Pangram has raised $9 million to develop systems capable of detecting AI-generated text and images. The startup has also partnered with Substack, which uses its technology to provide signals on AI usage across newsletters.
The question seems simple — human or machine? — but technically it is increasingly non-binary. An author might generate a draft, rewrite half of it, use AI only for a few sentences, or run the text through multiple tools.
The detector must measure uncertainty
Detectors have often suffered from false positives, which are particularly problematic when used in schools, workplaces, or content moderation. Because of this, the challenge is not to produce an infallible stamp, but to build probabilistic signals reliable and explainable enough to be used with caution.
If the generative web continues to grow, content origin will become increasingly valuable information. Pangram is betting that this provenance can become infrastructure, not just a curiosity.
