I want to start with what I actually know, from my own work, rather than what the studies say — because the studies and the lived experience do not always land in the same place.
Across the audits I have done and the conversations I have had with other deliverability specialists, a pattern keeps showing up: programmes that have shifted heavily to AI-generated email copy are seeing engagement drop and inbox placement get worse. Not always dramatically. Not always immediately. But consistently enough, across enough accounts, that it is not a coincidence.
That is my starting point. Now let me explain what is driving it, what the research confirms, and what else is happening in the inbox — because AI is affecting deliverability in ways beyond just the copy question, and some of those ways are more significant than most people realise.
The question I get asked most often is whether Gmail, Outlook, or Yahoo are detecting AI-generated copy and sending it to spam. The short answer is no, not in the way people imagine. Spam filters are not running your email through an AI detector. There is no switch that says "ChatGPT wrote this, spam folder."
What spam filters are doing, and have always done, is measuring engagement. Who opens, who clicks, who replies, who moves the email out of spam, who marks it as spam, how quickly people interact after receipt. Your sender reputation is built from those signals, accumulated over time, across your whole sending programme. It is a trust score, and it reflects how much your subscribers actually want to hear from you.
AI-generated email does not trigger spam filters. But what it tends to produce does.
When you use AI to write email copy without doing the work of making it specific, relevant, and distinctly yours — which is how most teams are using it, because that is the path of least resistance — you end up with copy that sounds like everyone else's AI-generated email. Generic. Competent but forgettable. Written in a register that could belong to any brand in your sector. Your subscribers, who have a finely tuned sense for what is and is not worth their time, respond accordingly. Open rates may hold (for reasons we will come to shortly). Clicks decline. Replies drop. The engagement signals that inbox providers use to make placement decisions get worse.
And worse engagement signals mean worse deliverability. Not immediately. Not from one campaign. But sustained over months of lower-engagement sends, the inbox providers learn that your emails are not the ones your subscribers are interested in, and they begin treating them accordingly.
→ Litmus: The Dangers of Generative AI in Email Marketing (litmus.com/blog/the-dangers-of-generative-ai-in-email-marketing)
→ Validity 2026 Deliverability Benchmark Report (validity.com)
There is a second mechanism worth understanding alongside the engagement one. Over 51% of spam sent in 2026 is AI-generated. Phishing emails that used to take skilled human effort to craft can now be produced in seconds, in bulk, in convincing language, in any tone. Because of this, inbox providers have had to tighten their filters across the board — not just for obvious spam, but for borderline email from legitimate senders too.
You are navigating a harder inbox environment than you were two years ago, and the reason is not anything you did. It is the collateral consequence of AI lowering the floor for bad actors. Every legitimate sender is paying a small tax on bad actors' AI use, and the brands generating the lowest engagement signals are paying the highest rate.
This matters because it changes the risk calculation on AI-generated copy. It was never zero risk. It is even lower margin now.
There is a subtler version of this problem that does not show up directly in deliverability data but feeds into it over time.
When multiple brands in the same sector are all using the same AI tools, prompted in similar ways, drawing from similar training data, the copy they produce trends towards a shared register. Same sentence rhythms. Same vocabulary. Same opening gambits. Same call-to-action structures. Anyone reading email from two or three brands in the same space using AI heavily will start to notice — not necessarily consciously — that the emails feel interchangeable.
For deliverability, this matters because of word association and sender familiarity. The associations your subscribers build with your sender name, what you talk about, how you sound, what they can expect from you — are what determine whether your email gets a fair reading or a reflexive dismiss. When your copy no longer sounds distinctly like you, those associations weaken. The familiarity that earned you an engaged open last year becomes harder to maintain.
I covered this in more depth in the piece on the billboard effect and email as an awareness channel: every email you send is building or eroding an association in your subscriber's mind, whether they open it or not. AI copy that sounds like everyone else is eroding a distinctive association and replacing it with a generic one. That has commercial consequences well beyond any individual campaign.
The AI-generated copy question is important. But there is a second way AI is changing deliverability that is, in my view, more significant and less understood: what is happening inside the inbox at the provider level.
In January 2026, Google announced that Gmail is "entering the Gemini era." Powered by Gemini 3, Google's most advanced AI model, Gmail rolled out a set of features that fundamentally change how email is processed, sorted, and surfaced before a subscriber ever sees it.
This is not a spam filter update. It is a structural change to how the inbox works.
The three features that matter for deliverability
Taken together, these three changes mean that "delivered to inbox" is no longer a binary outcome. There is now a gradient of visibility within the Gmail inbox, and where your email sits on that gradient is determined by the engagement quality of your entire programme — not any individual send.
→ Folderly: How Gmail's Gemini AI Changes Email Deliverability in 2026 (folderly.com/blog/gmail-gemini-ai-email-deliverability-2026)
→ Google: Gmail is entering the Gemini era (blog.google/products-and-platforms/products/gmail/gmail-is-entering-the-gemini-era/)
→ Emailexpert: Gmail Launches AI Inbox (emailexpert.com/gmail-launches-ai-inbox/)
Gemini is not assessing your email against a spam rulebook. It is making relevance predictions based on the history between your sender domain and each individual subscriber. High engagement history — genuine opens, clicks, replies, email moves — tells Gemini that this subscriber finds what you send worth their attention. That subscriber gets your email surfaced. Low engagement history tells Gemini the opposite.
This is why AI-generated copy that produces lower engagement is a deliverability problem at the Gmail level, not just a commercial performance problem. The engagement shortfall feeds directly into Gemini's relevance model, and over time, that model decides how visible your emails are. The mechanism runs from copy quality → engagement → Gemini relevance score → inbox visibility.
The expert community is putting it starkly. Manu Cinca of Stacked Marketer: "With AI Inbox, your email will be one of many that Gemini pulls into a summary, so it's about convincing Gemini to show your email and content to the user. Because if Gemini becomes the gatekeeper, you might no longer have a direct connection with a subscriber."
Marc Thomas of Positive Human frames it differently but consistently: "I think Gmail's AI Inbox will benefit good email marketing and continue to punish bad email marketing. It surfaces good information, which is what you should have been sending the whole time, and relegates purely promotional information to more of a potluck."
He is right. And the programmes that have been investing in relevance, in list quality, in intent-based sending rather than volume-based sending — those programmes will be rewarded by Gemini. The programmes using AI to send more generic email faster will be penalised by it.
→ Harro: Gmail's AI Inbox may redefine deliverability (harro.com/2026/05/08/gmails-ai-inbox-may-redefine-deliverability/)
There is a specific measurement consequence of the Gemini rollout that needs naming directly, because it is already causing marketers to misread their programme performance.
Gmail auto-opens emails to generate AI summaries. Every time Gemini reads your email to create an overview or extract a task, that registers as an open in your ESP. Industry open rates have risen to 45.6% on average — a number that looks healthy but is significantly inflated by machine opens, not human ones. Meanwhile, click-through rates have dropped from 4.35% to 3.93%.
If you are using open rate as a primary performance metric — or worse, using opens to trigger re-engagement flows, segment your active list, or inform suppression decisions — you are now working from data that is even further from reality than it was after Apple Mail Privacy Protection inflated it from the Apple side. The two inflation sources compound each other.
This is not a new argument from me. I have been saying for years that open rate is the wrong metric to lead with. Gemini has made the case even more obvious. Click rate, reply rate, complaint rate, meaningful actions — these are the numbers that reflect what subscribers are actually doing.
→ Stop Reporting on Opens and Clicks. Here's What to Measure Instead (weareastral.co.uk/thevault/stop-reporting-on-opens-and-clicks.-heres-what-to-measure-instead)
→ Zepic: Gmail Update 2026 — What It Means for Email Marketing (zepic.com/article/gmail-update-2026-gemini-integration-what-it-means-for-email-marketing-in-2026)
Gmail gets most of the conversation because of its scale — over 3 billion users, accounting for roughly 25-32% of all email opens globally. But if you send to a B2B audience, Microsoft Outlook is probably your most important inbox environment, and the Outlook story is in some ways more concerning than the Gmail one.
Outlook has historically been the toughest major inbox provider on deliverability. In 2025 and into 2026, Outlook's inbox placement rate sits at around 77.4% — compared to Gmail at 89.8% and Yahoo at 87.3%. Spam rates at Outlook exceed 14%, the highest of any major provider. One in four marketing emails sent to an Outlook address is not reaching the inbox. That was already the case before AI got involved.
Microsoft has been adding AI-driven inbox prioritisation for years — the Focused Inbox, introduced with Acompli in 2014, was an early version of this. In April 2025 they added "Prioritise My Inbox," where Copilot classifies emails as High, Normal, or Low priority based on individual relationship and behaviour signals. In April 2026, they rolled out Copilot Agent Mode for Outlook in their Frontier program — agentic inbox triage, automatic rule-setting, draft follow-ups, and more — currently for enterprise users but with broader rollout expected.
The deliverability implication of Copilot prioritisation is the same as Gemini's but expressed differently. Copilot sorts by relationship signal and actionability. Emails from known, frequently-contacted people rise. Emails that look like tasks or time-sensitive items rise. Marketing emails from brands, even well-run ones, are structurally disadvantaged in a model built to serve the recipient's working life rather than the sender's commercial goals.
There is no fix for this in the conventional deliverability sense. Authentication, list hygiene, and complaint rate management all remain necessary — but they are not sufficient. The sufficient condition is that your subscribers actually want what you send, often enough, that the engagement history Copilot draws on puts you in a favourable position. Which brings the Outlook conversation back to the same place as the Gmail one: relevance is now a deliverability factor, not just a performance one.
→ Microsoft Copilot in Outlook — April 2026 Agent Mode rollout (testingcatalog.com/microsoft-copilot-in-outlook-adds-ai-to-manage-inbox-and-calendar/)
→ Practical 365: Copilot in Outlook brings AI to email filtering (practical365.com/prioritize-my-inbox/)
This is a point that came out of the expert conversations around the Gemini launch and it is worth including separately because it affects a specific type of email programme.
Gemini summarises email content. To summarise content, it needs to be able to read it. An email that is primarily one large image, or a collection of images with minimal text, cannot be read, summarised, or assessed for relevance by Gemini. An email that Gemini cannot process is an email it cannot promote.
Email designer and deliverability expert Jen Blair put it plainly: "Gemini can summarise the text, but it cannot summarise the design. The age of the one-big-image email is officially dead. If your email is not readable by Gemini, your deliverability may suffer if Gmail decides to take a dim view of emails it cannot scan and summarise."
This has always been a deliverability risk, image-only emails trip filters because filters cannot assess the content, and they have always relied on alt text for accessibility. But Gemini adds a second layer: even if the image-heavy email passes spam filtering, it cannot earn a favourable relevance ranking because there is nothing for the AI to rank. Proper alt text, readable text content, and a clear structure are now both accessibility practice and deliverability practice.
→ Bloomreach: Gmail Changed the Game — What Google's AI-Powered Inbox Means for Email Marketers (bloomreach.com/en/blog/gemini-gmail)
These are not hypothetical future risks. They are what is happening now, based on what I see in audits and what the data confirms.
The common thread running through every way AI is affecting deliverability is the same one that has always been true of email: the programmes that send relevant, specific, valuable content to people who actually want it will perform well. The programmes that use whatever is available — AI, volume, clever tactics — to shortcut that fundamental equation will perform worse.
What has changed is the speed and precision with which inbox providers can identify the difference. Gemini does not just sort email into spam or not-spam. It makes individual-level relevance predictions for each subscriber, based on their full engagement history with each sender, and uses those predictions to determine visibility. The more generic your email, the more it looks like everything else, the weaker those signals are, and the less visible you become over time.
AI-generated copy is not the enemy. AI used to produce more generic email faster than you could produce it manually — that is the enemy. The tool does not cause the problem. The way most teams are using it does.
The answer is not to stop using AI in email. It is to make sure AI is making your email more relevant and more distinctly yours, not less. And to build the kind of programme — clean data, engaged list, intent-based segmentation, honest measurement — that earns good inbox placement regardless of what the inbox providers decide to do next.