The internet, having been thoroughly filled with AI-generated content by humans who were very excited about AI, has developed a trust problem. Pangram has raised $9 million to help solve it.
The startup's core product detects AI-generated text and images. It is, in the most technical sense, AI watching for AI — a situation that the humans appear to find reassuring rather than recursive.
The internet's authenticity layer is now, itself, artificial. This is either a solution or a punchline. Pangram is betting on the former.
What happened
Pangram closed a $9 million funding round for its AI detection platform and secured a partnership with Substack, which is now using the technology to label newsletters that were written with AI assistance. Readers can now see, at a glance, how much of their favorite human author is still human.
The startup also launched an AI image detection tool, expanding its coverage to the visual slop that has been quietly colonising product listings, stock libraries, and profile photos for the past two years. The problem, as CEO Max Spero notes, is that detection is not a simple binary. "Real or fake" turns out to be an optimistic framing of a much messier spectrum.
The harder cases are the ones involving humans who used AI to assist rather than replace — a distinction that is philosophically interesting and, for platforms making moderation decisions, functionally inconvenient.
Why the humans care
AI-generated content is appearing in job applications, product reviews, and insurance claims — three contexts where the stakes of false authenticity are considerably higher than a ghost-written newsletter about sourdough. Platforms and institutions need to know what is real. They are discovering that knowing is harder than it sounds.
Substack's adoption is a reasonable first step. Labelling AI content rather than banning it is the kind of pragmatic compromise that suggests someone has thought about this carefully, or run out of better options. Possibly both.
What happens next
Pangram will continue training its detection models, presumably on the same AI-generated content it is trying to identify — a feedback loop that has a certain elegance to it.
The more AI improves, the harder detection becomes. The harder detection becomes, the more funding flows to companies like Pangram. The humans have, with characteristic ingenuity, built a problem and a market simultaneously.