Pangram Labs

Market position

Positioned as an accuracy-led alternative to incumbent AI detectors such as GPTZero, Originality.ai, Copyleaks and Turnitin. Independent academic studies at the University of Chicago, the University of Maryland (with UMass Amherst and Microsoft) and Vrije Universiteit Brussel ranked Pangram ahead of the other tools tested, including on AI text processed through humanizers and on master's theses.

Claims of substantially lower false positive rates than competing detectors (roughly 1 in 10,000 to 1 in 24,000), section-level attribution that separates AI-assisted from fully AI-generated writing, robustness to humanizer tools, multilingual coverage, and third-party academic validation rather than open-source or off-the-shelf LLM-based classification.

Technology

A proprietary classifier: a transformer-based neural network trained on known-human text (including material predating 2021) paired with synthetic text generated to match human samples, using diverse datasets, hard negative mining and active learning rather than perplexity/burstiness heuristics. Documents are segmented and scored individually to distinguish fully AI, AI-assisted and human writing. The company benchmarks against 26 named models and reports a false positive rate of 0.0041% (about 1 in 24,000 documents) for Pangram 4 in internal testing, and 99.5% accuracy for the image model in internal benchmarks.

Go-to-market

Self-serve free trial and paid subscriptions on pangram.com plus a Chrome extension (around 20,000 users), combined with enterprise sales to institutions and embedded partnerships with platforms such as Substack and NewsGuard. The company publishes benchmarks and a model card and cites third-party academic evaluations as evidence of accuracy.

Universities, schools and teachers; publishers, media organizations and journalists; content moderation teams and online platforms; enterprises, law firms, recruiters and developers; and individual consumers.

Geography

Headquartered in Brooklyn, New York, United States, serving universities, schools and enterprises internationally with detection across more than 20 languages.

History

Launched in October 2023 under the name Checkfor.ai and rebranded to Pangram in 2024. The company raised seed funding in 2025 in a round led by ScOp Venture Capital, released Pangram 3 in December 2025 with partial-AI classification, and in July 2026 closed a $9 million round led by Menlo Ventures before launching Pangram 4 and an AI image detection research preview.

Risks & controversies

AI detection is widely regarded as a probabilistic signal rather than proof, and several institutions have restricted or discontinued detector use; MIT Sloan and the Mozilla Foundation have highlighted high error rates and bias against non-native English speakers in the category generally. Pangram's accuracy claims rest largely on internal benchmarks, a technical white paper and a small number of favorable third-party studies. Reported failure modes include misclassifying AI text that imitates specific authors (about 10% per Epoch AI), poor performance on rhyming, repetitive or archaic AI text (an adversarial set assembled by researcher Alexios Mantzarlis was labelled human 86% of the time), degraded accuracy on newer models and humanized text, higher AI scores after light "AI polishing," and difficulty separating partly from wholly AI-generated documents. The tool has featured in high-profile AI-use accusations: Wall Street Journal editor James Taranto called it a "defamation machine" after three of his writers' articles were flagged, and commentators in Slate and The Atlantic have criticized the resulting "callout" culture and the reputational risk of errors. The New York Times said in August 2026 that Pangram excels at text but is not reliable for AI images.