Pangram Raises $9 Million and Launches New AI Text and Image Detectors

Pangram has raised $9 million to expand software that distinguishes human-created material from generated content. Menlo Ventures led the round with participation from Haystack, ScOp, Script Capital and Cadenza. The company is also releasing Pangram 4 for text and a research preview of an image detector.

Pangram says its new text model exceeds 99% accuracy on AI-assisted and mixed human-AI writing and is better at recognizing output modified by so-called humanizer tools. The image system is not yet generally available and will be tested with researchers before a wider release.

Founders Max Spero and Bradley Emi trained the detector using tens of millions of human documents and created a synthetic counterpart for each one with a frontier model. The system learns recurring differences in word choice and style rather than depending only on watermarks or file metadata.

Detection is becoming a product category as platforms, schools, publishers and courts confront undisclosed generated work. LinkedIn, Substack and other services are introducing reporting or detection features. Pangram will compete with GPTZero, Originality.ai, Copyleaks and Winston AI.

Pangram's customers may use detection for very different purposes, from feed ranking to academic discipline. The company should set limits on high-stakes use and require additional evidence before a model score triggers punishment. A detector designed to identify broad internet trends is not automatically suitable for deciding whether one person acted dishonestly.

Detection companies face an adversarial cycle. As classifiers improve, model developers and content farms can adjust wording to reduce a score, while human writing is too varied for perfect identification. Pangram should publish performance by language, genre and model and avoid a single certainty label. Customers need thresholds tied to the consequence of an error. Using the system to prioritize moderation is different from using it to accuse a student or employee. The company can build credibility by treating detection as evidence in a broader review, not a machine verdict about authorship.

No detector should be treated as unquestionable evidence. False positives can damage students, employees and writers, particularly those using translation or accessibility tools. Pangram's strongest path is to provide confidence scores, explain limitations and combine model results with human review. The demand is real, but the company will be trusted only if it is as careful about uncertainty as it is confident about accuracy.