More Articles Are Now Created by AI Than Humans
🔗https://shre.ink/more-articles-created-by-AI-than-humas
1. Key Takeaways
Graphite’s study finds that at a point in late 2024 the volume of articles published on the open web that are AI-generated surpassed the volume of those written by humans. However, that growth has since plateaued rather than continuing to accelerate. A further disturbing but important nuance: while AI-written articles may now be more numerous, they are not proportionally visible in search engine results or AI tool citations—human-written content still dominates the top ranks. The study also flags that the statistics don’t account for AI-assisted human editing (where humans take AI drafts and heavily revise them) which may mean the real influence of AI is even larger.
2. Motivation
The study notes many companies have adopted AI-content generation as a cost-efficient alternative to paying human writers, seeking to drive traffic via Google Search, social channels and answer-engines. Also, improvements in AI quality (sometimes matching or exceeding human-written text) and difficulty distinguishing human vs AI writing have made the deployment tempting.
3. Results
By November 2024, AI-generated articles crossed the threshold of being more numerous than human-written ones, in the sample of ~65,000 English-language URLs studied.
The growth of AI-written content spiked after the launch of ChatGPT and related models: from modest percentages to ~39% of new articles within 12 months of ChatGPT’s debut.
But the upward growth flattened out: the proportion of AI-generated content has remained stable since around May 2024
When it comes to performance (visibility): Only about 14% of articles ranking in Google Search were AI-generated, and about 82% of articles cited by systems like ChatGPT and Perplexity were human-written.
Human-written articles tend to outrank AI-written ones when controlling for keywords (i.e., the methodologically paired comparison shows human articles rank higher).
4. Methodology
Graphite’s research follows this approach:
Sampled ~65 000 English-language URLs from the publicly available Common Crawl archive, published between January 2020 and May 2025, and classified as “article” or “listicle”.
Applied the AI-detection tool Surfer to each article, classifying each 500-word chunk as human- or AI-written, then aggregating. If >50% of content is AI-detected → article is “AI-generated”.
Evaluated detector accuracy: For the human-written baseline (pre-ChatGPT articles), false positive rate ~4.2% (i.e., human articles wrongly classified as AI). For AI-generated (GPT-4o generated) false negative rate ~0.6%.
5. Limitations
Graphite outlines important caveats:
The study did not assess articles that were AI-drafted and human-edited (hybrid content). The prevalence of that may be even higher than pure AI-written.
The detection algorithm (Surfer) may be less reliable for other models or newer ones; accuracy might decline as AI advances.
The sample uses Common Crawl data—this may exclude many pay-walled or proprietary webpages (likely human-written) and thus may bias the sample toward more publicly available AI-written content. This implies the actual human-written share may be higher.
The performance metric (visibility/ranking) is associated but causation cannot be definitively claimed (many factors affect search ranking).