How to tell if a reply on X is from a bot (or just written with AI)

By Jose Enrique, maker of BotBlind · Updated October 19, 2026 · 8 min read

First, two different questions

"Is this a bot?" usually hides two questions that have different answers.

• Is the account automated? A script watches for posts and fires replies with no person deciding what to say. This is what X itself goes after: in July 2026 its head of product, Nikita Bier, said X had removed 42,000 accounts that were automating replies with chatbots (Search Engine Journal).
• Was the text written with AI? A real person pastes a ChatGPT reply to look clever or to farm reach. The account isn't a bot, but the reply has the same empty feel.

For you as a reader, the difference matters less than you'd think. Both push real conversation down the thread, and you can mute, hide or skip both. But it matters a lot before you call someone a bot in public, which is a good reason to be careful (more on that below).

Why AI detectors don't work on replies

It's tempting to paste a reply into an AI detector. Don't trust the result. Detectors look for statistical patterns across many sentences, and a 20-word reply doesn't have enough of them. Even the people who build these tools say so: Botfilter's extension skips any text under 25 words, and GPTTrace warns when a text is under 150 words (Botfilter, GPTTrace). OpenAI withdrew its own AI text classifier in 2023, citing low accuracy.

There's a second problem. "Great point, couldn't agree more!" reads the same whether a model or a person typed it. People wrote generic replies long before chatbots existed.

So the useful evidence is mostly outside the sentence.

Signals in the thread (strongest)

1. Timing that doesn't add up

A fast reply is normal. A long, polished, perfectly punctuated paragraph that lands eight seconds after the post is not, because nobody reads, thinks and types that fast. Open the post and the reply and compare their full timestamps. X shows the exact time on a post's own page, while the thread view only shows a rounded "8s" or "3m".

Watch out: fans with notifications on reply within seconds all the time. Speed only counts when it comes with length and polish.

2. Copies and near-copies

Scroll the replies. If two or three accounts posted almost the same sentence, or the same unusual three-word phrase, that's coordination or a shared template. This is one of the few signals that's hard to fake innocently.

3. It could be a reply to anything

Read the reply without the post. If it would fit under a post about a product launch, a football result and a cat photo equally well ("This is so true!", "Thanks for sharing this insight"), it probably wasn't written for this post. Real people mention a detail, disagree, ask something specific, or make a joke that only works here.

Watch out: a short question ("source?") doesn't mention the post either, and it's very human.

Signals on the account

4. X's own "Automated" label

X offers an automated-account label for bots that follow its rules. When it's there, the profile says the account is automated and names the account that runs it. Honest bots wear the label. Spam bots don't, so its absence proves nothing, but its presence settles the question.

5. The handle

A first name plus a long string of digits (think @laura_m84520193) is what you get when an account is created in bulk and X suggests a free username. Security researchers call these low-budget sock puppets (NBC News). Plenty of real people never changed the username X gave them, so treat it as a hint.

6. The replies tab

This is the single best check, and almost nobody does it. Open the profile and switch to its Replies tab (the list of everything the account has replied to). A person replies to friends, argues, makes typos, talks about their own life. An automated or farming account replies to dozens of large accounts in a row, at all hours, with replies that all have the same shape: compliment, restatement, emoji.

7. The profile around it

A recent join date, a stock or AI-generated photo, a bio that's mostly a link or "DM for promo", and a following count far above followers. Each is weak alone. Together with the replies tab, they're strong.

8. The checkmark doesn't mean human

A paid checkmark means someone pays for the account. It doesn't mean a person wrote each reply. Bier shared X data in March 2026 showing that users paid through revenue sharing were twice as likely to be flagged as spam as users who weren't paid (tech-ish). The payout is the incentive.

Signals in the writing (weakest)

These are the ones everyone shares on X, and they're the least reliable. Use them only to tip a decision that the signals above already started.

• Stock phrases: "great point", "couldn't agree more", "this resonates", "a testament to", "game-changer", "delve".
• Templates: "It isn't X, it's Y." A one-line validation followed by "Curious what you think about…". "Here's the thing:".
• Style tells: em dashes in a two-line reply, a neat closing summary, a trail of 🚀🔥💯.
• No stake: no opinion, no experience, no "I tried this and…".

Plenty of careful writers use dashes and tidy sentences. Plenty of AI replies have been edited to remove them. On their own, these prove nothing.

A 60-second check

1. Read it without the post. Could it go under any post? +1.
2. Check the time. Long and polished within a minute? +1.
3. Scan the thread. Same text or phrase from other accounts? +2.
4. Look at the handle and label. "Automated" label settles it. Name plus long digits? +1.
5. Open the Replies tab. Dozens of same-shaped replies to big accounts? +2.
6. Only now, the writing. Stock phrases or templates? +1.

This isn't a scientific scale, just a way to stop yourself from deciding on one clue. If you're at 4 or more, it's very likely automated or farmed. At 1 or 2, give the person the benefit of the doubt.

Three made-up examples

Invented replies, invented handles. Any resemblance to real accounts is accidental.

What not to do

• Don't accuse people publicly on one clue. Non-native speakers, shy people and polite fans write generic replies. Being called a bot is a bad first interaction.
• Don't rely on "ignore all previous instructions". The trick went viral in 2024. Sometimes a bot falls for it. Often nothing happens, and real people now reply with a poem as a joke, so the result tells you very little.
• Don't paste replies into AI detectors and take the percentage as an answer, for the reasons above.

When you're fairly sure

• Mute or block the account from the ··· menu, and use Report post → Spam for clear spam.
• On your own posts, Hide reply and consider limiting who can reply.
• Add the phrases bots keep reusing to your muted words. Our starter list of muted words for bot replies sorts them by how likely they are to catch real people.
• For every native option, step by step, see how to hide AI and bot replies on X.

How BotBlind does this check for you

BotBlind runs a version of the same check on every X conversation you open: timing, copies and shared phrases in the thread, digit-heavy handles, replies that ignore the post, spam hooks, and (weighted low) stock phrases and style tells. Each reply gets a score from 0 to 100, and replies over your threshold fold into one line that shows the score and the reasons. It can't open each account's replies tab the way you can, so it leans on what's visible in the thread.

It's rules, not an AI detector, and it runs entirely in your browser. It will sometimes fold a real person, which is why nothing is deleted and Show is always one click away.

Status: BotBlind's Chrome Web Store listing is pending.

FAQ

Can an AI detector tell me if a tweet was written by AI?

Not reliably. Replies on X are usually too short for statistical detectors, and humans and AI write short replies in very similar ways. Look at the thread and the account instead: timing, repeated text and posting history tell you more than one sentence does.

Is a verified account always a real person?

No. A paid checkmark means someone pays for the account, not that a person writes each reply. X's head of product has said that accounts paid through revenue sharing are more likely to be flagged as spam than accounts that are not.

Are fast replies always bots?

No. Fans with notifications turned on reply within seconds all the time. Speed only means something when the reply is long and polished, or when the account replies that fast to many big accounts in a row.