How to tell if something was written by AI, without a detector

The easiest way to know if something was written by AI is to read it and look for specifics. If all it has is generic advice and generic examples, no specifics, every sentence the same length, hedging and a tidy wrap-up at the end, that's a good hint. If it also has words like "genuinely" or "roughly," does the "rather than X" correction or ends with a summary, any one of those isn't a guarantee. But if several are there and the text lacks specifics, it's a strong hint.

How can you tell if something was written by AI in 60 seconds?

Run all seven, in this order. The first matters most, and it takes about a minute.

1. Contains specific info? (15 seconds) Is there one bit of info only the author could know? A name? A number? A date, or something that went wrong? Humans writing about real work include specifics. AI, unless it's in the prompt, writes the general case.
2. First and last line (10 seconds). Does the first line stay very close to the prompt ("I wanted to follow up on our conversation")? Is the last line overly upbeat, without content ("Let me know if you have any questions!" or "Excited for what's ahead!"), or a summary of the entire text? AI often does both.
3. 2026-isms (10 seconds). Does it contain words that, in our data, AI uses far more than humans? When AI polishes text: "genuinely" (12.5 times as often), "roughly" (7.7 times), "plenty of" (7.2 times), "essentially" (7.1 times) or "for instance". When AI writes from scratch: "honestly" (8.3 times), "whether you're" (7.1 times), "roughly" (4.7 times) and "here's" (3.9 times). One of these can be normal. Four in a short email isn't.
4. Read three sentences out loud (10 seconds). Do all the sentences have the same length and cadence? Humans vary the length of their sentences. AI often doesn't.
5. Are the examples real? (5 seconds) Often, the AI will include placeholder examples ("For example, a growing team might use this to save time"). Real examples tend to have a who and a when.
6. Is there hedging? (5 seconds) AI often hedges. Look out for lots of it ("this may help," "based on available information"). Also look for things framed as a correction of something else ("should be seen as X rather than Y"). Wikipedia's editors flag these corrective framings as a new pattern in AI text since mid-2026.
7. Is the formatting right for the medium? (5 seconds) If it's an email or a LinkedIn comment, does it have lots of formatting (bold, bullet lists, headings)? People often paste straight from chatbots. In our study of AI drafts, models added bold without being asked, even in emails (5.8 bold phrases per 1,000 words).

Zero or one hit: probably nothing. More: likely written or polished by AI, especially if there's no specific info. It's not proof, but it's a good indicator. If you want the entire list, with data, check out all the signs of AI writing we track.

What should you look for in emails, LinkedIn posts, cover letters and reports?

What to watch out for is pretty much the same across formats, it's just that how much certain tells mean will vary by context. For example, an em dash in an email is unusual (we counted 0.06 em dashes per thousand words in people's emails prior to ChatGPT) but normal in a news story (1.62). Here are some specifics by format.

Emails

Look out for an opening of "I hope this email finds you well," a recap of a meeting you both went to, next steps that aren't specific, an ending of "Looking forward to hearing from you!" and, basically, anything that could be sent to anyone. Most people are specific in their emails to the one person they're sending them to, and generally only talk about one thing.

LinkedIn posts

Look out for an opening one-liner to hook the reader, paragraphs that are only one sentence long, lists of "here's what I learned," an ending question to prompt engagement, and lines like "It's not about X. It's about Y." Tons of people write on LinkedIn like this on purpose, so it's not a surefire way to know if someone is using AI. Still, the "it's not X, it's Y" pattern showed up in the AI-drafted versions of a bunch of LinkedIn posts we studied (0.51 per thousand words) but not in the original, human-written versions people had posted. Oddly, AI didn't write anywhere near as many exclamation marks in those LinkedIn posts as people did (3.4 vs. 8.3 per thousand words).

Cover letters

A big giveaway of an AI-written cover letter is if it only includes generic things like praise for "your commitment to innovation," skills picked up from the job ad and generic statements of enthusiasm. It probably won't include specifics about a project the applicant worked on, what the result was, and why they want this job specifically. That being said, lots of people use AI to help polish cover letters they wrote, and that's totally reasonable. If you're hiring, ask specific questions about what the candidate included in their cover letter instead of trying to detect whether they used AI. (Or, if you're a recruiter looking for help writing outreach to candidates that doesn't read as AI, check out our page for recruiters.)

Reports

With longer pieces of writing like reports, it's usually the structure that gives it away. Look out for really short sections all with their own headings, bold text at the beginning of every bullet point, groupings of three, and conclusions that rehash what was said in the introduction. (Pangram found that AI-generated text used bullet lists 9 times as often as humans, and groupings of three 4 times as often. In our AI drafts, it was 8.9 bullet lines per 1,000 words against 2.4 for people, and 1.4 times as many groups of three.) Another thing to look out for in longer pieces is a claim like "studies have shown" without citing any studies, or a statistic without saying where it came from.

What does an AI-written email look like?

Below is an example of the sort of short follow-up email AI might produce every day of the week. Next to it is the same email after it's been run through Zavus. Below that is a list of the tells that were picked up on.

AI draft

Hi Sarah,

I hope this email finds you well! I wanted to follow up on our conversation from last week regarding the Q3 marketing plan. I genuinely appreciated your insights — they were truly valuable. Rather than waiting for our next meeting, I thought it would be helpful to share a few thoughts now. For instance, we could streamline our approval process, align on key metrics, and strengthen collaboration across teams. It would take roughly a week to set up. Let me know if you have any questions or would like to discuss further. Looking forward to hearing from you!

Best,
Alex

After Zavus

Hi Sarah,

I hope all is well with you. I'm writing to follow up on our discussion last week about the Q3 marketing plan. I really appreciated all the information you gave me. I wanted to share a couple ideas I had before we meet again. Some of the things we can do to make this work include: streamlining our approval process, aligning on key metrics, and strengthening collaboration across teams. It'll take about a week to get this set up. Please let me know if you have any questions. I look forward to hearing from you!

Best,
Alex

Tells picked up on in this example:

1. Opener. "I hope this email finds you well!" followed by a recap of a conversation Sarah was a part of.
2. Intensifiers. "Genuinely" and "truly" are both in the same sentence, and both are among the words AI overuses most when it rewrites in our data.
3. Praise with no content. Sarah's insights might be "valuable," but we don't find out which one.
4. Unprompted correction. The idea of waiting until the next meeting has not been mentioned, but the writer feels it necessary to argue against it: "Rather than waiting for our next meeting."
5. Generic triad. Streamline, align, strengthen: a triad of abstract verbs that could apply to any team, in any company, and it's introduced by "for instance," something the bloggers in our rewriting study almost never used.
6. Vague number. "Roughly a week" to set up what?
7. Closer. Two stock phrases in a row, "Let me know if you have any questions" followed by "Looking forward to hearing from you!"

Note that the em dash has not been included, as it's such a weak signal in this example: you probably wouldn't notice it if all the others had been removed. Also note that while Zavus can remove the intensifiers, the em dash and the unprompted correction, it can't add the detail only Alex would know, for example which of Sarah's ideas was useful. That part is still up to the writer.

Are em dashes a sign of AI writing?

Em dashes aren't a tell of AI writing overall, but they can be a hint in some types of writing. People do use em dash characters (0.73 per thousand words in a corpus of 22 million words of pre-ChatGPT human writing), and the four bloggers in our rewriting study used them even more frequently (2.56 per thousand words). But when we had AI write from scratch, it used them a lot (4.31 per thousand words, about six times as often as in the humans' writing). Still, people use dashes nearly as often as AI does, they just type them as "--", as a hyphen with a space on each side, or as an en dash. In a sample of 4,000 human-written emails, we found 3.8 of all kinds of dashes per thousand words but only 0.06 em dash characters. Our AI-drafted emails had 2.9.

On the other hand, some AI models do tend to use an unusually high number of em dashes in their writing. GLM-5.3 used 10.75 em dashes per 1,000 words when rewriting and 11.13 when writing from scratch. Claude barely used them: 0.10 for Claude Opus 5.5 when rewriting, 0.09 for Claude Sonnet 5.5 when drafting. Moreover, OpenAI has let users disable em dashes in their custom instructions since November 2025. So, while the presence of em dashes might be a signal for some AI models, their absence doesn't mean anything. Wikipedia now lists em dash overuse as a "historical" indicator of AI writing. You can find more details in our article on why ChatGPT uses so many em dashes.

Can you ever be sure something was written by AI?

No. Style alone can't prove who wrote something, and there's no point getting cocky. People and software can make mistakes.

First, don't think you're cleverer than you are. In Medianet's 2026 survey of 803 Australian journalists, 48% thought they could "almost always" tell if a pitch was AI. But the Wikipedia guide to spotting AI cites a 2025 study where people did no better than random chance, and it says: "Do not rely too much on your own judgment." It also notes that people are starting to write like AI, which makes it harder.

Second, the software isn't perfect either. According to Wikipedia, tools such as GPTZero and Pangram are better than chance but have "non-trivial error rates," and can be fooled by paraphrasing and by AI models they weren't trained on. False positives happen: people have been accused of using AI for writing they did themselves.

Third, some popular tells don't work. Wikipedia says perfect grammar, formal or "fancy" prose, bland prose and transition words on their own are ineffective. Lots of careful human writers do all four.

What should you do if you think AI wrote it?

If the writing is clear and gives you what you need, it doesn't much matter how it was written.

If you really want to know how it was written, ask specific questions like "Which part of the Q3 plan are you referring to here?" People who were there will be able to answer. Avoid making accusations about how the writing was created, because if you are wrong the cost falls on a person, and your evidence is weak.

If you are the one sending AI-assisted writing, follow the checklist above in reverse: add the details only you know, cut the stock openers and closers, and vary your sentences. For more detail, here is our guide to making AI writing sound human. Zavus does the rewriting inside the agent you already use, so the emails and posts you send sound like you wrote them.

Frequently asked questions

How can you tell if an email was written by AI?

Look for something only the sender would know. AI-written emails tend to open with "I hope this email finds you well," offer to do something without saying what, list three generic benefits and close with "Let me know if you have any questions." One of these doesn't mean it's AI. All of them together probably does.

Are em dashes a sign of AI writing?

Not on their own. Our AI drafts used the em dash character about 6 times as often as people, but people type dashes in other ways ("--" or " - "), and some models, like Claude, barely use them. In an email, where people almost never type the character, it's a slightly stronger hint.

Are AI detectors accurate?

They do better than chance, but they make mistakes, including flagging human writing as AI. Wikipedia's editors recommend not relying on them alone. If you use one, treat it as one input next to your own reading.

Can people really spot AI writing?

Not as well as they think. Wikipedia cites a 2025 study that found people were no better than chance, though a 2025 preprint found that a group of expert annotators caught 90% of AI text. It's something you can get better at with practice.

Sources

1. AI writing tells in 2026: data from 7,554 rewrites and 1,177 drafts, Zavus, November 2026
2. Wikipedia: Signs of AI writing, Wikipedia, accessed November 2026
3. 9 Signs of AI Writing, Backed by Data, Pangram, 2026
4. Journalists embrace AI despite rising concerns and threat to jobs, report reveals, Medianet, March 2026
5. OpenAI says it's fixed ChatGPT's em dash problem, TechCrunch, November 2025