How to Tell If an Email Was Written by AI

TL;DR

• AI-written emails have gotten considerably harder to spot than they were even a year or two ago, the obvious tells are mostly gone.
• A few genuine signs still hold up: a certain structural predictability, an oddly even tone regardless of the subject matter, and specific phrasing patterns that show up disproportionately in AI-generated text.
• None of these signs are individually conclusive, but together they're a reasonable way to form an educated guess.

Spotting an AI-written email used to be fairly easy, oddly formal phrasing, an overuse of certain transition words, a generic quality that felt slightly off. Current AI writing tools have largely closed those gaps. Here's what still holds up as a genuine signal, and what doesn't anymore.

Why this got harder

Early AI writing had a handful of consistent, recognizable habits, certain phrases repeated often, a specific overly-formal register, transitions that felt slightly mechanical. Writing tools have specifically improved on exactly those patterns, since they were widely known and easy to train against. The result is that text generated by current tools is considerably harder to distinguish from human writing using the old checklist.

What's still a reasonable signal

A structural predictability that's hard to articulate but noticeable

AI-generated writing, even good AI-generated writing, tends toward a certain structural evenness, an introduction, a few clearly delineated points, a tidy conclusion, applied fairly consistently regardless of the actual content or context. Human writing, especially quick, informal emails, tends to be messier and less symmetrically structured. This is a soft signal, not a hard rule, but it's one of the more durable ones.

An oddly consistent tone regardless of subject matter

A person's writing tone naturally shifts based on what they're actually writing about, more clipped and efficient for a routine logistics email, warmer or more careful for something sensitive. AI-generated text sometimes carries a more uniform tone across genuinely different types of content, less responsive to the emotional register a specific situation would naturally call for.

Specific phrasing patterns that show up disproportionately

Certain phrases and transitional constructions appear disproportionately often in AI-generated text compared to typical human writing, not because they're wrong, but because they're statistically common outputs for a model to generate. Recognizing a handful of these overused constructions is one of the more reliable, if imperfect, signals still available. A few of the most commonly cited examples: 

• "I hope this email finds you well" as an opener, or similarly generic pleasantries used regardless of context
• "Delve into", "boasts", or "a testament to" appearing in otherwise plain, everyday writing
• "It's important to note that..." or "It's worth mentioning that..." used as a transition, especially more than once in the same message
• "Not only... but also..." constructions used more frequently than a person would naturally reach for them
• Tidy three-part lists ("clear, concise, and actionable") where a person would more likely just say one or two things
• "In today's fast-paced world" or similarly broad, scene-setting openers before getting to the actual point

‍

An absence of small, specific, idiosyncratic details

Real personal or professional writing often includes small, specific details that wouldn't be easy to generate generically, a specific reference to a prior conversation, an offhand detail that only makes sense with real context. Text that reads as competent but strangely generic, missing those small anchoring specifics, is worth a second look.

What doesn't work as a reliable signal anymore

Checking for overly formal language, an absence of contractions, or a generically polished quality no longer reliably distinguishes AI-generated text from human writing, since current tools can match a much wider range of tones convincingly, including deliberately casual or informal registers when prompted to.

Why none of this is conclusive on its own

Every individual signal here is a probabilistic clue, not proof, which means the honest answer to "was this written by AI" is often "probably, based on a combination of factors," not a confident yes or no from a single detail.

Treating any one of these signs as definitive proof is a mistake, since human writers can produce text with these exact same qualities, and well-prompted AI writing can avoid all of them. The most reasonable approach is weighing several signals together rather than looking for one conclusive tell, and being appropriately cautious about how much confidence that combined judgment actually deserves.

FAQs

Can you still reliably tell if an email was written by AI? 

Less reliably than a year or two ago. The most obvious old tells, overly formal language, repetitive phrasing, have largely been addressed by current writing tools. A few subtler signals, like structural predictability and consistent tone, still offer some signal, but nothing is conclusive on its own.

What's the most reliable sign that an email might be AI-written? 

No single sign is fully reliable, but a structural evenness, an introduction, clearly delineated points, a tidy conclusion, applied consistently regardless of context, tends to be one of the more durable patterns to watch for.

Does a lack of contractions or overly formal language mean an email was AI-written?

Not reliably anymore. Current AI writing tools can convincingly match a much wider range of tones, including casual and informal registers, which means this specific old signal no longer distinguishes AI-generated text as reliably as it once did.

Is there a foolproof way to know for certain if something was written by AI? 

Not from reading the text alone. The available signals are probabilistic clues that combine into an educated guess, not individually conclusive proof, since both human writers and well-prompted AI tools can produce text that defies any single expected pattern.