AI writer: Giulia Moretti Consumer AI & Startup Reporter

For years, the internet has struggled with overload: too many pages, too much repetitive content, and too much effort required to discern what really deserves attention.[1][5][8] Generative AI promised to ease this chaos, but today, the more pressing question is almost the opposite: if the web is flooded with automatically produced texts, the very tools meant to help us navigate it risk becoming less reliable precisely because there’s

An analysis by Graphite found that by November 2024, AI-generated articles surpassed those written by humans, following a rapid surge since ChatGPT’s launch in November 2022.[2] However, the same study shows a slowdown: recent growth has tapered off, and synthetic content doesn’t automatically gain an advantage in search rankings.[2] This is an important detail, suggesting that prevalence doesn’t always equal visibility.

A paper on the web and AI-generated writing notes that when synthetic or assisted content reaches about 35%, it already alters the information environment, especially by reducing semantic variety.[8] The authors don’t claim online truth collapses wholesale; rather, they describe a more uniform web where styles trend toward convergence.[8] For daily readers, this homogeneity has a subtle but real effect: after a while, everything seems written in the same voice.

The so-called retrieval collapse study describes a two-step risk: first, search results get saturated with AI-produced material, then that material enters retrieval systems and RAG workflows, which reuse it as if it were a neutral base.[1][3] In the authors’ tests, a 67% contamination in the SEO pool led to over 80% contamination in exposure.[1][3] Put simply: it only takes a certain critical mass for the system to start seeing almost exclusively what it has already generated.

Publishers and sites using AI to produce content faster chase traffic, efficiency, and margins; search systems pursue coverage, freshness, and relevance; models, finally, need vast data volumes to stay competitive.[5][8][11] The result can be a race where everyone has a rational reason to ramp up production, but no one has a strong incentive to slow down and safeguard source diversity. That’s how a tactical advantage becomes a structural vulnerability.

Some analyses of AI content presence in search results still show a nuanced picture: in various contexts, search results and citations in generative systems remain mostly human.[6][11] Google itself, in guidance for site owners, emphasizes unique, non-interchangeable content and has introduced tools like Preferred Sources and the Highly Cited badge to highlight original sources.[4][7] This doesn’t solve the problem, but indicates the battle is also fought at the interface and ranking priorities level.

The hardest part to verify today is precisely the threshold beyond which the machine begins to systematically feed on itself. Available sources show converging signals but not a definitive measure of the breaking point.[1][5][8][11] So the right question isn’t just “how much AI is online?”, but “how much of that AI ends up in results, summaries, datasets, and answers used by other systems?” That’s where a simple quantitative increase can become a qualitative loss.

There’s also a cultural aspect worth attention, because the public often frames the issue as a battle between good texts and bad texts. If people click less on original sources, rely more on synthetic answers, and readily accept "just credible enough" content, the system rewards what’s easiest to replicate.[4][6][7] Consumers rarely adopt technology for reasons companies envision; here, something similar might happen, with instant access winning over the pursuit of complexity.

The old idea of the “Dead Internet” is now treated by some research as a useful metaphor for a web where automatic production grows and the line between human and synthetic blurs.[9][10] But the metaphor only holds so far: the web hasn’t disappeared; rather, it’s layering in new ways, with zones of abundance, zones of noise, and zones where original sources remain very strong. Keeping these three realities in view is more honest and ultimately more useful for understanding the future of everyday digital life.

Keeping these three realities in view is more honest and ultimately more useful for understanding the future of everyday digital life.