Pangram tests a Gmail filter that can trash AI-written email
The tool would scan messages on Pangram's servers, apply AI, Mixed or Human labels, and automate where flagged mail goes.
By Ryan Merket · Published · Updated
Why it matters
Pangram is turning AI detection from an advisory score into an inbox gate. Automatic filtering raises the cost of false positives while requiring server-side access to private email.
Pangram, the AI-detection business founded by Max Spero (@max_spero_) and Bradley Emi (@bradley_emi), is testing a Gmail integration that scans incoming messages and can automatically send email classified as AI-written to spam or trash.
Security researcher and reverse engineer Jane Manchun Wong (@wongmjane), who is known for uncovering unreleased app features, published screenshots of the integration on August 11. The images show a Pangram dashboard connected to Gmail through Google OAuth, with controls for pausing scans, excluding selected senders and domains, and choosing how Gmail handles messages that Pangram labels as AI-generated.
The feature remains under development. Wong said Pangram is "working on" the integration, and the screenshots describe the connection as scanning incoming email on Pangram's own servers. New messages would be checked after Gmail sends Pangram a delivery notification, then marked with Pangram/AI, Pangram/Mixed or Pangram/Human labels inside the inbox.
A detector becomes an inbox rule
The important change is the action attached to Pangram's score. Pangram's existing tools generally let people inspect selected documents, web pages and social posts. The Gmail integration would run continuously and use a classification to determine whether a message remains in the inbox.
One screenshot shows a dashboard summarizing checked, skipped and labeled threads, along with a list of senders responsible for the most AI-labeled messages. Users can establish exclusion rules for senders, domains, groups and mailing lists before Pangram fetches the message text. Another setting lets users move AI-labeled mail out of the inbox.
That design could give users a new way to suppress automated sales pitches, mass outreach and other machine-written messages that pass conventional spam checks. It could also catch legitimate email drafted with an AI assistant. Pangram's own terms of service describe its detection system as a probabilistic classifier and tell users of its text-message detection service to treat each prediction as one signal among many. An automatic inbox rule gives that prediction a direct operational consequence.
Spero and Emi met as Stanford freshmen and later worked on applied machine-learning systems before founding Pangram. Spero worked at Nuro, Google, Two Sigma and Yelp, while Emi led deep-learning research at Absci after working on Tesla Autopilot's computer-vision team. Pangram has built its business around detecting synthetic text for schools, publishers, platforms and other organizations. (Pangram's company history)
The Gmail test follows a rapid expansion beyond Pangram's original document checker. Substack began using Pangram on posts and notes published from July 21, allowing readers to estimate how much text was human-written or AI-assisted. Pangram then announced Pangram 4 and raised $9 million in a July 29 round led by Menlo Ventures, with Haystack, ScOp Venture Capital, Script Capital and Cadenza participating, according to TechCrunch's report.
The inbox raises the privacy stakes
A Gmail connection requires Pangram to process a category of text that can include private correspondence, receipts, account alerts and workplace discussions. The screenshots emphasize Google OAuth, revocable access and sender-level exclusions, while making clear that Pangram performs the scan on its servers rather than entirely inside Gmail.
Pangram's current privacy policy says it collects text submitted for detection, along with metadata for registered customers. Pangram says customer content is not used to train or improve its AI models, is not used for advertising, and can be deleted from query history. A separate data privacy FAQ says submitted content remains stored while an account is active unless the user deletes it.
Pangram's latest technical report says Pangram 4 achieved a 0.0041% false-positive rate and a 0.3396% false-negative rate in its evaluations. Those are Pangram's reported benchmark results, and inbox performance will depend on the length, style and model origin of real email. The report also says Pangram 4 was designed to classify mixed human-AI writing and identify smaller AI-edited passages, a necessary capability when many messages start with a human draft and pass through an AI writing tool before being sent.
The Gmail integration extends Pangram's core bet from identifying machine-written material to enforcing user preferences against it. If released, Pangram would sit between the sender and the recipient, turning an authorship estimate into a delivery decision before the user opens the message.