Do AI-written emails go to spam? What Gmail and Outlook check, and what readers notice

Not under any rule Gmail or Outlook has published. Neither Google's nor Microsoft's rules for senders say anything about how a message was written. They check whether it really comes from you, how many people mark it as spam, whether they can unsubscribe, and whether the subject and headers are honest. As far as those rules show, an email written with ChatGPT or Claude lands in spam for the same reasons any email does. Whether anyone answers it comes down to something no published rule checks: whether it was written to them.

What Gmail and Outlook check

Google's sender guidelines ask every sender for SPF or DKIM, valid forward and reverse DNS, a TLS connection and a reported spam rate below 0.3%. Senders of more than 5,000 messages a day to Gmail accounts also need both SPF and DKIM, plus DMARC, a From domain that aligns, and one-click unsubscribe on marketing mail. The content rules are about honesty: no misleading subject, and no "Re:" on a message that is not a reply.

Microsoft's rules for high-volume senders to Outlook.com, announced in April 2025 and enforced from May, require SPF, DKIM and DMARC, and recommend a working unsubscribe, a clean list, accurate subject lines and consent. Neither page mentions AI.

Both cover personal mailboxes. Google's apply to @gmail.com addresses, not to Google Workspace accounts, and Microsoft's to Outlook.com, Hotmail and Live addresses. A cold email to a company address goes through that company's own mail filters instead, and SPF, DKIM and DMARC are how any receiving server checks that a message comes from who it says. The free email domain checker reads all three.

Can Gmail detect AI writing?

Google's sender guidelines say nothing about it, and we found no Google statement that Gmail tries. Filters read the text of a message, and more of what they read is machine-written: Barracuda, working with researchers at Columbia and the University of Chicago, estimated that 51% of spam was AI-generated by April 2025. Microsoft's threat intelligence team, writing about an AI-assisted phishing campaign, said its detection signals are largely unaffected by whether a human or a model wrote the attack code, though AI-written code can leave traces that become signals themselves.

Recipients can now filter it themselves. Pangram's Gmail add-on, launched in September 2026, labels messages it judges AI-written, and its users can choose to archive them or send them straight to spam. The same report notes that detectors can flag human writing by mistake.

In our view, the risk with AI-written outbound is the volume it makes cheap: one email to thousands of people who did not ask for it, and the complaints that follow, which is what the rules above measure. At Google's line, three complaints in a thousand messages is 0.3%.

Can people tell an email was written by AI?

From the writing, mostly not. In six experiments with 4,600 participants, people could not reliably tell AI-written self-descriptions from human ones: in the three main experiments they were right 50 to 52% of the time, close to chance, and they leaned on the wrong cues, such as contractions and first-person pronouns (Jakesch, Hancock and Naaman, PNAS 2023). The study used online profiles written by GPT-2 and GPT-3, not emails, so read it as a direction.

What readers can tell is whether it was written to them. An email that could have gone to anyone reads like one, whoever wrote it, and the old trick for hiding that has worn out. In randomized field experiments with three companies that emailed millions of people, adding the recipient's name to the subject line lifted opens in the main experiment from 9.05% to 10.80% (Sahni, Wheeler and Chintagunta, Marketing Science 2018). A 2023 replication found no positive effect from the first name, while the surname did have one, and its authors suggest people have grown used to the tactic (Defau and Zauner, Marketing Letters).

A first name shows only that a database had a row for you. A dated fact about your company shows someone read about you, and a model can put a true one in an email only if it is given one or finds one.

What to do instead

• Authenticate the domain you send from, and keep the complaint rate where Google can see it. The deliverability checklist has the order.
• Send to a list you can write for. Fifty to two hundred named accounts, each with a reason, is a list a person can read every email for (how to build one).
• Give the model a person. A research note with three sourced facts turns a draft for a job title into a draft for one reader. The method is in how to write cold emails with ChatGPT and Claude for sales prospecting.
• Read every draft before it sends, and look for the merge tag that did not fill. The free mail-merge preview shows every message as it will arrive.

Questions

Does Gmail penalize emails written with ChatGPT?

No published Gmail rule says so, and nothing Google has published treats them differently from any other email. What the rules judge is the sender, the complaints and the headers.

Is it bad to send AI-generated emails?

Not in itself. The email that hurts is the one sent unread, at volume, to people it was not written for, and that was true before AI wrote any of it.

How do I keep cold emails out of spam?

Authenticate the domain, keep the list small enough that people expect to hear from you, and make unsubscribing easy. The deliverability checklist runs the checks in order.