Technology
AI detection startup Pangram faces industry scrutiny amid publishing controversies
The company's text scanner has contributed to canceled book deals and platform integrations, while critics warn of false positives and potential demographic bias.
The short version
- Brooklyn-based startup Pangram has gained traction across the publishing sector and online platforms, raising $13 million to develop software designed to flag AI-generated text.
- Pangram scores have played a role in public controversies and canceled deals involving books and articles, leading some literary agents and platforms like Substack to adopt its tools.
- Critics and researchers warn that AI detectors risk producing false positives and disproportionately flagging non-native English speakers or neurodiverse writers.
- Major publishers remain cautious, with some stating that AI-detection tools serve only as advisory components rather than final arbiters.
Key facts
- Pangram was cofounded in 2023 by Max Spero and Bradley Emi under the initial name Checkfor.ai, subsequently raising $13 million in total funding, including $9 million in July.[Wired]
- The company trains its detection software using 'synthetic mirroring' alongside 'hard negative mining' to recognize AI text patterns and minimize false positives.[Wired]
- Pangram's public AI-likelihood ratings were tied to high-profile industry events, including Hachette canceling the publication of the novel Shy Girl after Spero reported it was 78 percent AI-generated.[Wired]
- Substack integrated Pangram technology to allow readers to check posts for potential AI generation.[Wired]
- A survey conducted by Gotham Ghostwriters indicated that 61 percent of polled writers use AI tools, with 7 percent stating they have published AI-generated content.[Wired]
- Penguin Random House stated that approved AI detectors are used by its editors merely as an additional, non-determinative tool within a wider editorial workflow.[Wired]
What remains uncertain
- The extent to which AI detectors exhibit systematic bias against non-native English or neurodiverse writers remains disputed, with critics citing past studies while Pangram cites internal research to the contrary.[Wired]
- Substack declined to disclose the financial structure or terms of its partnership with Pangram, including whether it pays per scan.[Wired]