Every few months, someone claims Gmail can “detect AI-written email” and penalize it outright. It’s a tidy story, and it’s mostly wrong — or at least, wrong about the mechanism. What actually happens is less dramatic and more useful to understand: spam filters in 2026 are built almost entirely on authentication, sender behavior, and recipient engagement. Writing style is, at most, a weak and indirect signal that flows through those same channels.

What Gmail and Outlook actually check

Gmail’s published sender guidelines are explicit about what determines inbox placement: SPF, DKIM, and DMARC alignment; consistent sending volume rather than sudden bursts; a low spam complaint rate; and recipient engagement — opens, replies, and whether people delete a message without reading it. None of that is about the words in the email. It’s about whether the sender behaves like a legitimate correspondent and whether recipients treat the mail like something they wanted.

Outlook’s SmartScreen filtering works on a comparable set of signals: sender reputation, authentication, and — more narrowly than Gmail — complaint data via the Bulk Complaint Level score, which measures how often recipients mark a sender’s messages as junk. Content matters here too, but mostly for structural red flags (broken HTML, mismatched links, spam-trigger phrasing) rather than an assessment of whether a human or a model wrote the sentence.

Neither published system describes a classifier whose job is “detect AI-authored prose and penalize it.” That specific claim doesn’t have first-party documentation behind it, on either platform, as of 2026.

That’s worth sitting with, because it cuts against the intuition a lot of senders have. It would be reasonable to assume a company that builds large-language-model detectors for academic integrity tools could just as easily point one at incoming mail. What the published guidance describes instead is closer to a credit score than a lie detector: providers track how a sender and their content have behaved over time, and a new signal — even a striking one, like a sudden switch to AI-drafted copy — mostly matters to the extent it changes engagement, not because the filter recognizes the tool that produced it.

Where AI-written copy can still hurt you — indirectly

The nuance that gets lost in the “does AI trigger spam filters” debate: generic AI-written copy doesn’t get flagged for being AI-written. It gets flagged for the same reasons generic human-written copy gets flagged — because it performs badly on the engagement signals filters actually measure.

Low-effort, unpersonalized email at volume tends to produce exactly the behavior that erodes sender reputation: fewer opens, fewer replies, more deletes without reading, and — the signal that matters most — a higher spam complaint rate. Spam complaint rate above roughly 0.3% is the threshold where most major providers start meaningfully suppressing a sender, regardless of what produced the copy. A prompt that outputs the same generic paragraph for every recipient is a fast way to get there, not because a model wrote it, but because nobody who receives it feels like it was written for them specifically.

This is the actual mechanism, and it matters because it points to the right fix. The fix isn’t “write it yourself instead of using AI.” The fix is the same one that’s always applied to outreach at scale: real personalization, real relevance, and a low enough volume-to-quality ratio that recipients don’t feel like they’re getting spam-blasted — why emails go to spam covers the fuller list of causes, and authorship isn’t one of them. Prompting for copy that doesn’t read as generic AI output is worth doing for the reply rate on its own; that it also keeps you further from the complaint-rate threshold is a side benefit, not the main reason to do it.

The attacker side complicates the narrative

Part of why “AI detection” gets brought up so often is that AI-generated phishing and spam volume has genuinely grown, and providers have responded by investing more in classifiers like RETVec. But that investment is aimed at the adversarial-manipulation and malicious-intent side of the problem — content designed to defraud, obfuscate, or evade filters — not at flagging legitimate B2B outreach for the crime of being drafted with AI assistance. A well-targeted, honestly personalized cold email written with AI help and a well-targeted, honestly personalized cold email typed by hand should land in the inbox at comparable rates, because the filter isn’t evaluating which one it’s looking at.

Why the myth persists anyway

Part of the reason “Gmail detects AI” keeps circulating is that the timing lines up in a way that feels like causation: a team adopts AI drafting, sees reply rates drop a few weeks later, and connects the two events. What usually happened in between is less mysterious — the AI-assisted campaign was sent to a broader list, with less per-segment customization, than the hand-written campaigns it replaced. The reply-rate drop tracks the drop in relevance, not the switch to AI. Attributing it to a filter that specifically targets AI writing is a simpler story, but it’s not the one the evidence supports.

The other reason is that “the algorithm is against me” is a more comfortable explanation than “the copy wasn’t good enough for this audience.” Both explanations produce the same feeling of being filtered unfairly. Only one of them points at something you can actually fix.

What this means for how you use AI to draft

Run every draft, AI-assisted or not, through the same checks: does the personalization reference something true and specific, is the volume reasonable for how relevant the message actually is, and does the copy avoid the structural red flags — broken links, spam-trigger language, missing authentication — that filters do check. The spam word checker catches the mechanical version of that last check before a draft goes anywhere.

A quick self-check, before any AI-assisted draft goes out at volume:

Does this reference something true and specific about the recipient? If the same paragraph could go to every name on the list unchanged, the filter risk isn’t that it’s AI — it’s that it’s generic, and generic gets reported.
Is the authentication actually passing? SPF, DKIM, and DMARC alignment matter more than any wording choice, and they’re the first thing worth verifying if a campaign is underperforming.
Is the volume proportional to how warmed-up the sender is? A sudden jump in send volume looks the same to a filter whether a human or an agent decided to make it.
Would a recipient recognize this as relevant within the first line? If not, the risk is a delete-without-reading, and enough of those is its own reputation signal regardless of what wrote the email.

None of these checks ask “was this written by AI.” They ask the questions a filter is actually built to answer.

The honest summary: spam filters are not judging your prompt. They’re judging your sending behavior and your recipients’ reactions to it. AI can make either one better or worse depending on how it’s used — faster research and drafting if you keep a human checking relevance, or faster generic-content-at-volume if you don’t, which is the same trade-off worth weighing before giving an agent broader sending authority. The filter doesn’t care which path you took. It only sees the result.