Will AI Destroy the World Wide Web?
Moshe Y. Vardi · Communications of the ACM · 2025
This topic was addressed in a July 2024 Nature paper titled "AI Models Collapse when Trained on Recursively Generated Data." c "We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear", reported the authors.In other words, an LLM-generated Web is useless in providing data to train LLMs.A counterargument d is that model collapse is not inevitable.It can be avoided if LLM-generated content is added to human-generated content rather than replace it.Are we in an incremental content regime or a replacement content regime?Only time will tell.Cory Doctorow coined the phrase "enshittification" e to describe the business strategy by online platforms of hooking users with tempting products, only to degrade them later by shifting value away from users.Is the Web undergoing enshittification?The Web became immensely useful because many people generated content, and Google found the best-quality content in response to our questions.But the current quality of GenAI-generated answers is such that people argue that "The Entire Internet Is Reverting to Beta." f In fact, at the bottom of each AI-generated answer Google adds the disclaimer, in a small font, "AI responses may include mistakes."I have to teach my students to always click the link and go to the source, but I doubt many people do that.Will AI destroy the Web?Even ChatGPT agrees it could erode its value unless regulated and used responsibly.