Pangram has raised $9 million to expand its tools for identifying AI-generated text and images. The New York-based startup also launched Pangram 4, its latest text detection model, alongside an image detector that is initially available through a research preview.
Menlo Ventures led the funding round, with participation from Haystack, ScOp, Script Capital and Cadenza. Pangram was founded about two years ago by Stanford graduates Max Spero and Bradley Emi.
Pangram 4 Targets Mixed and Edited Writing
Pangram says its new model can identify fully generated text, AI-assisted writing and passages that combine human and machine-written sentences. It is also designed to recognise output that has been processed by “humanizer” tools intended to bypass AI detectors.
The company’s Pangram 4 technical report recorded an AUROC score of 0.9916, a false-positive rate of 0.0041% and a false-negative rate of 0.3396 across its evaluation data. The report was written by Pangram researchers, so its performance claims should not be treated as a guarantee for every type of real-world content.
Pangram trains its system using large collections of human-written documents. It then creates matching AI-generated versions with similar subjects, lengths and styles, allowing the model to learn recurring differences between human and machine writing without relying on watermarks or hidden metadata.
Image Detector Enters Research Preview
Pangram Image analyses pixel-level patterns to estimate whether an image was generated by AI. The company says the model is designed to work across outputs from multiple image generators rather than detecting only content carrying a specific provider’s watermark.
The system can also highlight areas it believes contain generated imagery, including an AI-created picture displayed inside an otherwise real photograph. Pangram plans to make the image detector more widely available after the research preview.
AI detection systems remain imperfect. Edited writing, unusual personal styles and unfamiliar content types can produce incorrect classifications, making the tools more appropriate as an initial signal than as the sole evidence in academic, employment or legal decisions.
Substack Adds Pangram Detection
Substack recently integrated Pangram into its platform, allowing readers to scan eligible posts, Notes, comments and replies for estimates of human-written and AI-assisted text. The feature applies to content published from July 21, while publishers can disable detection for individual posts.
Pangram also sells access through a $20 monthly plan, an API and a browser extension. Its extension scans content on services including X, LinkedIn, Substack and Medium and provides an estimate of how much visible text may be AI-generated.
The funding arrives as publishers, schools and research platforms introduce stronger rules for unchecked AI output. ArXiv, for example, has said authors may face a one-year submission ban when papers contain clear evidence that generated material, such as fabricated citations or leftover chatbot instructions, was not reviewed.
Featured image credits: Magnific.com
For more stories like it, click the +Follow button at the top of this page to follow us.