

Fourth, AI technology should move from the “bit world” to the “atom world.” According to Peter Thiel’s classification, computers and the internet belong to the bit world, while physical domains such as transportation, energy, and biopharmaceuticals belong to the atom world. AI’s entry into medical drug discovery is a groundbreaking link between these two realms. Such directions—including smart robotics—are the true way forward for AI, whereas applications that merely replace humans are insignificant.

Real-World Barriers and Exploration

However, there is a large gap between the ideal presented in this essay and reality. The reason AI-driven human replacement is progressing is not solely due to corporate laziness. The B2B market is easier to monetize, and for companies seeking short-term ROI, replacement AI is the most reliable investment target.

On the other hand, fields like AI drug discovery have long development cycles and low success probabilities. Only venture capitalists willing to take risks and some large technology companies can enter this area. Yet the fact that Silicon Valley money is flowing into biology laboratories indicates that long-term investment decisions have begun.

Editorial Perspective

In the short term, the momentum of “replacement AI” under the existing B2B model is likely to continue. Corporate pressure to cut costs remains strong, and the revenue structure of major AI firms will not change overnight. However, the risk of declining social efficiency and consumption collapse from “So-So Automation” pointed out by Acemoglu could affect corporate decision-making in the medium term. Especially if regulators focus on this issue, changes in companies’ AI adoption policies may occur.

From a long-term perspective, the editorial board believes that applying AI to “what humans cannot do,” as Sun advocates, holds the key to sustainable development. Successful cases like AlphaFold demonstrate AI’s potential to generate truly creative value. The key question is whether AI can undergo a paradigm shift from a “tool of efficiency” to a “tool of discovery.” If AI applications in physical domains such as healthcare, energy, and materials development advance, they have the power to transform industrial structures. However, this shift requires long-term investment and decisions that go beyond the logic of capital pursuing short-term profits.

This raises a question: Can companies, within the constraint of maximizing shareholder value, truly prioritize long-term social value in their AI investments? And is it possible to encourage a shift from replacement-type to creation-type AI through regulation or incentive design? As AI evolution accelerates, it seems that the direction of its application is determined not only by engineers and managers but also by the collective choice of society.

References

• Huxiu: “AI Should Not Do What Humans Do, But What Humans Cannot Do or Don’t Want to Do” — Published June 10, 2026