Pangram

Pangram was founded in Brooklyn with Max Spero as CEO and Bradley Emi as CTO. They met in a freshman dorm at Stanford before pursuing applied AI work in industry. Spero led active-learning work at Nuro and had worked at Google, Two Sigma and Yelp. Emi led deep-learning research at Absci and worked on computer vision for Tesla Autopilot.⁠[1][2]

The company announced more than $2.7 million in additional financing, bringing pre-seed and seed funding to $3.98 million after an earlier $1.25 million pre-seed round. ScOp led the new round, joined by Script Capital, Cadenza and individual investors; Haystack had led the pre-seed financing. Pangram planned to expand higher-education partnerships alongside business, media and review-authenticity applications.⁠[3]

Pangram released two lightweight models based on EditLens research, using Llama 3.2 3B and RoBERTa-large, together with weights, training data and source code. They distinguish human-written, AI-generated and AI-edited text and are offered for noncommercial use under CC BY-NC-SA 4.0. The company presents them as research baselines and advises against using these open versions to enforce educational or workplace AI policies.⁠[4]

Substack launched Pangram scanning on the web and iOS for posts, notes, replies and comments longer than 100 words published from that day onward. Analysis is shown to the person requesting it. Authors can scan drafts, explain their process and report or remove scans they consider mistaken. Substack noted that detection cannot establish how much human thought went into a text or whether AI was used only as a source.⁠[5]

Pangram announced $9 million in financing led by Menlo Ventures, with Haystack, ScOp, Script Capital and Cadenza participating. In Menlo’s investment account, Deedy DasDeedy DasDeedy Das, also known as Debarghya Das, is a partner at Menlo Ventures investing in AI, infrastructure and enterprise software. Previously an engineer at Facebook and Google and an early builder of Glean and Glean Assistant, he helped launch the Anthology Fund with Anthropic.Open the full entry identified the investment as part of the Anthology FundAnthology FundAnthology Fund is a $100 million AI investment initiative launched by Menlo Ventures and Anthropic in 2024 and financed by Menlo. It backs early-stage AI companies with capital, Anthropic resources and company-building support, with scope for follow-on investments.Open the full entry and framed the company’s role as helping readers and platforms identify generated content.⁠[6][7]

Pangram 4 succeeded version 3.3.2 with finer-grained mixed-authorship analysis and detection of text modified by evasion tools. Its model card describes a sparse mixture-of-experts language-model backbone that labels segments Human, AI-Assisted or AI-Generated, alongside confidence levels and document fractions. It accepts natural-language prose of at least 50 words, primarily in complete sentences. The reported confidence levels are not calibrated probability estimates.⁠[8]

CTO Bradley Emi described evaluations using large held-out datasets, manually checked samples and challenging cases, with active learning, hard-negative mining and threshold calibration used to improve the models. The company’s roughly one-in-10,000 false-positive claim reflects its evaluation distribution; performance varies with genre, length and language. Aggregate figures do not transfer directly to short answers, outlines, poetry, recipes or highly templated material.⁠[9][8]

Max Spero published guidance explaining that the displayed AI percentage describes the share of text flagged, rather than the probability that a judgment is correct; users should also inspect segment confidence levels. He recommends discussing results alongside writing history, drafts and course rules, with further investigation before sanctions. Detecting AI editing does not by itself establish a violation of a course policy.⁠[10]

Pangram offers web scanning, browser extensions, Google Docs and learning-management-system integrations, APIs, image detection and similarity checking. Text classification learns differences between human and model writing rather than querying an author’s chat history; similarity checking is a separate feature. In a public interview, Spero acknowledged that the deep-learning classifier remains difficult to interpret fully and that displayed linguistic clues help users understand the result without completely reproducing the model’s decision process.⁠[1][11]

Website: https://www.pangram.com/ (opens in a new window). X: https://x.com/pangram (opens in a new window). LinkedInLinkedInLinkedIn is a social networking platform for professionals. It was founded in 2002 by Reid Hoffman and others, launched in 2003, and listed on the New York Stock Exchange in 2011; after Microsoft acquired it for approximately $26.2 billion in 2016, it became a wholly owned Microsoft subsidiary. Its revenue comes from membership subscriptions, advertising, and recruitment solutions.Open the full entry: https://www.linkedin.com/company/pangramlabs/ (opens in a new window).⁠[2][12][13]