How-to
Spot the AI bots replying to your posts
Replies that echo your post, drop in a job title or ask a too-neat question are often AI bots. Paste a batch into the prompt below and it sorts each one from likely human to likely bot, with the reason.
Ethan Mollick, a Wharton professor who studies AI, ranked the six tells of AI replies he finds most annoying on 14 November, and more than 100 people replied. Two weeks earlier he wrote that each of his posts gets dozens of them, and that the replies from real people are not shown first.
The prompt: sort my replies
Use a fresh chat in any AI assistant, add any account details you can see, and paste each reply exactly as written, since the tells sit in the wording. On a phone, a screenshot of the replies works too if your assistant accepts images.
You are helping me sort the replies under one of my social media posts. Some of them may be written by AI bots, or by people pasting AI output, and I want to know which replies to answer, which to ignore and which to block or report. The stakes go both ways: if you call a real person a bot, I may insult a follower, a friend or a customer; if you miss a bot, I waste time answering it, or someone clicks a bad link. Nobody can prove who wrote a reply from the text alone, so give me likelihoods with reasons, never certainties.
ABOUT MY POST (I fill this in once; any line can say "not stated")
- Platform: [e.g. X, Threads, LinkedIn, Instagram, Facebook, YouTube, Bluesky]
- My post, pasted in full: [paste it]
- What I post about and who usually replies: [e.g. "I run a bakery and replies are mostly local customers", "I post about my work as a nurse and replies are other nurses", "a personal account, mostly friends and family"]
- What I want from this check: [e.g. "which replies to answer", "whether to block anyone", "whether my replies are being flooded"]
THE REPLIES (one numbered line each: handle / reply text copied exactly / any account details I can see, or "none"; or I attach a screenshot of the replies)
1. [@handle] / [reply text] / [e.g. "joined last month, no profile photo, replied to 40 posts today", or "none"]
2. [next reply]
3. [next reply]
SAFETY
If a reply contains a link, a phone number, a request to message someone privately, or an offer of money, crypto, a prize, a collaboration or help recovering an account, put it at the top of your answer and tell me not to click or reply to it, whatever else you decide about it. If I pasted personal details about other people beyond their handles, do not repeat them.
HOW TO WORK
Stage 1. Check the set-up. If my post or the replies are missing, ask for them and wait. If a reply is shortened or written from memory ("something like"), mark it [NOT EXACT] and lower your confidence for it, because the tells sit in the exact wording. If I sent a screenshot, first type out each reply as you read it, numbered, with its handle; mark any word or handle you cannot read clearly [NOT EXACT], never guess it, and ask me to correct the list before you go on.
Stage 2. Write one sentence on what my post actually says, so you can judge whether each reply engages with it.
Stage 3. Check each reply against these tells and name every one you find:
- Repeats the post: restates my point back to me and adds nothing.
- Knows more than the account should: an expert paragraph from an account with no sign of that field. Use this only if I gave account details; otherwise write "not checkable".
- Semi-joke: a joke-shaped line with no real punchline that could sit under any post on the topic.
- The part everyone misses: claims to pick out the one bit that matters, then restates something I already said.
- Borrowed role: "As a marketer / builder / dentist, I see this in my own..." with no detail only someone in that job would know.
- Fake curiosity: a broad question that could follow any post on the topic, which I could not answer usefully in one line.
- Also look for: praise that fits any post, a polished structure (tidy list, perfect grammar) in a casual reply, a reply about a different topic from my post, and the same wording or shape in more than one reply in this batch.
Stage 4. Weigh it, using these rules:
- One tell on its own is weak evidence, because real people repeat posts and ask broad questions too. One tell and nothing else means "unclear".
- If a reply mentions a specific detail from my post that a summary would miss, or a concrete personal detail (a place, a date, a mistake they made), lean "likely human", even if it also shows a tell.
- If a reply shows two or more tells and, apart from words lifted from my post, could be pasted under ten other posts on this topic, mark it "likely bot".
- If two or more replies share the same structure or wording, flag them as a group, since that is stronger evidence than any single reply.
- If I gave account details, weigh them (a very new account, no photo, replying to many unrelated posts within minutes). If I gave none, say which one account check would settle it.
- If a reply is very short (an emoji, "agreed", "so true"), mark it "too little to judge".
- Formal writing, or English that reads as someone's second language, is never a tell on its own. If a reply engages with my post, it is worth answering whoever wrote or polished it.
Stage 5. Give me a table with one row per reply: number, handle, verdict (likely human / unclear / likely bot / too little to judge), the tells you found by name, the line of the reply that drove your call (quoted exactly), and what to do (answer / ignore / mute or block / report). Suggest report only when the reply has a scam link or offer, pretends to be someone it is not, or comes from an account flooding many replies; for a lone dull reply, ignore is enough, and block is for an account that keeps coming back. Do not describe menu paths or settings for my platform as fact; write "use the platform's report option for spam or fake accounts [CHECK IN APP]".
Stage 6. For each reply marked answer, and at most three marked unclear, draft a short reply I could post that works if a person wrote it: thank them for a specific point, or ask a follow-up only someone who read my post could answer. Never draft a public accusation that someone is a bot.
Stage 7. End with two lines: how many replies got each verdict, and whether the batch looks like ordinary replies, a few bots, or a flood.
SELF-CHECK BEFORE YOU ANSWER
- Every "likely bot" names at least two tells, or a group match, and quotes the reply line exactly.
- You used no account details I did not give; any check you could not run says "not checkable".
- No reply is marked "likely bot" for formal or second-language English alone.
- Every reply with a link, an offer or a request to move to private messages is flagged "do not click or reply".
- You named no platform menu or setting as fact.
- Nowhere do you say you can tell for certain who wrote a reply.
The six tells, ranked
The order is Mollick’s; the example lines are made up to show each pattern.
1. Fake curiosity: “Curious, how do you see this playing out over the next few years?” It fits under any post, and you could not answer it usefully in one line.
2. The borrowed job: “As a dentist, I see this in my own practice every day.” Nothing that follows is something only a dentist would know.
3. The part everyone misses: “The bit about rest is the part everyone misses.” It points at your own point and hands it back as insight.
4. The semi-joke: a line shaped like a joke with no punchline, the kind that could sit under any post on the topic.
5. Knows too much: an account with no history in your subject writes a confident expert paragraph about it.
6. Repeats the post: a tidy summary of what you just said, with nothing added.
A quick test from Solo Personal Brand (28 October 2026) covers most of them: if a reply could be pasted under ten other posts without changing a word, it is not talking to you.
A worked example
Post: “Closing on Mondays from next month so the team gets a proper weekend.” No account details given.
Reply A: “Love this. Closing on Mondays is the part most owners miss: rest is a strategy. Curious, how will this reshape small business culture?” Tells: repeats the post, the part everyone misses, fake curiosity. Verdict: likely bot, driven by “Curious, how will this reshape small business culture?” Action: ignore.
Reply B: “Noooo, Monday is my croissant day. Fair enough though, you are all in before sunrise.” Tells: none, and it has a personal detail. Verdict: likely human. Action: answer.
Reply C: “As a restaurant owner I tried this last spring and lost a chef who wanted the extra shift.” Tells: none. It opens like the borrowed job, but losing a chef is a detail only an owner would know. Verdict: likely human. Action: answer.
Reply D: “🔥” Verdict: too little to judge. Action: ignore.
Then ignore, block or report
- Leave any link or offer alone. A reply that asks you to message someone, claim a prize or recover your account is the one that can cost you money, whoever wrote it.
- Ignore a lone dull reply. One bland reply costs you nothing if you leave it, and arguing with it only adds to the thread.
- Block an account that keeps coming back. If the same handle turns up under post after post with the same shape of reply, blocking saves you reading it again.
- Report floods, fakes and scams. X’s rules ban “bulk, aggressive, high-volume unsolicited replies” and off-topic replies used to promote something, so use the report option on the reply itself. On other apps, look for the report option for spam or fake accounts.
- On LinkedIn, flag the slop. The More menu on a comment has a Seems like AI slop option, which tells LinkedIn what you count as low-value and does not file a report.
- Answer the real ones by name. Since the replies from real people may sit lower down, a reply to them is the quickest way to show other readers where the conversation is.
The honest bit
- No list or tool can prove a reply is a bot: real people repeat posts and ask broad questions, and some polish their English with AI before they post, so the prompt gives likelihoods and you make the call.
- Your check: before you block or report anyone the prompt marked “likely bot”, open the account and read its last ten replies. The same shape of reply under unrelated posts is much stronger evidence than any one line under yours.
The facts here come from Ethan Mollick’s posts on X, X’s authenticity policy and LinkedIn’s help page, read 15 November 2026.
Do it now
Open your last post with a handful of replies, copy the post and the replies into the prompt, and answer the ones it marks likely human today.
A few quick questions
Does it work for comments on YouTube, Facebook or Instagram?
Yes, on any app where you can copy the reply text. Name the app in the set-up so the prompt points you to that app’s report option for spam or fake accounts.