Pangram Labs funding reaches $9M for AI text and image detection

Pangram Labs funding now includes a new $9 million round led by Menlo Ventures, giving the AI detection startup fresh capital to improve text detection and move into image analysis. The company said the round will support its core detector and a new research-preview product for identifying AI-generated images, a category where accuracy claims matter because false accusations can carry real consequences.

Haystack, ScOp Venture Capital, Script Capital and Cadenza also joined the financing. Pangram said the round brings its total funding to almost $13 million, following a $2.7 million raise in June 2025. The company did not disclose its valuation, revenue, headcount or the structure of the round. The deal is a venture capital bet on AI provenance software at a time when publishers, schools and businesses are trying to tell whether content was made by people, models or some mix of both.

Pangram, formally Pangram Labs Inc., builds software that estimates whether text was generated by AI. The company says its existing platform has an industry-leading false-positive rate of 1 in 10,000, meaning it would incorrectly flag human-written material as AI-written at that rate under its stated testing conditions.

How accurate is Pangram's AI detector?

Pangram said its new Pangram 4 model produced a 0.0041% false-positive rate in internal benchmarks, or about one incorrect AI label per 24,000 documents. The company also said the model cuts false negatives, where AI-written text is missed, and performs better against “humanizer” tools that try to rewrite AI text so it appears human.

Those numbers remain company claims. Pangram’s accuracy case is based on internal testing, a technical white paper and a limited set of favorable third-party studies, according to SiliconANGLE. That distinction is important because AI detection is probabilistic: the systems estimate likelihood, rather than prove who or what wrote a passage.

Pangram describes its text system as a classifier model, a neural network trained to distinguish patterns associated with large language model output from patterns associated with human writing. The company says it trained the model using known human-written text from before 2021 paired with AI-generated samples. Pangram has also acknowledged that language use changes over time, which means the training approach has to adapt as writing styles and model outputs shift.

What is Pangram Image detection?

Pangram also introduced Pangram Image detection in research preview. The company said the image model can detect pictures generated by systems including OpenAI’s GPT Image, Google’s Gemini Nano Banana, Midjourney, FLUX and Grok Imagine, as well as output from some AI video providers including Kling AI, Seedance, Google’s Veo and Wan.

The company argues that image detection tools are needed because deepfakes and AI-generated images are increasingly presented as real. Watermarking systems such as Google’s SynthID can help when model providers add those signals, but Pangram said that approach is limited to systems that embed them in the first place.

Human review is also becoming less dependable, according to figures cited by SiliconANGLE from Let’s Enhance, a blog focused on AI creative tools. The report put overall human detection rates at 63.7%, with performance dropping to about 29% for images from higher-performing generators such as FLUX. The same research said average human ability to separate AI images from natural ones is moving closer to 50%.

The reliability problem is still unresolved

Pangram is entering a market carrying reputational baggage. MIT Sloan Teaching and Learning Technologies has said AI detectors are far from foolproof and can produce high error rates that lead to false accusations. The Mozilla Foundation has found that detector tools do not perform as reliably as they claim, and research reported by The Markup has shown that some tools can be biased against non-native English speakers.

That is why false positives are the central business risk for AI detection vendors. Some educational institutions have stopped using or restricted AI detectors, and others advise instructors to treat them as one signal rather than a verdict. The University of Waterloo discontinued use of Turnitin’s AI detection function in September, while MIT Sloan’s guidance says policy and expectations should do more of the work than detection software alone.

This story draws on original reporting from SiliconANGLE.