Is AI-generated content rising faster on some platforms?

The paper measures how much social media text is machine-written instead of assuming it. It assembles SM-D, about 2.4 million posts (1,170,821 Medium articles, 245,131 Quora answers and 982,440 Reddit comments) collected from January 1, 2022 to October 31, 2024. It classifies the posts with OSM-Det, the detector that performed best on its AIGTBench benchmark (accuracy 0.979, F1 0.980), and reports the share flagged as AI-generated as the "AI Attribution Rate" (AAR). The headline result is the difference between platforms: Medium's AAR rose from 1.77 percent to 37.03 percent and Quora's from 2.06 percent to 38.95 percent, while Reddit's rose only from 1.31 percent to 2.45 percent. The excerpt places the break in December 2022. Before then "the AAR across platforms remains stable"; from December, "Medium and Quora show significant surges, while Reddit shows only a slight increase."

The reasons the paper gives are interpretive, and it offers them as such. It reads the simultaneous Medium and Quora surge as "widespread and diverse LLM adoption on social media," and it ties the platform differences to what each platform is for. The paper describes Medium, Quora and Reddit as places where "users emphasize the depth and quality of the information they share," and it treats each platform's "distinct characteristics" as the reason the trends differ. Within Medium, "Technology" and "Software Development" have the highest AARs from December 2022 onward, which the paper attributes to people in those fields being "more likely to know about LLMs and frequently interact with them." On Quora, the AAR fell from its early-2023 peak to a low near 19.79 percent in September and October 2024. The paper's likely explanation is that Poe, Quora's LLM platform, produced an initial surge that faded once users found its capabilities insufficient.

The web-scale figure in How much of the internet is AI-generated now? comes from the same kind of detector run, but on a sample of websites and as a single aggregate. This paper adds a per-platform time series, which is what shows the rise is uneven. The two numbers rest on different samples and definitions, and the excerpt does not compare them, so neither checks the other. The AAR is also a count of detector flags. Why do fake news detectors flag AI-generated truthful content? shows that such flags can follow linguistic style rather than authorship, and this paper's own analysis fits a style reading: it finds sentence-level patterns "provide more distinctive characteristics," with AIGTs "more objective and standardized" than human-written texts.

The excerpt does not establish how accurate OSM-Det is on these posts. The 0.979 accuracy comes from AIGTBench, which mixes public datasets with the paper's own generations from social media texts by 12 LLMs. No accuracy, false-positive rate or human-labeled check on SM-D appears in the excerpt. The Discussion's claim is that the detector "exhibits strong generalization capability" to previously unseen LLMs, which is narrower than accuracy on real posts. The limitations add that the platforms are English-dominated and that the models are only those released after November 2022. The implication is that the direction of the rise is well supported across these platforms, while the level (37.03 percent on Medium) is best read as a detector-flagged share, not a measured share of authored text. Making the number mean authorship would need a human-labeled sample drawn from the same posts.