Jorge LewisDisclosure: GTM Bud is an AI outreach product, so this article scrutinizes the category we sell in. We wrote it honestly anyway, including the parts where AI-written email genuinely fails, because the only version of this piece worth your time is the one that does not flatter us.
Can prospects tell if AI wrote your cold email? Usually yes, but not the way most senders fear. Recipients are poor at detecting AI prose and very good at detecting the absence of research, and almost everything people call an “AI tell” is really a shallow-research tell. A generic compliment reads as mass production whether a model or a tired SDR wrote it, and a line that could only have been written about this one prospect reads as human even when a model drafted it.
That distinction is not theory for us. Our parent agency, Referral Program Pros, has run more than 4,000 outbound campaigns and booked over 7,000 meetings, and GTM Bud is built on that playbook with a written guarantee behind it: 5 percent positive replies on LinkedIn or 1.5 percent on email, or a full refund. A guarantee priced in replies does not survive output that prospects smell as automated, so the question in this article’s title is one we answer with money, not opinions.
This piece covers what the evidence actually says about detection, the specific tells recipients notice, what the circulating reply-rate numbers really are (anecdotes, quoted as anecdotes), and the question buyers should be asking instead.
Can prospects actually tell when AI wrote a cold email?
Recipients cannot reliably detect AI prose, but they detect shallow research instantly. The most relevant controlled evidence is a 2023 study by Jakesch, Hancock, and Naaman published in PNAS, which ran six experiments with 4,600 participants and found people identified AI-written self-presentations at roughly chance accuracy, about 50 to 52 percent. Worse, the cues people trusted as human, such as first-person pronouns and contractions, were unreliable and easy for AI to exploit. So the sentence-level question, did a machine type this, is one prospects mostly cannot answer. The question they answer in seconds is different: did anything in this email require knowing me? A compliment that fits any company, a fact lifted from a profile headline, warm enthusiasm from a total stranger. Those read as mass production whoever produced them, and mass production is what gets archived.
One honest caveat on that study: it tested self-presentations in professional, hospitality, and dating contexts, not cold email specifically. No controlled study we could find tests AI detection on cold email itself. What exists for cold email is practitioner experience, and the practitioners largely agree with the lab: nobody is parsing your sentence rhythm, everybody is checking whether you did the work.
The tells recipients actually notice
Consultant Lilach Bullock, who has written repeatedly about why AI cold emails kill reply rates, names a consistent set of tells, and notice that every one of them is a research failure wearing an AI costume. Here is her list, against what a researched email does instead:
| The tell | Why it reads as automated | What a researched email does instead |
|---|---|---|
| Generic compliments (“your impressive work at X”) | The sentence fits any company in the segment | Names one verifiable event: a hire, a launch, a post, an open role |
| Fake specificity | Restates the prospect’s own profile headline back at them | Cites a fact that took actual digging and connects it to a reason for writing |
| Over-enthusiastic stranger tone | “Thrilled” and “truly inspiring” from someone you have never met | A neutral, colleague-level register that treats the reader as a peer |
| Stock openers | Restating the subject line, or “I hope this email finds you well” | Makes the research itself the first line, so the opener does the proving |
Run any of these through the detection question and the pattern holds. “I love what you are building at Acme” is not suspicious because an LLM has a recognizable way of loving things. It is suspicious because it carries zero information that required knowing the recipient, and a buyer who gets 30 of these a week has pattern-matched the emptiness, not the authorship. We walk through how to build the opposite kind of opener in our guide to writing cold emails that get replies.
What do the reply-rate anecdotes really show?
Numbers circulate in this debate, and they deserve honest labels: they are anecdotes from practitioners, not controlled studies, and most come from people selling either services or software.
Bullock reports a small split test of 120 sends in which the human-written batch generated more than triple the replies of the AI batch, and a client engagement in which reply rates went from 2.7 percent to 12.7 percent after switching from AI-generated to hand-written emails. She also claims prospects now recognize AI-written outreach in roughly two seconds. All three figures are self-reported by one consultant whose business is teaching people to write, the samples are small, and she acknowledges herself that her tests are not scientific. Quote them, as we just did, as exactly that.
The counter-position comes from vendors. Hunter, which sells an AI writing assistant alongside its email tooling, published a skeptical piece titled “AI Personalization Isn’t Ready For Primetime” arguing that reasoning-based AI personalization was not yet reliable, while its guide to sending cold emails with AI argues the real problem is generic output rather than AI itself. Hunter’s own survey research reports that 67 percent of decision makers do not mind whether a cold email was written with AI, while 47 percent of B2B professionals experimenting with generative AI said they would be less likely to reply to an email they believed was AI-written. Both figures come from a company with a product in this fight, and so does our read of them. Industry press is warier still: MarTech ran its analysis under the headline AI-powered cold email: A nightmare in the making?, which tells you how the scaled, lazy version of this plays out in reputational terms.
Put the anecdotes and the vendor positions side by side and something useful emerges. The people who hand-write report big lifts over bad AI output. The people who sell AI say the fix is better inputs. Both are describing the same variable from opposite ends: specificity. Nobody on either side claims that a researched, accurate, specific email loses replies because a model drafted the sentences.
Where both sides land: AI researches, specificity decides
The practical consensus across practitioners and vendors is a hybrid workflow: AI does the research and the drafting, and the specificity of the output decides the reply, with a human owning strategy and reviewing high-value sends. The division of labor follows from what each side is bad at. A human cannot research 50 prospects a day deeply enough to write a genuinely specific line for each one, and a model given no research can only produce the generic patterns that get archived. A model given real inputs, a funding round, a hiring spike, a post the prospect actually wrote, produces copy that passes the only detection test recipients run. Bullock’s own recommended fix, after all her split tests, is to let AI handle research and structure while a human shapes the final message. That is the same architecture the AI-tool side describes, disagreeing only about how much of the final pass software can safely own.
What that workflow looks like layer by layer is its own article, and we have written it: our breakdown of how AI cold email writers actually work covers the research, writing, and quality-control stages, and our guide to personalization at scale covers the signal-based playbook that feeds them. This article’s job is narrower: the detection fear, examined, points you at research depth, and research depth is a workflow choice.
It is also where the risk concentrates. The one failure mode where AI-written email is categorically worse than a lazy human template is hallucination: a model that invents a podcast appearance or cites the wrong company’s funding round produces an email that is worse than generic, because it proves the research was fake. Any AI system you let near prospects needs source-traced claims and quality checks before send. This is the standard we built GTM Bud’s AI cold email writer against, and the reason the reply-rate guarantee can exist: every personalized claim traces to a real data source, because one invented detail costs more trust than a hundred plain emails.
The better question: did the email require knowing the prospect?
If you are evaluating your own campaigns, or a tool, or an agency, “can prospects tell it is AI” is the wrong test, because the answer hinges on a variable nobody can measure from the outside. Replace it with a test you can run on any draft in ten seconds: does this email contain at least one sentence that could not have been written without knowing this specific prospect? Swap in the name of a competitor of theirs. If the email still works, it was never personalized, whatever wrote it.
That test explains every result in this article at once. It is why the lab subjects in the PNAS study could not spot AI self-presentations: the AI text contained plausible specifics. It is why Bullock’s hand-written batches won: her hand-writing process forced research the AI batch skipped. It is why vendors insist failed personalization is the problem: their product survives the test only when the research layer is real. And it is the test we apply before anything sends under our guarantee, because across the 4,000+ campaigns behind GTM Bud, the first line proving real research has moved replies more than any other single variable our agency has touched.
Frequently asked questions about AI-written cold emails
Do AI detection tools catch AI-written cold emails?
Not reliably, and it rarely matters. Detection tools are built for long-form text and produce noisy results on a four-sentence email, and no prospect runs your message through a detector before deciding whether to reply. The judgment that matters happens in the first line: either the email proves research or it does not. Optimizing to beat a detector solves a problem your prospect does not have.
Should you tell prospects that AI wrote your cold email?
No one expects a disclosure line, and adding one solves nothing. What you are accountable for is accuracy: every claim in the email must be true and verifiable regardless of who or what drafted it. A hallucinated funding round or a misattributed blog post destroys trust in a way no disclosure can repair. Hold AI-assisted email to the same standard as anything you would sign personally.
Do AI-written cold emails hurt deliverability?
Indirectly, yes. Mailbox providers score engagement, so a batch of near-identical, low-relevance emails that nobody answers drags sender reputation down over time, whether a model or an intern wrote them. The fix is the same as the reply-rate fix: fewer, more relevant sends. Pairing researched copy with a cold email automation tool that manages warmed infrastructure and pacing protects the domain while the relevance earns the engagement signals.
Is it better to write cold emails yourself at low volume?
At a handful of sends per week to named accounts, yes. Manual research plus manual writing is the quality ceiling, and at that volume you can afford it. In our agency campaign experience, the economics flip somewhere around 20 to 30 prospects a day, where manual research stops being feasible and the real choice becomes AI research with human-grade specificity versus templates. Our comparison of AI tools for personalized cold email maps which tools actually do the research half. The volume decides the workflow, not ideology.
Can ChatGPT alone write a cold email that gets replies?
Only if you hand it the research first. A general-purpose model knows nothing about your prospect until you paste in the specifics, so a bare prompt produces exactly the generic output recipients archive. Used as a drafting layer on top of real research you supply, it can produce a solid first draft. The research step is the part that decides the reply, and it is the part a bare prompt skips.
Stop worrying about detection and start proving research
So, can prospects tell if AI wrote your cold email? They can tell when nobody did the research, and that is the whole detection story. The controlled evidence says humans cannot spot AI prose at better than a coin flip. The practitioner anecdotes, honestly labeled, say lazy AI output loses badly to researched writing. The vendors, us included, say the fix is a real research layer. All three are pointing at the same sentence: the one in your email that could only have been written about this one prospect. Write that sentence, or build a system that writes it, and the authorship question dissolves.
If you would rather not build that system yourself, done-for-you outbound runs the researched version of this for you, on the same playbook behind 7,000+ booked meetings, with the 5 percent LinkedIn and 1.5 percent email positive-reply guarantee standing behind every send. The guarantee only works because the emails pass the test this article is about.