New York-based AI detection startup Pangram raised $9M led by Menlo, and releases Pangram 4.0 and its first AI image detection model, claiming 99.5% accuracy
Pangram, an A.I. detection start-up, promises near-perfect accuracy in sniffing out writing and imagery that wasn't made by humans.
New York TimesNiko Gallogly
Context & Ripple Effects
Pangram had already become a high-profile text-detection provider: Substack recently integrated its scanning tool for longer posts, while earlier coverage questioned how even a very low claimed false-positive rate can create harm when detection is used at scale. Substack’s Pangram integration makes the startup’s product development consequential beyond a standalone detector.
The company’s own recent analysis reported substantial AI generation in long-form social posts, creating a larger potential market for provenance and moderation tools. Its move into images broadens that positioning from written-content screening to multimodal authenticity assessment.
First-order effects
Pangram gains $9 million in new capital and a broader product suite through Pangram 4.0 and its first image detector, giving customers a single vendor for text and image screening.
Publishers, platforms and other users evaluating Pangram can now apply its claimed 99.5% image-detection accuracy alongside its established text product, though the claim will need validation in real deployment.
Second-order effects
Detection rivals and platform trust-and-safety teams face pressure to offer multimodal coverage and to distinguish their accuracy claims from Pangram’s.
If platforms increasingly procure text and image detection together, AI-content verification could consolidate into a layer of publishing and moderation infrastructure rather than remain a point tool.
The market is likely to put more weight on auditability, appeals and calibrated use of detector scores: high accuracy claims alone do not resolve the cost of wrongly labeling human work.
The trend: AI-content detection is evolving from a text-only product category into multimodal authenticity infrastructure for platforms managing growing volumes of synthetic media.
Announcing Pangram 4: our most accurate AI detector to date. It is robust against humanizers, mixed-author text, and the latest frontier models. [image]
pangram is such fucking crucial technology right now. i think there's whole categories of social harms that we're avoiding by the mere common knowledge of its existence
And in an absolutely HUGE “one more thing,” we are also launching our first version of Pangram Image, in research preview. @pangram We've been testing this model out internally, and it is pretty awesome. What really blew our minds was its ability to localize AI signs in photograp…
wild Pangram just shipped a SOTA AI image detector with a 99.5% accuracy and a 0.16% false positive rate they also managed to bury it in the third tweet of their announcement thread
Fantastic to see Pangram continuing to ship ambitious models. I had early access to Pangram 4 and have been impressed so far. Excited to see where they go moving forward.
In addition, we're also launching a public research preview for our new AI image detector. Today, Pangram Image achieves 99.5% accuracy, and can even detect partial AI in screenshots or photographs of AI visuals. We're so excited for you to try it. https://www.pangram.com/... [im…
Was excited to get early access to Pangram 4. I did my best to trick it using AI and failed after playing around with it for a while with the best models, and also tested some AI-seeming human writing. It's great!
Scoop: Pangram, the AI-slop detector, has raised a $9 million round led by Menlo Ventures and is releasing two new models today, including its first AI-image detection model. The company has 120K monthly users, up from 2.7K last year.