Human beings have always found ways to pool what they know. Oral tradition passed accumulated wisdom across generations before writing existed. Libraries made that knowledge persistent. The printing press made it distributable. Wikipedia made it editable and alive. Each step expanded the number of people who could contribute to and access a shared body of human knowledge.

AI is the next step in that sequence. Every capability a large language model has today came from human output. The writing it understands, the code it can help debug, the music theory it can reason about, all of it originated with a human. The model organized, synthesized, and made accessible what humans had already created. Remove the human contribution entirely and there is nothing left to train on. The output is a function of the input, and the input has always been us.

The more familiar framings of AI are economic, existential, or political. AI as productivity multiplier. AI as existential risk. AI as regulatory challenge. Each is legitimate in its own domain, but none of them accurately describes what the technology is at its foundation. A more accurate description is that AI is a tool for human collaboration at civilizational scale.

The Cultural Shift That Distorted the Conversation

For most of the twentieth century the cultural imagination around AI was relatively optimistic. The computers aboard the Enterprise in Star Trek were capable, helpful, and oriented toward supporting human curiosity and exploration. The relationship between human and machine was collaborative. The future felt like something to move toward.

The 1990s changed that. The Terminator had arrived in 1984, but it was the decade that followed which produced a sustained wave of dystopian AI narratives. The Matrix arrived in 1999. Academics and journalists began documenting the shift, with some framing it as a choice between a Star Trek utopia or a Terminator dystopia as the two poles of how society imagined its AI future. The same technology that once inspired optimism about human potential became the cultural symbol of human obsolescence, and the story society told about AI reshaped how it was built, regulated, and feared.

Daniel Kahneman's research on System 1 and System 2 thinking is useful here. System 1 is fast, automatic, and emotionally driven. System 2 is deliberate, analytical, and effortful. Fear activates System 1 efficiently while nuanced analysis of a complex technology requires System 2. Social media platforms, scaling rapidly through the same period, discovered that System 1 activation drove more engagement than System 2 content. Anxiety and outrage traveled faster and further than curiosity and analysis, and the commercial incentive structure of the attention economy was built around the faster brain. AI fear was abundant, accessible, and engagement-generating. The result is a cultural mental model of AI that is substantially more adversarial than the technology warrants.

What Happens When You Remove the Human Input

A 2024 study published in Nature examined what happens when AI models are trained on AI-generated content rather than human-generated content. The researchers found that models trained recursively on their own outputs experience what they termed model collapse, a degenerative process in which the model progressively loses the diversity, nuance, and accuracy of the original human-generated training data. The outputs become narrower, less representative, and eventually meaningless, and the only reliable corrective is continued access to genuine human-generated content.

As of April 2025, an estimated 74% of new web content was already AI-generated, meaning the training pipeline for future models is already under pressure from its own outputs. Human participation in the creation of content, code, ideas, and culture is the feedstock that keeps AI capable and improving. Reduce that participation and the quality of AI degrades in ways that compound over generations of training.

The Layoff Logic and Its Limits

In 2025 alone, companies attributed 55,000 layoffs in the United States to AI, twelve times the number from two years earlier. The logic driving these decisions is financially intuitive: AI can perform tasks that humans previously performed, and human labor carries costs that AI does not.

A Gartner study published in May 2026 surveyed 350 global business executives at companies with at least one billion dollars in annual revenue. It found that 80% of companies deploying AI had reduced their workforce, and that there was no correlation between those reductions and higher ROI. The companies that eliminated the most roles were not outperforming those that did not. The short-term arithmetic looked clean. The business outcomes did not follow.

There is a longer-term dimension that the ROI data does not yet fully capture. When people are displaced by AI and develop resentment toward it, they stop contributing to it. They stop using it, stop generating the human content that feeds it, stop advocating for its responsible expansion. The network effect that concentrated AI capability and investment into a small number of platforms runs in reverse when the population contributing to those platforms contracts. When a company replaces an employee with AI to execute a function, it is choosing to have the AI platform, and by extension any competitor using the same platform, execute that function. The institutional knowledge, domain judgment, and organizational context that employee carried does not transfer to the model but disappears instead.

The Concentration Problem

Network effects find points of least friction and concentrate growth there. OpenAI and Anthropic together capture approximately 89% of all AI-native revenue globally while the remaining 32 significant AI companies split the other 11%. This is the predictable outcome of network dynamics in a market where scale produces capability advantages that attract more users, which produces more data, which produces more capability.

The problem is what this concentration means for who gets to participate in building and shaping AI. A small number of platforms now effectively determine whose creative output, whose writing, whose code, and whose knowledge gets included in training data. That is a governance question with significant long-term consequences for how representative and capable the resulting models are.

If the population contributing to AI narrows, the AI narrows with it. Iterated across multiple training generations, a tool that was built on the collaborative output of a broad human population begins to reflect something smaller and less diverse. The trajectory matters as much as the current state.

The Reframe That Is Available Right Now

Leaders making decisions about AI deployment and workforce strategy have access to a frame that most public discourse does not use: AI as a human collaboration infrastructure that requires broad human participation to remain capable and improving.

This frame requires a different set of questions in the planning process more so than new legislation, though accountability for training data sourcing and the distribution of economic value from AI is a legitimate and under-examined policy area.

An enterprise asking only what AI can replace is asking a narrower question than the situation warrants. The fuller question is what human participation is required to keep the AI improving, and whether the workforce decisions being made are preserving or degrading that participation.

The organizations that treat AI as a collaboration tool and invest in the human infrastructure around it will have a different relationship with the technology in five years than those that treated it as a cost reduction mechanism. The models will reflect what we put into them. They always have.