Pangram’s Gmail Test Split 108 Threads: 8 AI, 37 Human, 63 Mixed

In my test, Pangram’s $20-per-month Gmail integration labeled 8 of 108 recent threads as AI, 37 as human and the rest as mixed.

Pangram makes an AI text and image detector. It now offers a browser extension inside Gmail that runs incoming messages through algorithms designed to detect AI writing. The Gmail option requires at least the $20-per-month individual plan, and Pangram does not offer the integration on its free tier.

The product targets a problem created by the spread of generated writing. AI can now produce restaurant menus, student essays, full-length movies and social posts, as well as messages sent to friends, clients and colleagues.

That shifts the stakes inside an inbox. A newsletter may simply feel automated, but a machine-written note from a friend or client can make trust harder to judge.

I connected the service through Pangram’s Gmail integration page. The same option appears as Gmail Integration in the web dashboard’s left navigation. After clicking the button labeled “Connect your Gmail,” I approved the standard authorization screens.

Pangram says, “The contents of your emails aren't stored after being assessed and won't be used to train models designed to look for AI text.” That policy does not remove the need for trust, but it narrows what Pangram says it retains and trains on.

The connection remains reversible. Users can remove the Pangram plug-in from Gmail at any point, although that action does not automatically delete any labels the service has already applied.

What does Pangram scan in Gmail?

After I connected my account, Pangram scanned the 100 most recent email threads. It then continued checking new messages as they arrived. It cannot inspect every email I have ever received.

Those automatic Gmail checks did not appear to consume the monthly scan credit limit, as far as I could tell. The feature therefore covered both the initial backlog and the ongoing flow of incoming mail without reducing the credits listed for that month.

The Gmail Integration view then summarized the results in three buckets:

• Suspected AI
• Most probably human
• Mixed

I could apply those categories as Gmail labels, but doing so was optional.

Labels persist after disconnection, which is worth knowing before anyone enables them. If I removed Pangram later, the labels would remain in Gmail, and I would have to delete them myself.

Pangram could also route incoming AI-written mail straight to the archive or spam folder. That is handy for keeping suspected machine writing out of sight, though the choice remains with the user.

I could exclude four kinds of traffic from the scan:

• Specific domains
• Gmail labels
• Subject keywords
• Individual senders

Those controls could spare a boss, a relative or another sensitive conversation from scrutiny, while preserving automated checks for the rest of the inbox.

In practice, the integration turns an ordinary inbox into an ongoing triage system, sorting conversations according to whether Pangram thinks a machine wrote them, a person wrote them, or both kinds of signals appear.

How did the detector perform on a real inbox?

The finished scan covered 108 threads. Eight received the AI label, 37 received the human label, and the remaining 63 were marked as mixed. Press releases were the most common source of the AI-labeled messages.

Counting the mixed threads, more than half of the scanned conversations contained at least some suspected AI writing. That result felt dispiriting.

The more encouraging finding involved messages I knew were human. Pangram did not flag any email from friends, family or colleagues as fully or partly AI. Every one of those messages passed its human test.

That result comes with a major qualifier: this was a small personal sample, not a controlled benchmark. It shows that the detector performed well on messages whose authorship I could verify, but it cannot establish a broader accuracy rate.

I also sent two test messages to myself from separate accounts. One used Microsoft Copilot, and Pangram correctly marked it as AI. The other was a section of this article, written by my own brain and keyboard taps, and Pangram marked it “Human.”

The controlled checks produced the right labels in both directions. They also showed the practical value of checking known authorship against the tool, rather than treating a single score as a universal verdict on machine writing.

What can Pangram miss in a thread?

Short messages and later replies create two different blind spots. Pangram refuses to classify an email below its unspecified minimum length, while the Gmail setting examines only the opening message in each conversation.

The minimum exists because AI is harder to detect with a smaller sample size. The thread rule creates a sharper problem: a correspondent can switch to AI halfway through, and Pangram will not spot that change.

These limits matter because the dashboard condenses each thread into a single label. A mixed label signals uncertainty, but a human label cannot tell users how much evidence stood behind it.

The service runs quietly in the background, and I like the way it stays out of the way while it works. I could see myself using it long term, mainly because the controls are easy to reach and the scan continues without manual action.

It also changes how a message feels. Knowing that a note was assembled entirely by AI can alter the way I read it. That reaction is the feature’s real effect.

Is the $20-per-month Gmail plan worth it?

For people who want a running view of machine-written email, the plan offered a credible second opinion in my test. It handled 100 recent threads, kept watching new mail, applied labels and offered archiving, spam routing and exclusions without obvious friction.

The $20-per-month price was the main barrier, because the feature remains unavailable on the free tier. The broader Gmail privacy claim also depends on user trust, even though Pangram says assessed email contents are neither stored nor used to train its AI-text models.

The detector’s blind spots are structural rather than temporary. It skips short messages, reads only the opening email in each conversation, and can leave labels behind after removal, so users still need to understand what each verdict does and does not establish.

With Gmail already able to generate an AI reply, Pangram’s next challenge is keeping its judgment useful as generated messages become routine, without presenting a probabilistic label as proof of who actually typed the words.