AI has been writing for only a few years, but it could soon surpass all human writing — in volume — from the last 500 years.
That’s the implication of data shared by Brett Winton, Chief Futurist at ARK Invest, showing that AI annual written output crossed the human threshold in 2025 — and that cumulative AI output is on track to eclipse the entire written record of human civilization by the end of this decade.
The Chart That Changes the Frame
ARK’s chart tracks the annual written output from 1500 to 2030 on a logarithmic scale. The human line climbs steadily over five centuries — from the printing press era through the internet age — reaching roughly 200,000 trillion words by 2025. The AI line barely registers until around 2022, then goes nearly vertical, crossing annual human output in 2025.
The logarithmic scale is critical context here. On a log chart, what looks like a modest uptick represents orders-of-magnitude jumps. The AI curve isn’t just growing fast — it’s growing at a rate that compresses centuries of human output into years.
What “Written Output” Actually Means
The scope is everything: every postcard, memo, whitepaper, legal brief, business email, and news article humanity has produced since Gutenberg. This is the baseline AI is racing against — and the pace of AI model development suggests the race won’t last long.
Large language models today generate text at speeds no human writer can approach. A single model, queried millions of times a day across enterprise tools, consumer apps, and automated software pipelines, produces volumes that dwarf any individual or newsroom. Aggregate that globally, and the numbers become difficult to reason about intuitively — which is exactly why the log chart is the right lens.
The Cumulative Milestone Still Ahead
It’s worth being precise: what crossed in 2025 was annual output. AI is now writing more words per year than humans do. The larger milestone — cumulative AI output surpassing the full 500-year archive of human writing — is what Winton projects for the late 2020s.
That second crossing requires AI to sustain and accelerate its current trajectory. Given that AI capabilities are compounding rapidly, with models becoming faster, cheaper, and more widely deployed each year, the projection isn’t especially aggressive.
What It Means
The implications extend beyond a striking data point. Training future AI models requires text data. If AI-generated content now dominates the written internet, models will increasingly train on their own outputs — a feedback loop with uncertain consequences for quality, diversity, and drift. Researchers have begun calling this “model collapse,” and it remains an open problem.
There’s also the question of what a synthetically dominated information environment does to how humans find, trust, and use written knowledge. The 500-year human archive wasn’t just large — it was the substrate for education, law, culture, and science. A world where that archive is a rounding error against AI output is genuinely new territory.