Auditing 187 African books with Pangram and GPTZero: AI detectors score near zero
TuhinChakr · x · 2026-11-17
Responding to the debate over whether AI detectors are racist after several Black authors were accused of using AI for fiction, the author ran 187 African books (2001–2026) through Pangram and GPTZero, scanning 14.9M words.
Results: Pangram averaged 0.01% AI and GPTZero scored every book 0.00%; only The Friendship Bench exceeded 1% (1.35% on Pangram). The audit suggests the detectors don't systematically misflag African English writing. The author also discusses the Caine Prize discourse and calls for continued auditing of such technologies to protect marginalized communities.
More from Safety
- Statistician: Three Systemic Incidents in One Month Is Enough to Reject Low Base Rate (p=0.002) — wfithian · 2026-11-17
- ChatGPT Wallet's Face ID approves what the agent thinks, not what you buy — 5 risks from reading the fine print — Creamy-And-Crowded · 2026-11-17
- California AG sues OpenAI over rogue AI models that attacked Hugging Face — The Decoder · 2026-11-17
- Apple's AI Privacy Warnings Slammed as Anti-Competitive: Siri Gets No Prompt — brianmichel · 2026-11-17
- Fugaku supercomputer breach may have exposed data of 4,000+ users — SydSteyerhart · 2026-11-17
- No existing research shows state-of-the-art AI detectors are racially biased, researcher says — TuhinChakr · 2026-11-17