AI is eating the web's memory and people are noticing
A thread on AI consuming the web's collective memory got traction, with commenters debating whether Google Search has gotten better or worse. The split was interesting: some said Google is the most useful it's been in years, calling it a 'lucky side effect' of the move to AI mode. Others said AI-generated answers are aggressively surfaced even when you're looking up something specific, and that the underlying web of linked human-written content is degrading.
This isn't just a search quality debate. It's a feedback loop problem: AI trains on web content, AI-generated content fills the web, future AI trains on AI output. The concern is that the original signal, human experience and expertise, gets diluted. The sunset anecdote in the article (someone got the wrong time from Google and missed it) was called out as weak evidence, but the underlying anxiety is real and keeps resurfacing.
The Claude watermarking thread connects here directly. Anthropic is rolling out imperceptible text watermarks in the EU and signed provenance metadata for generated files. The stated use case is catching academic cheaters, but the deeper function is maintaining a distinction between human and machine-generated content as that line blurs.
So what?
If your product depends on web search quality or user-generated content, the degradation of the underlying web is a real risk to your data pipeline. For founders building AI products, the watermarking and provenance trend is worth watching: regulatory requirements around content labeling are coming, and building for them now in the EU means you're ahead of the curve when they arrive in the US.