Over 50 Percent of the Internet Is Now AI Slop, New Data Finds (futurism.com)

A new analysis from SEO firm Graphite — summarized by Axios — finds that just over half (52% as of May 2025) of newly published English-language articles are at least half-written by large language models. Graphite sampled 65,000 articles from January 2020 to May 2025 and used the Surfer AI detector to label pieces with ≥50% LLM-generated content as “AI.” The data show a rapid jump after ChatGPT’s launch (roughly 10% in late 2022 to over 40% by 2024), peaking in November 2024 and since leveling out around a 50/50 split between AI and human-written content. The findings matter for the AI/ML community because they affect training data quality, detection work, and how models interact with web knowledge. Methodological caveats are important: Graphite relied on Common Crawl, which misses many paywalled (likely human-written) pages, and Surfer exhibited a 4.2% false positive rate on human articles and 0.6% false negatives on AI text — underscoring detector fragility. Other data suggest search engines surface far more human content (Graphite found 86% human in Google Search vs. 14% AI), implying AI “slop” may be getting demoted. The trend may therefore reflect both real saturation and growing hybrid workflows where humans use AI tools, blurring detection and raising questions about dataset contamination, evaluation, and content-moderation strategies.