ABAB Pedia

Pangram

Pangram, also known as Pangram Labs, is an AI content detection company founded in Brooklyn, New York, in 2023 by Max Spero and Bradley Emi. Its products classify text as human-written, AI-assisted or AI-generated and include image detection, browser extensions and APIs for education, publishing and content platforms.

Contents11 sections
Key facts

founding and team

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]

$3.98 million in pre-seed and seed funding

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]

Open Pangram release

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 integration

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]

$9 million round led by Menlo

Pangram 4

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]

false positives and usage limits

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]

guidance for educational use

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]

Topics: products and classification method

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]

Official website and public accounts

Sources

  1. Pangram — AI content detection (opens in a new window)
  2. About Pangram (opens in a new window)
  3. Pangram Closes $4 Million in Seed Funding for AI Detection Technology (opens in a new window)
  4. Introducing Open Pangram (opens in a new window)
  5. Against Claudefishing (opens in a new window)
  6. Pangram Labs raises $9M to launch more accurate AI detection for text and images (opens in a new window)
  7. Investing in Pangram to Stop AI Slop on the Internet (opens in a new window)
  8. Pangram 4 Model Card (opens in a new window)
  9. All About False Positives in AI Detectors (opens in a new window)
  10. What to do when a student submission is flagged as AI (opens in a new window)
  11. Q&A: Pangram CEO (opens in a new window)
  12. Pangram (@pangram) (opens in a new window)
  13. Pangram Labs | LinkedIn (opens in a new window)