I Scanned My Gmail for AI Writing. The Results Need Context
AI writing is easy to spot in a polished press release, but it's harder to judge in the middle of an ordinary email thread. A message could be drafted by a person and cleaned up with an AI assistant, pulled from a template, or stitched together from multiple sources. In an inbox full of newsletters, work updates, and personal messages, you usually can't tell - which is exactly why I tried a detector.
Pangram, an AI text detection service, added a Gmail integration that scans incoming messages and labels them as AI, human, or mixed. I ran it on a small set of email threads and got results that were interesting - but also a reminder that an AI detector shouldn't be treated like a verdict. In my test, most messages landed in the mixed category, and a small personal sample can't tell you how accurate the tool is across other inboxes, other writers, or other writing styles.
This guide explains what the integration scans, how its labels and settings work, what my inbox scan showed, and how to use detection without overreacting. The real question isn't whether a tool can catch every AI-written email. It's whether its signal is good enough to change how you manage your inbox.
What does Pangram's Gmail integration do?
Pangram's Gmail integration is a browser-based feature that checks email text for signs of AI-generated writing. After you connect it, it scans the 100 most recent email threads and then keeps checking new incoming messages. It doesn't do a full "scan my whole mailbox" sweep on demand, so think of it as ongoing triage rather than a forensic audit of everything you've ever received.
In the dashboard, scanned threads fall into three buckets: AI, human, and mixed. Gmail labels can also be applied to messages, but that part is optional. Even with these categories, you still don't get a clear breakdown of how much text was machine-generated - or which specific sentences might have been touched by a tool.
That distinction matters. Calling something "AI" is a result of a classification system, not proof of the sender's process. Someone might use an assistant to edit a draft without handing over the whole message. A company might reuse templated language across many emails. And sometimes you simply don't have enough text for a detector to make a meaningful call. The mixed label is probably the most honest outcome here: it suggests the text doesn't fit cleanly into one side or the other, and it shouldn't be treated like a precise percentage breakdown.
The interface also gives you options for what to do with the results. You can route emails flagged as AI straight to the archive or spam. You can also exclude specific domains, senders, Gmail labels, and subject keywords from scanning. Those controls matter if you're trying to avoid processing sensitive messages or you don't want a particular sender's emails analyzed.
This feature is part of a paid Pangram plan, not the free tier. In the test described here, Gmail access required the individual plan priced at $20 per month. Pricing and terms can change, so it's worth checking before you decide it's worth paying for. The main point isn't the label itself - it's that you get a lightweight classification step added to incoming mail.
How do you set it up and control what it scans?
The setup is simple, but connecting any third-party tool to your email deserves a little caution. From Pangram's dashboard, you open the Gmail integration option and connect your account. Gmail then shows authorization screens - so don't just click through. Read the access request and decide whether scanning your inbox makes sense for the account you're connecting.
Once authorized, Pangram scans the 100 most recent threads and then processes new incoming messages. In the version I tested, the scan didn't seem to use monthly scan credits. That means the feature could be left running without hitting a visible rationing limit, but you should still check the current plan terms rather than assume the billing behavior won't change later.
You can keep results in Pangram's dashboard or turn on Gmail labels for the three categories. Labels are a good middle ground if you want to review what it's doing without changing your workflow. Automatic archiving or spam placement is riskier. If the detector is wrong, it can hide messages you actually wanted to see. If you're trying the tool for the first time, I'd start with labels only, watch what it does, and only automate handling if the results genuinely prove useful in your inbox.
There are also exclusion controls for domains, individual senders, Gmail labels, and subject keywords. Use them to draw a clear boundary. For example, you might exclude messages from your employer, a client, a doctor, or family. It's not that those messages are automatically "more sensitive" than everything else - it's that broad inbox access should be a conscious choice, not something you accidentally enable by leaving defaults in place.
Pangram says email contents aren't stored after assessment and aren't used to train models designed to detect AI text. That's reassuring as far as their stated approach goes, but it still comes down to trust and permissions. If your account contains confidential business information, regulated data, or client material, you should check the privacy terms and your organization's policies first. You can disconnect the integration later, and any Gmail labels it applied can be removed manually. Disconnecting a tool helps - but it doesn't replace paying attention to what access you originally granted.
What did the inbox scan find?
In my test, Pangram scanned 108 threads. Eight were labeled AI, 37 were labeled human, and the remaining 63 were marked mixed. Press releases showed up most often in the AI category. That's plausible - corporate announcements tend to use polished, standardized language, and companies are increasingly using writing tools for drafts and revisions. What the scan didn't do is tell me why any particular sender wrote one specific email in one specific way.
The numbers are also easy to misread. Only eight threads got the fully AI label. Those 63 mixed results aren't "63 confirmed AI-written emails." They mean the detector didn't place those messages cleanly into either the AI or human buckets. Without a deeper explanation of how Pangram calculates labels, converting "mixed" into a confident claim about authorship would overreach.
The personal checks were encouraging, but they're still limited. Messages from friends, family, and colleagues - people I knew wrote their emails themselves - were labeled human. A self-sent message drafted with Microsoft Copilot came back as AI. A passage I wrote manually was labeled human. Those examples suggest the detector could pick out a few known cases in this small sample. But that's not the same thing as an accuracy rate, because this wasn't a controlled evaluation across different writers, languages, and email formats.
There are also practical coverage limits. Pangram needs enough text to make a classification, so very short messages may never get a label. That's reasonable: with just a few words, there's not much signal to analyze. Also, the integration scans only the first email in each thread. If someone changes how they write later in an ongoing conversation, the detector may miss that shift.
Overall, I'd treat these results as a snapshot of one inbox - not a measurement of how much AI people use in email. Still, they're useful for asking better questions: Which kinds of messages get flagged? Do the results stay consistent for senders I know well? If I see a label, do I actually change how I respond? What they aren't good for is accusing an individual of outsourcing their writing.
Can you trust an AI detector in your inbox?
Whether you can "trust" it depends on what you're trying to do. If you're looking for a low-stakes sorting aid, it might help you spot patterns in bulk email or decide which messages you want to review more closely. If your goal is to prove that a person used AI, then no - detection isn't strong enough for that. These systems infer things from the text. They don't observe how the message was created, and a label can't reliably separate full generation from editing, templates, or just a sender's naturally polished style.
This matters even more in business contexts. A company might draft routine customer updates with an AI assistant, then have an employee review and send them. In that situation, the email can still be accountable human communication, even if an AI helped produce the first draft. On the flip side, a fully human-written message can sound generic and get flagged anyway. If you care about service quality, accuracy, or whether someone fulfilled an obligation, you should evaluate the content and the process - not a detector label in isolation.
Privacy is a separate issue from accuracy. Email is a dense record of personal and business information. Before turning on an integration, think about what it can access, which messages you should exclude, what the provider says about retention and model training, and whether your workplace allows third-party processing. Pangram says it doesn't retain email contents after assessment and doesn't use them to train detection models. Even so, if your organization has stricter requirements, it's worth reviewing the exact terms and internal policies.
A cautious rollout also reduces the downside of being wrong:
- Start with a personal or low-sensitivity mailbox, not a shared inbox full of confidential material.
- Use labels instead of automatic archiving or spam placement while you learn how the tool behaves.
- Exclude senders, domains, and labels that shouldn't be scanned.
- Compare its classifications with messages where you truly know the writing process.
- Treat AI and mixed labels as prompts for review, not evidence for a personnel or customer decision.
- Reassess after a few weeks. If it adds noise or risk without saving time, turn it off.
The broader lesson is simple: AI detection is most defensible when the cost of a wrong label is low and you keep a human in charge. An inbox filter can help you focus. It shouldn't become an authorship tribunal. If you need to know whether a message was AI-assisted for a compliance or contractual reason, you're better off setting up a clear disclosure process with the sender rather than trying to infer intent from prose.
Conclusion: Treat detection as a signal, not a verdict
Pangram's Gmail integration adds AI classification into the inbox workflow. It scans recent threads and incoming mail, applies optional labels, and can filter messages based on your settings. In my test, it did classify a few known examples correctly - but most threads ended up in the mixed category, and the sample was too small to prove broad accuracy.
For me, the best use case is modest: watch for patterns, reduce repetitive triage, and keep a human review step when mistakes actually matter. Start with limited access. Avoid automatic filtering until you've seen how it behaves. And don't treat a label as proof of authorship.
If those boundaries make the feature useful, keep it. If they don't, it may just be adding complexity where you wanted clarity.