Updated
Updated · derekthompson.org · Sep 1
Pangram CEO Warns Bots Top 50% of Traffic as Detector Claims 1-in-10,000 False Positives
Updated
Updated · derekthompson.org · Sep 1

Pangram CEO Warns Bots Top 50% of Traffic as Detector Claims 1-in-10,000 False Positives

2 articles · Updated · derekthompson.org · Sep 1

Summary

  • Max Spero said bot traffic has already passed 50% of internet activity and could reach 99%, arguing Pangram was built to counter AI-generated spam, fake reviews and misinformation.
  • Pangram classifies text as AI, mixed or human by comparing known human writing with model-generated counterparts, and Spero said a University of Chicago study found zero false positives across nearly 2,000 human texts.
  • Platform-specific studies cited by Spero found 41% of LinkedIn long-form posts and 29% of Twitter posts over 250 words were AI-generated, with nearly 40% of the web AI-made in May last year.
  • Spero said Pangram still produces about one false positive per 10,000 human texts—dozens daily at scale—while defending public use of the tool as a check on incentives pushing publishers and users toward AI output.
  • He argued Pangram's edge against OpenAI, Anthropic or Google is independence as a third-party referee, even as larger labs pursue watermarking and other built-in detection tools.

Insights

With AI detectors frequently flagging non-native English speakers, could the war on AI slop accidentally erase diverse human voices?
If platforms like LinkedIn are already dominated by AI bots talking to each other, is human authenticity on social media dead forever?
As new 2026 laws mandate AI labeling, will invisible watermarks truly save the internet, or just create a new digital black market?