Home / Technology / Brooklyn Startup Cracks AI Writing Code
Brooklyn Startup Cracks AI Writing Code
2 Sep
Summary
- Pangram detects AI-generated text with a percentage score.
- The startup gained prominence after flagging a novel.
- Concerns exist about AI detection's accuracy and bias.
Pangram, a 24-employee startup based in Brooklyn, has raised $13 million to develop AI that detects AI-generated text. The company's technology assigns a percentage score indicating the likelihood of AI involvement in writing.
Pangram gained significant attention in January when its CEO asserted a self-published novel was 78 percent AI-generated, leading to its traditional publication deal being canceled. This incident was followed by accusations against The New York Times and a Commonwealth Short Story Prize winner, both scoring 100 percent on Pangram's detection.
Despite its growing influence, particularly after Substack announced an integration, Pangram faces criticism. Authors and publishing experts express anger and distrust, citing concerns about the software's accuracy and potential bias, especially against non-native English speakers and writers of color.
The company uses methods like 'synthetic mirroring' and 'hard negative mining' for detection. While Pangram claims a low false-positive rate of 0.0041 percent with its latest model, critics argue its effectiveness is limited, especially on shorter texts or those altered by 'humanizer' tools.
Pangram aims to be the definitive arbiter of originality across industries like publishing and education. CEO Max Spero emphasizes transparency and community engagement to build trust, even as the company navigates the evolving landscape of AI-generated content and potential legal challenges.
