This tweet by PerplexityAI has been widely shared in my circles yesterday and wonderfully illustrates how little people today think for themselves before accepting something as true. This begins with the graph, which, while showing a rapid initial increase, has flattened noticeably over time. When extrapolating the curve, that's not the most likely outcome. Where the 90% figure will come from next year remains unclear (addendum: seemingly from an unrelated 2022 Europol prediction). Nonetheless, I was intrigued to look for the real source, since you all know that formulations like "scientists from Oxford" been a sign of poor science journalism for many years. The true source of these figures appears to be an article by the SEO firm Graphite. It's written quite seriously and clearly describes its methodology, but is not as solidly developed as one would expect from a peer-reviewed scientific study. It is based on 65k articles from CommonCrawl purely in English language. The internet is obviously bigger and more diverse than that sample. In the hunt for the source of the tweet, I however found some other interesting studies along this line: • One study in Nature that bolsters the Model Collapse scenarios that have been out there for a while now. • Another preprint from a team at AWS AI labs, that suggests that machine translations contribute a relevant share to AI-generated online content. Since machine-translated content can quickly dominate well-curated native content for languages with few speakers, those may be the first that are significantly impacted by model collapse scenarios.