TLDR
Pangram Labs closed a US$9 million funding round led by Menlo Ventures, bringing total capital raised to about US$13 million. On the same day, the company launched Pangram 4, a text detection model it says catches AI-humaniser tools 98.83 per cent of the time across 13 commercial services, with a false positive rate of just 0.0041 per cent. A companion image detector, released in research preview, claims 99.5 per cent accuracy on internal tests and flags outputs from tools including GPT Image, Midjourney, FLUX and AI video generators such as Kling and Veo. Schools, publishers and platforms are the primary customers as pressure mounts to distinguish human-authored material from machine output.
KEY TAKEAWAYS
A $9 million bet on content authenticity
Pangram Labs, the San Francisco-based AI detection company, closed a US$9 million funding round on 29 July 2026, with Menlo Ventures leading and Haystack, ScOp Venture Capital, Script Capital and Cadenza joining the raise.[1] Total capital raised now sits at roughly US$13 million, a figure that reflects how quickly institutional appetite for detection infrastructure has grown as generative AI tools spread across the open web.
The funding announcement arrived alongside two product launches, not one. Pangram 4, the company's most capable text detection model to date, and Pangram Image, a research-preview tool for identifying AI-generated visuals, both went live the same morning. That timing is deliberate: the twin release signals that the authenticity problem has outgrown text alone.
What Pangram 4 actually does differently
The core challenge for any AI text detector is not catching a raw GPT output. That has become relatively tractable. The harder problem is catching text that has been deliberately run through an AI humaniser, a class of commercial tool designed to rewrite AI drafts so they evade detection. Pangram 4 detects AI-humaniser tools 98.83 per cent of the time across 13 popular commercial text-generation services, and is over six times larger than its predecessor model.verifiedVerified Source: pangram.com[2]
Chief executive and co-founder Max Spero said the model works by learning patterns that persist even after humanisation. Spero said the model is learning the stylistic differences and the choices that AI makes consistently, and is able to use that to identify what makes something AI-generated with high confidence.[1] Rather than matching surface-level statistical fingerprints that humanisers are specifically designed to scrub, the model is trained to identify deeper generative choices that remain consistent regardless of post-processing.
The benchmark numbers Pangram Labs published are self-reported on internal test sets, which is worth flagging. On those internal benchmarks, Pangram 4 records a false positive rate of 0.0041 per cent and a false negative rate of 0.3396 per cent.verifiedVerified Source: pangram.com[2] Bushletter could not independently verify the figures. A false positive rate near zero matters enormously to the customers most likely to use the product: an educator or editor who falsely accuses a human writer of using AI faces real reputational and legal exposure, so the low false positive claim is the number they will scrutinise hardest.
The image detector: scope and accuracy claims
Pangram Image is the more technically ambitious of the two launches. The research-preview model claims to identify AI-generated images from a wide range of generators, including OpenAI's GPT Image, Google's Gemini Nano Banana, Midjourney, FLUX and Grok Imagine, as well as individual frames extracted from AI video tools such as Kling, Seedance, Veo and Wan.[3] The breadth of that coverage list reflects how quickly the image generation market has fragmented: a detector that handles only one or two providers becomes obsolete almost as fast as it ships.
On internal benchmarks, Pangram Image claims 99.5 per cent accuracy, a false positive rate of 0.16 per cent on real-world pre-2022 images, and over 99 per cent AUROC on the external NTIRE 2026 AI Image Challenge.verifiedVerified Source: pangram.com[3] The NTIRE challenge provides at least some external validation that the internal figures are not purely self-serving, though the model remains in research preview and independent peer review has not yet been published.
Pangram Labs has bundled image scanning into existing text-detection plans at no additional cost. Individual subscribers receive up to 300,000 words and 100 image scans per month, while Pro subscribers receive up to 1.5 million words and 500 image scans.[2] By making image detection a free inclusion rather than a separate product, Pangram Labs reduces the friction for existing customers to start testing the image tool, which in turn accelerates the data the company needs to move the model out of research preview.
Who is buying and why demand is accelerating
Pangram Labs serves schools, publishers, content moderators and journalists.[1] Each group faces a version of the same underlying problem, but the stakes differ. For educators, the question is academic integrity; for publishers and media organisations, it is audience trust; for content moderators on social platforms, it is volume, with AI-generated submissions growing faster than any manual review process can handle.
Spero framed the product's purpose in terms of epistemic trust rather than policing. Spero said it is incredibly valuable to know whether what you are looking at is AI-generated, especially text, because it changes how people approach what they are reading and whether they need to watch for hallucinations and engage sceptically, or whether it is something they trust was well-researched from an actual journalist.[1]
That framing positions detection not as censorship but as metadata. The argument is not that AI-generated content is inherently bad; it is that readers and institutions deserve to know what they are dealing with before deciding how much scrutiny to apply.
The arms race dynamic and what it means for publishers
The humaniser-detection capability in Pangram 4 reveals a dynamic that anyone publishing AI-assisted content professionally needs to understand. The detection market and the humanisation market are co-evolving, with each improvement in detection accuracy creating commercial incentive to improve humanisation tools, and vice versa. Pangram Labs is wagering that its model architecture can stay ahead of that cycle long enough to build a durable customer base and data advantage.
Publishers integrating AI into editorial workflows now face a credibility question that is no longer hypothetical: if a detection tool flags a piece of content as AI-generated, regardless of whether the flag is accurate, the downstream reputational cost falls on the publisher, not the tool vendor. A very low false positive rate combined with near-complete humaniser detection is precisely the specification those publishers need before they can deploy detection as a trust signal rather than merely an internal quality gate.
Pangram Labs has not yet published third-party audit results for either Pangram 4 or Pangram Image, and both the text and image accuracy figures cited here come from the company's own benchmarking. Independent validation will be the next test of whether those numbers hold at scale across real-world content distributions rather than curated test sets.
SOURCES & CITATIONS
FREQUENTLY ASKED QUESTIONS
What did Pangram Labs raise and who backed the round?
What makes Pangram 4 different from earlier AI text detectors?
What can Pangram Image detect?
Are the accuracy figures independently verified?
How is image detection priced?
Ava Nguyen writes about internet culture and the platforms younger Australians actually use. She reports from inside the feeds rather than about them.
