As AI content floods the internet, Pangram raises $9M to detect it

Pangram, a New York company building tools to identify machine-made media, has secured $9 million in new funding as publishers, schools and online platforms confront a growing volume of synthetic material. The company is pairing the investment with the release of Pangram 4, its newest system for examining text, and a research-preview product designed to assess whether images were generated by artificial intelligence.

Menlo Ventures led the financing round. Haystack, ScOp, Script Capital and Cadenza also took part, according to the company. The funding gives Pangram additional resources to expand products aimed at a basic but increasingly difficult question online: whether a piece of writing, or an image, was made by a person, produced by software, or created through a combination of both.

The announcement arrives as generative AI tools make it faster and cheaper to create articles, marketing copy, social posts, illustrations and other digital material at scale. That expansion has increased demand for systems that can help organizations review content before it is published, submitted or relied upon. Pangram is positioning its software as part of the infrastructure for that review process rather than as a tool for making new content. For organizations handling large volumes, the appeal is not necessarily a final verdict from software, but a way to decide which material deserves closer human examination and what follow-up questions should be asked.

Pangram 4 is the company’s latest text-analysis model. Pangram says it can recognize AI-assisted prose as well as documents in which human authorship and generated passages are mixed together. The startup also says the model achieves accuracy above 99% in finding such writing, a performance claim that will be important to test in the range of real-world settings where the product may be used.

A notable focus for the new model is software sometimes marketed as an AI “humanizer.” Such programs attempt to alter generated language so that it sounds less formulaic or less likely to trigger automated checks. Pangram says Pangram 4 is better able to identify content that has gone through those tools, underscoring the continuing contest between systems that create or disguise synthetic text and systems intended to spot it.

The company has also introduced Pangram Image, an image-detection model. It is not yet broadly available: Pangram is offering it as a research preview while it gathers feedback and evaluates the product. A wider release is planned in the coming weeks, which will make the image tool a key next test of whether Pangram can extend its approach beyond written material.

Text and images present different challenges for detection. Written work can contain a blend of original drafting, editing assistance and generated passages, while an image may be altered, compressed, cropped or combined with other material before it reaches an audience. Those realities make clear disclosure and careful interpretation important for users of detection scores, especially when the result could influence editorial decisions, academic reviews or assessments of online credibility.

Detection products are being developed amid a wider debate over how much confidence institutions should place in automated judgments about authorship. Supporters see them as a way to add evidence when reviewing large amounts of content that cannot realistically be checked by hand. Skeptics, however, are likely to look closely at how systems perform across writing styles, languages, editing practices and the many ways people and AI tools can work together.

For Pangram, the new financing and product releases set up a period in which adoption and independent scrutiny may matter as much as technical claims. The company will need to show how Pangram 4 performs outside controlled testing and how its image model develops after the preview phase. With synthetic material becoming a more visible part of the internet, the value of reliable detection will depend not only on identifying AI output, but also on communicating the limits of what a detection result can establish.

Source: TechCrunch