AI’s destructive cycles threaten the industries that feed it

The systems feeding the models

An internal Microsoft document called the pattern a “destructive cycle”: AI-generated content, it argued, can degrade both models and the internet. The unusual part, the document said, was that the end product threatened the economic foundations of key suppliers.

That tension is already visible in publishing. Journalists once expected online platforms to send readers to their work. ChatGPT and Google’s Gemini can instead summarize articles without sending audiences to publishers’ sites, deepening pressure on independent media. The open internet has reached a bleak threshold: AI-generated content now exceeds material created by humans. Newspapers, media companies and local news stations are struggling to stay afloat as companies, particularly Google, change how people use the web.

Court filings in The New York Times’ copyright case offer another view of the same dependency. People inside the companies have acknowledged that the system is built on taking others’ material. One Microsoft document warned that millions of people could soon see large models’ “absorption” of their work as theft on an unprecedented scale.

The pressure extends beyond publishing:

When both sides automate

Hiring shows how quickly these loops can compound. More desperate applicants use AI to improve résumés and fool automated screening systems. Recruiters respond by wading through a flood of nearly identical applications.

Daniel Chait, CEO of applicant-tracking company Greenhouse, told Wired that both sides face the same problem: each uses AI to fix its own side of the process, but ends up making the situation worse. The more applicants and recruiters rely on AI, the more widespread its use becomes—and the worse the outcome for everyone.

Schools are developing a similar pattern. Students increasingly hand homework to AI and outsource part of the thinking; teachers use AI to prepare assignments, leaving less human involvement in the learning process. Educators who resist the technology can find themselves in conflict with school administrators.

A recent study by the Center for Universal Education at the Brookings Institution concluded that the risks of generative AI in education far outweigh its benefits. Teachers are also warning that college students are rapidly losing the ability to read.

The bill for dependence

AI advocates, including lawyers for OpenAI and Microsoft in the ongoing case with The New York Times, argue that models create something new. Their position is that outputs differ enough from training material to qualify as fair use, a long-standing principle in US copyright law. The courts have yet to settle that argument. Anthropic, meanwhile, has already agreed to pay billions of dollars after being found to have infringed copyrights.

My concern is that winning the legal argument would not resolve the economic one. Even if model outputs count as fair use, the technology still depends on cultural and educational systems that produce the material it learns from. The announcement of a legal defense is not a plan for keeping those systems viable.

That is the deeper risk in the destructive-cycle framing. An industry can extract value from publishers, workers and educators for a while; it cannot assume those sources will remain healthy while their economics deteriorate. If the gains stay with those at the top as others lose the means to participate, AI may weaken the very supply of human work it needs.