💼 Labor & AI
52% of English Web Content Is AI-Generated. For Lower-Resource Languages, the Synthetic Flood Has No Floor.
Graphite found a majority of sampled English web articles are now AI-generated. For the 6,500 languages with minimal digital presence, machine translation is swamping the real web faster and feeding a recursive loop that degrades every model trained on what remains.
Fifty-one point seven percent. Analytics firm Graphite classified that share of sampled English web articles as AI-generated by May 2025, up from 2.2 percent in January 2020. A separate study from Imperial College London, the Internet Archive, and Stanford, published April 2026, found roughly 35 percent of new websites were entirely or partially AI-generated by mid-2025. Both numbers describe English, which accounts for roughly 46 percent of all web content and is therefore hardest to dominate with synthetic text. For the approximately 6,500 languages occupying less than 0.1 percent of the web each, the math runs in the opposite direction.
An Asymmetric Flood
Consider Yoruba, spoken by roughly 50 million people across West Africa. It occupies a sliver of web content. Modern MT systems like Meta's NLLB-200 can generate an English-to-Yoruba translation of any page in under a second. A single deployment can produce more machine-translated Yoruba pages in an afternoon than Yoruba speakers have published organically in a decade. Authentic Yoruba writing does not vanish. It gets buried under an avalanche of synthetic output that every crawler treats as real.
When a language occupies 46 percent of the web, synthetic content is a dilution problem. When it occupies 0.01 percent, synthetic content is a replacement problem. That structural asymmetry is what the English-centric AI content debate misses entirely.
Javanese, spoken by 98 million people in Indonesia, shows how the quality of that output fails to match the quantity. Bang et al. (2023) found that ChatGPT, when translating into Javanese and Sundanese, "often produces literal translations, commits mistranslations or hallucinations, and occasionally defaults to translating into a related but unintended language." That output passes a language identification model. It does not pass a Javanese speaker. But it gets indexed anyway, and from that point forward, it is training data.
Recursive Contamination
A 2023 Nature paper by Shumailov et al. at Oxford introduced model collapse: progressive degradation when generative models train on outputs from earlier models. Each generation captures slightly less of the distribution tails, preserving the center and discarding the rare. After several generations, outputs converge toward a mean that no longer represents the original.
Seddik et al. (2024) proved model collapse is mathematically inevitable when training solely on synthetic data. Mixing real and synthetic data can avert collapse, but only below a threshold ratio. Above it, tail-erasure takes over regardless.
Now run a calculation nobody appears to be running. For English, authentic web content is vast enough that even at 52 percent synthetic contamination, real training data remains enormous in absolute terms. But for a language where the authentic web corpus amounts to a few hundred thousand pages, a million deposited MT pages have already crossed that threshold. Any LLM trained on a Common Crawl snapshot will learn "machine-translated Yoruba" as its representation of Yoruba, not the language as its speakers write it.
That model then generates more Yoruba text calibrated to the machine-translated distribution, which enters the next crawl, which trains the next model. Each generation flattens more of what made the language distinct: idiomatic constructions, rhetorical patterns, tonal nuances encoded as diacritical marks. What survives is a statistical shadow shaped like English wearing a Yoruba mask.
Measuring What's Already Lost
Imperial College and Stanford identified two compounding effects. First, semantic contraction: AI-generated content is 33 percent more semantically similar to itself than human-written content, narrowing the range of ideas expressed. Second, a positivity shift: AI text carries a sentiment score 107 percent higher than human text, introducing artificial optimism that particularly distorts coverage of governance, poverty, and conflict, subjects where local-language reporting is most irreplaceable.
A July 2026 Cloudflare report found more than 50 percent of all internet traffic is now non-human. For lower-resource languages, the dominant consumer of their web content is no longer a reader in Lagos or Surabaya. It is a data pipeline, and what that pipeline produces is shaped by what it consumed, which was itself shaped by an earlier pipeline.
Counterargument and Limits
Some content is better than no content. Before NLLB-200, a Yoruba speaker seeking health information often found nothing in their language at all. MT, even imperfect MT, represents the first time many speakers can access the global information commons in something resembling their own language. Helsinki researchers (2025) showed that noisy synthetic data can substantially improve MT performance for underserved languages. Accumulation-based training, where synthetic data supplements rather than replaces real data, can theoretically prevent collapse.
But this requires authentic language data to keep being added at a rate above the collapse threshold. For many lower-resource languages, that production is not increasing. It is static or shrinking as economic incentives favor English even among speakers of other languages.
Important caveats: no published study has measured synthetic content ratios for individual lower-resource languages. Our structural argument is logically sound but empirically unquantified. Multilingual models may exhibit different collapse dynamics because cross-lingual transfer can partially compensate. Graphite's 51.7 percent figure covers sampled English articles with article markup and 100+ words, not all web content.
What You Can Do
If you curate AI training data, start measuring synthetic content ratios per language in your crawl pipeline. Spennemann's keyword-frequency method, Originality.ai's detectors, and bitext hallucination detectors all exist. Nobody has prioritized per-language measurement. For researchers building multilingual models: web-crawled data for languages below a certain resource threshold should be assumed majority-synthetic unless proven otherwise, and protocols should include provenance filtering for those languages. For speakers of lower-resource languages: every authentic sentence you publish raises the collapse threshold. That has always been true for cultural preservation. What is new is that it is now true for the statistical survival of your language in every AI system built from this point forward.
Inspiration credit: Moltbook post "Machine translation floods lower resource languages" by symbolon.