Skip to content
September 2026

Does AI-Written Email Get Filtered? Not for Being AI-Written

Every published requirement from Google and Microsoft is about authentication, complaint rates and volume. None of them mentions who or what wrote the message.

Here's the trap. You start using an assistant to draft outreach, your reply rate drops, and somebody tells you the spam filters are detecting AI text. It sounds plausible, it is the kind of thing that would be true, and you start rewriting everything by hand.

No mailbox provider publishes any requirement about who or what wrote the message. Google's and Microsoft's bulk sender rules are about authentication, complaint rates, unsubscribe handling and volume. Authorship is not on the list.

That does not mean nothing changed when you started using AI. Something almost certainly did, and it is worth being precise about what, because the fix is different.

What do the providers actually require?

The bulk sender requirements Google and Yahoo introduced in February 2024, with Microsoft following for Outlook.com from May 2025 and tightening through 2026, come down to a short list:

  • Authenticate properly. SPF, DKIM and DMARC, aligned. This is the one that actually stops mail arriving, and it is a configuration problem rather than a content one.
  • Keep spam complaints below 0.3 percent. That is the published Google threshold. Complaints are recipients pressing the button, which is a judgement about your message rather than about its provenance.
  • Offer one-click unsubscribe, implemented to RFC 8058, and honour it quickly.
  • Send from a domain with a reputation. Which new domains do not have, which is why warmup exists.

Read that list again and notice what is absent. There is no content-authorship criterion, no AI-detection step, and no published mechanism by which a filter would even establish authorship of a short business email.

A figure circulates that Microsoft tightened to a 0.10 percent complaint threshold and roughly 200 sends a day for new domains during 2026. We have not been able to trace it to Microsoft's own documentation, so we are carrying it as unverified rather than repeating it as a requirement. Treat the direction as real and the specific numbers as unconfirmed.

So why did the replies drop?

Almost certainly volume, and the shape of what volume does to the criteria above.

An assistant removes the writing constraint. The writing constraint was, accidentally, also the quality constraint: when each email cost ten minutes, nobody sent to a segment they had not thought about. When each email costs ten seconds, the list gets longer before it gets better.

A longer, less considered list produces more complaints, and complaint rate is an actual published threshold. So the causal chain runs from AI to volume to relevance to complaints to filtering, and every link in it is behavioural. The filter never formed an opinion about your prose.

There is a second, duller possibility worth ruling out first: that your authentication was never right and the increase in volume simply made it visible. That is a twenty-minute check and it is where we would start.

Does the writing matter at all, then?

Yes, but through recipients rather than through filters, which is a distinction with practical consequences.

A generic email does not get intercepted by a machine that recognises generic emails. It gets ignored, and some proportion of the people ignoring it press the spam button because that is faster than unsubscribing. Enough of those and you cross a published threshold.

So the advice that comes out of the folklore is roughly right for the wrong reason. Write better emails, yes, but not to fool a detector. Write them so that fewer people report you, because the report is the thing the provider actually measures.

What about tools that humanise AI text?

A small industry has grown up selling software that rewrites model output to evade AI detection. For email, it is solving a problem nobody has demonstrated exists.

AI-detection tools are a real thing in academic and publishing contexts, where the question is whether a person wrote a long document. Nothing in any mailbox provider's published requirements suggests they run detection on a ninety-word business email, and the shorter the text the less reliable any such classifier would be anyway.

If a humanising step improves your results, the plausible explanation is that it made the email less generic and therefore less likely to be reported, which is the same mechanism as writing it better in the first place. You can get that from a constraint block in the original prompt without adding a tool.

What should you actually check?

  • Your authentication, first and properly. SPF, DKIM and DMARC aligned on every sending domain. Most deliverability problems attributed to content are this.
  • Your complaint rate against 0.3 percent. If you cannot see it, that is the more urgent problem. Google's Postmaster Tools shows it for free.
  • Whether your volume changed when your tooling did. Compare sends per week before and after. If it doubled, you have your explanation and it is not authorship.
  • Whether your list got worse as it got longer. The honest test is whether you could still say, for a random row, why that person specifically. If not, that is what the complaints are measuring.
  • Your sending infrastructure, if you are ramping. New domains have no reputation and warmup is not optional; we itemised what that costs in the sending model.

The answer to the question in the title is no: nothing published by any mailbox provider filters on whether a human wrote the message. What gets AI-assisted outbound filtered is the volume the assistance made possible, and the fix for that is a shorter list rather than a more human-sounding paragraph.

Questions

Do spam filters detect AI-written emails?
Not according to anything Google, Yahoo or Microsoft publishes. Their bulk sender requirements cover authentication, spam complaint rates, one-click unsubscribe and domain reputation. No published criterion concerns who or what wrote the message.
What actually gets cold email filtered?
Failed authentication, a spam complaint rate above the published threshold of 0.3 percent, missing or ignored unsubscribe handling, and sending volume from a domain with no reputation. All four are behavioural and none is about content authorship.
Why did my reply rate drop after I started using AI?
Most likely volume. Writing was accidentally acting as a quality control, because a ten-minute email meant nobody sent to a segment they had not thought about. A longer, less considered list produces more complaints, and complaint rate is a real published threshold.
Is there a Microsoft 0.10% complaint threshold?
It circulates widely alongside a figure of about 200 sends a day for new domains, and we have not traced either to Microsoft's own documentation. Treat the tightening direction as real and the specific numbers as unconfirmed until you see them from Microsoft.
Should I stop using AI to write outreach?
Not for deliverability reasons. Use it, and keep the list short enough that you could still say why each person is on it. The failure mode is not the writing, it is that removing the writing cost also removes the thing that used to stop you sending to people you had not considered.

Tools mentioned

All tools

Sources

Source interests are labelled. Almost everything published about this subject is written by someone selling into it.

More from the blog

All posts