AI-Generated Articles Have Overtaken Human Writing—What’s Left of Our Thinking?

Prevalence of AI‑generated English articles

Graphite’s May 2026 study sampled 43,000 random pages from the CommonCrawl corpus and applied an AI‑detection algorithm with a 4.2 % false‑positive rate and a 0.6 % false‑negative rate. The analysis found that AI‑only content accounted for 39 % of all published English articles twelve months after ChatGPT’s launch and stabilized above 50 % in 2025, with no sign of decline.

Terminology

The mass‑produced, low‑quality AI text has been labeled “AI slop”; Merriam‑Webster selected “slop” as its 2025 Word of the Year to denote this phenomenon.

Historical precedents

In 1953, mathematician Christopher Strachey generated love letters on a Manchester computer, an early example of machine‑written prose. The same year, Roald Dahl’s short story “The Great Automatic Grammatizator” imagined a machine capable of producing half of the English‑language fiction within a year.

Decoupling of language and rational thought

Leif Weatherby, director of the NYU Digital Humanities Center, observes that contemporary models can generate language without any rational participation, effectively separating language from thought.

Model collapse

A 2024 paper in Nature (https://www.nature.com/articles/s41586-024-07566-y) documents “model collapse”: repeated fine‑tuning of AI models on their own generated outputs causes a generational decline in output diversity and quality, eventually collapsing into meaningless noise. The authors compare this process to genetic inbreeding and note that the resulting homogeneity of AI‑generated text reduces the incentive for human authorship.

Feedback loop and “fuel depletion”

Graphite’s data show that, while the proportion of AI‑only articles has plateaued since May 2024, the remaining niches for original human content are being gradually filled. This creates a feedback loop: increasing AI‑generated text becomes the primary training material for subsequent models, which in turn accelerates the homogenisation of output and further diminishes human writing.

Literary critic Matthew Kirschenbaum warns of a “textpocalypse” in which authentic human writing becomes a rare artifact.

Potential trajectories

Optimistic scenarios posit that a super‑intelligent AGI could learn without relying on human‑generated text, thereby breaking the feedback loop. Pessimistic scenarios predict that prolonged reliance on AI may erode human reasoning to the point where future generations lack the capacity to engage with advanced AI. A middle “warm‑water‑frog” state is described in which AI dominates discourse without having fully supplanted human thought, leading to a homogenised linguistic landscape.

Code example

来源：新智元
本文
约2500字
，建议阅读
5
分钟
2024 年 11 月，AI 生成的网络文章数量正式超过人类。Merriam-Webster 把「slop」选为 2025 年度词汇。当机器开始替人类说话，人类会不会忘记怎么思考？更麻烦的是，当人类停止书写，AI 用来学习的燃料也将一并耗尽。一场关于语言和思维的连环危机，正以多数人未曾警觉的速度展开。