The Reverse Big Push: Generative AI and Self-Fulfilling Automation

In the modern economy, the way businesses make money often depends on a delicate balance between what they spend to operate and the money that flows back in from customers. When a company hires workers, it pays them wages. Those workers then spend their paychecks at other local businesses, creating a cycle of demand that keeps the whole economy moving. This cycle is a fundamental part of how local markets function. For decades, economists have studied how new technologies change this balance. Sometimes, a new machine is so expensive to buy that only a few companies can afford it, and those companies must work hard to find enough customers to pay for the machine. This is the classic idea of a "big push," where high costs force a group of businesses to grow together to survive.

However, the rise of generative artificial intelligence has flipped this script. Today, the most expensive part of using advanced AI—the cost of training the computer system—is paid by the company that built it, not the company that uses it. A law firm or a design studio can simply rent the AI's power by the hour as they get new work. This makes automation look like a flexible operating expense rather than a heavy, upfront investment. But there is a catch. If a business decides to use this AI to replace its human team, it stops paying those wages. If many businesses do this at once, the workers who lose their jobs stop spending money in the local economy. This shrinks the market for everyone else, making it harder for the remaining businesses to survive. The question researchers are asking is whether this shift creates a trap where businesses, acting in their own self-interest, accidentally destroy the very market they rely on.

A team of economists from China has built a model to explore this specific dynamic, focusing on how the choice between using AI with human help versus using AI to replace humans plays out in a local service economy. They found that the economy can get stuck in two very different states, depending on what businesses expect will happen next. In one state, businesses keep their human teams, wages remain high, and the local economy is vibrant. In the other, businesses cut their teams to save money, wages fall, the local economy shrinks, and businesses cut even more teams to survive the smaller market. The researchers call this a "reverse big push." Unlike the old model where high costs forced growth, here the ability to cut costs easily can trigger a self-fulfilling collapse.

The core of their discovery is that the decision to automate is not just about whether the technology works well; it is about how many other businesses are doing the same thing. If a business believes that most other firms will keep their human teams, it makes sense for them to keep their team too, because the local economy will be strong enough to support the higher wages. But if a business believes that everyone else is firing workers to switch to AI, it becomes the smartest move for them to fire workers too, because the local economy will be too weak to support a full team. This creates a situation where the economy can settle into a "low-demand" equilibrium where automation is the only choice, even though a "high-demand" equilibrium where humans and AI work together would be better for everyone.

The researchers show that this problem does not disappear even if wages are allowed to fall. In a normal market, if demand drops, wages drop, which makes it cheaper to hire people again. The model confirms that falling wages do make hiring cheaper, but they also mean that workers have less money to spend. The study finds that if the loss of spending power is greater than the savings from lower wages, the cycle of decline continues. The economy gets stuck in a low-wage, low-employment state because the drop in demand is too severe to be fixed by cheaper labor costs alone. This means that simply letting the market adjust wages is not enough to prevent a collapse if businesses lose confidence in the future.

To understand how this plays out, the authors looked at how businesses make decisions over time. They imagined a scenario where firms get the chance to change their strategy at different moments. If a firm is pessimistic and expects others to automate, it will automate first, which validates the pessimism of the next firm, creating a cascade of layoffs. If a firm is optimistic and expects others to keep their teams, it will keep its team, encouraging others to do the same. The starting point matters, but so does the expectation. The researchers found that there is a specific range of conditions where both outcomes are possible. In this range, the economy is not determined by the technology itself, but by the collective mood of the business community.

The study also examined what happens when a public shock, like a sudden change in the economy's overall health, occurs. They found that if the economy is in that uncertain middle ground, a public fundamental capable of reaching both dominance regions can select a unique path. If the economy is already leaning toward automation, a negative shock can push it over the edge into a full collapse. Conversely, if it is leaning toward human-AI cooperation, a positive shock can secure that path. The researchers used a concept called "risk dominance" to predict which state is more likely to be chosen when the future is uncertain. They found that if the static tipping point for human-AI cooperation lies below one half, the economy will likely choose the cooperative path. But if the tipping point lies above one half, the economy will likely choose the automated path, even if the cooperative path would have been better for society as a whole.

The authors then looked at what this means for policy. They calculated that in many professional service sectors, the conditions for this "coordination failure" are real. Using data from the United States, they estimated that for certain types of high-autonomy AI tasks, the economy could easily fall into the bad equilibrium. They found that the gap between the "good" state and the "bad" state is not huge, but it is wide enough to matter. In their simulations, a shift in how much of a job can be done by AI without human help could move the economy from a stable state where humans and machines work together to a state where machines replace humans entirely, simply because the market shrank too much to support the teams.

The researchers argue that the solution is not to ban AI or force companies to keep workers. Instead, they suggest that a temporary government intervention could help the economy cross the gap. If the economy is stuck in the low-demand state, a small, temporary subsidy or a tax on pure automation could act as a bridge. This bridge would give businesses the confidence to keep their human teams until the economy grows large enough to support them on its own. Once the economy reaches a certain size, the subsidy could be removed, and the human-AI model would become the natural, self-sustaining choice. The key is that this support must be temporary and targeted; it is not about saving every job forever, but about preventing the economy from getting stuck in a bad state where everyone is worse off.

The paper also clarifies that this problem is specific to certain types of work. It applies most strongly to "defensive" automation, where businesses use AI to cut costs because they are afraid the market is shrinking. It is less likely to happen with "offensive" automation, where AI creates new products that expand the market. In the defensive case, the fear of a shrinking market drives the decision to cut costs, which in turn shrinks the market. The researchers show that if a business is cutting costs because it expects the market to shrink, it is often better for the whole economy if that business keeps its workers, because their wages will help keep the market alive.

In their final analysis, the authors used real-world numbers to show how sensitive this system is. They looked at the ratio of revenue to payroll in professional services and found that in many cases, the economy is right on the edge of the coordination region. A small change in the cost of AI or the ability of AI to do tasks could tip the balance. They found that wage adjustments, while helpful, compress the range where this coordination failure can happen, but they do not eliminate it. The study concludes that the future of work in the age of generative AI depends less on the raw power of the technology and more on the collective confidence of the businesses using it. If businesses can coordinate to keep their human teams, the economy can thrive. If they cannot, they risk a self-fulfilling collapse where the very tools meant to help them end up hurting the market they serve.

The work provides a clear warning that the transition to an AI-driven economy is not automatic. It requires navigating a narrow path where expectations and reality reinforce each other. The researchers show that without a mechanism to break the cycle of pessimism, the economy can settle into a state that is stable but inefficient, where everyone is poorer than they could be. The solution lies in understanding that the choice of technology is not just a technical decision, but a social one that depends on the health of the entire local economy. By recognizing this, policymakers can design interventions that help the economy find its way to the better outcome, ensuring that the benefits of AI are shared rather than lost to a cycle of decline.