Pangram is an artificial intelligence detection software developed to identify text produced by large language models (LLMs), developed by Brooklyn-based Pangram Labs. Compared to other AI detection tools tested, it is more accurate, though like other tools it struggles on certain types of text. It has been used in a number of high-profile accusations of AI use in published writing, and criticized for contributing to "witch hunts" for AI writing.
History Pangram Labs was founded by CEO Max Spero and CTO Bradley Emi in 2023, originally named Checkfor.ai until rebranding in 2024. In 2025, the company raised US$4 million in seed funding in a round led by ScOp Venture Capital. In December, they released Pangram 3, adding the ability to categorize text as partially AI generated. In July 2026, Pangram Labs finished a fundraising round of US$9 million led by Menlo Ventures, with participation from ScOp, Haystack Ventures and others. Shortly afterwards, they launched the fourth version of their text detection model (Pangram 4) and a new AI image detection model (Pangram Image). Between June 2025 and June 2026, Pangram's monthly users increased from 2,700 to 120,000, with annual revenue increasing by a factor of 35.
Description and usage Pangram assesses the probability that a given text is fully written by AI, AI-assisted, or fully written by a human. It uses a transformer-based neural network trained on a mix of human writing and text generated by LLMs to match the human samples' topic, length, and tone. They contrast this technique with some other AI text classifiers, which use statistical measures like perplexity and burstiness. Pangram also breaks down longer documents into sections and classifies each individually, to determine how much of a text is AI generated.
As of August 2026, free users got 2,000 words quota per day on Pangram's website, with a paid $20 per month tier providing 300,000 words a month and AI plagiarism detection. They also provide a Chrome extension and integration with Canvas and Google Classroom. In July 2026, Substack added an AI detection feature using Pangram. The browser extension NewsGuard added similar integration in March the same year. Quora is also a customer. According to The New York Times, "[t]he company has also signed contracts with a number of universities and publishers looking to crack down on A.I. writing."
Efficacy In a 2025 study comparing Pangram to OriginalityAI, GPTZero and a classifier built on the open source model RoBERTa, researchers at the University of Chicago found that Pangram significantly outperformed all the other detectors, having a zero false positive rate (FPR) and near-zero false negative rate (FNR) on longer passages and maintaining FPR and FNR below 0.01 on shorter passages. However, on a different sample from Chatbot Arena Pangram's false negative rate was closer to one in 70. In the University of Chicago study, Pangram also outperformed the other models in detecting AI generated text ran through the "humanizer" StealthGPT, and was cheaper than the other two commercial detectors. In 2025, researchers at the University of Maryland, UMass Amherst, and Microsoft found that Pangram was the only AI detector to match the performance of the majority vote among human evaluators with experience using LLMs. A study published June 2026 by researchers at Vrije Universiteit Brussel found Pangram to be the most accurate among GPTZero, Pangram, Copyleaks, and Turnitin for AI detection in master's theses; it gave a median score of 80% on fully AI-generated papers (varying between 60% and 100%), while the other tools had medians below 20%. Pangram's median score on partially AI-generated papers also "showed closest alignment to the ground truth" among the detectors.
Failure modes
Pangram is in an "arms race" with the developers of large language models, which aim to produce more humanlike prose; the detector tends to perform worse on newer LLMs not included in its training set. It also performs worse when the LLM text has been fed into a "humanizer" program. Like other AI detectors, Pangram was found to misclassify synthetic text as human if the AI was asked to imitate specific human authors. Pangram was also more likely to misclassify AI-generated text as written by humans when it rhymed, repeated itself, or used archaic language; an adversarial set of AI text examples built by researcher Alexios Mantzarlis was incorrectly labelled as human by Pangram 86% of the time. Pangram also has historically struggled to accurately distinguish if a text is partly or wholly AI generated; for example, author Freddie deBoer found adding LLM-generated text to the end of a human-written piece made Pangram treat the whole document as AI written. University of Maryland researchers found that like other detectors, applying "AI polishing" to change small portions of a text causes the likelihood Pangram detects the sample as AI-written to increase dramatically, especially when using older or smaller LLMs. Pangram Labs has said that its newest model (Pangram 4) is more robust to humanizers and adversarial prompting, and is better at distinguishing human and AI parts of a document.
… excerpt ends here. Continue reading the full article.
