How to Spot AI Comment Bots: The Tells That Hold Up, and the Ones That Just Insult Good Writers

A post you follow has forty comments. Every one of them agrees with it. Every one opens by restating the post's own thesis back at it, adds one adjacent observation, and ends on something faintly inspirational. Not one mentions anything specific that was actually in the post.

You know. You cannot prove it, you cannot report it as anything, and if you say so out loud in the replies somebody will accuse you of being paranoid about a person's writing style. That gap — between knowing and being able to check — is the whole subject, and it is worth being precise about which half of it is solvable.

The tells that hold up

The reliable signals are behavioural. They are about what the account does over time, not about how any single sentence is phrased. That is the crucial split, and almost every viral "how to spot AI" thread gets it backwards.

• Nothing in the comment proves the poster saw the post. This is the strongest single tell and the easiest to apply. Take the comment and ask: could this have been written from the headline alone? From the first paragraph? From the thumbnail? A generated comment is usually working from a summary, so it will engage with the topic and never with a detail — no reference to the third example, the chart, the aside, the thing you said at 4:12.
• Output volume that is not physically available. Open the profile. If there are thoughtful, paragraph-length comments across a dozen unrelated posts in the last hour, that is not a fast reader, it is a pipeline.
• No conversational memory. Reply to it. A person picks up the thread. A one-shot generator produces another self-contained comment that does not reference what you just said, or restates its first comment in new words.
• Structural sameness across unrelated topics. This is the one that really convicts. Read five of the account's comments on five different subjects. If they share a skeleton — validate, elaborate, gentle uplift — across cooking, geopolitics and a product launch, that skeleton is a template. Humans have habits; they do not have fixed structures that survive a change of subject.
• No stake and no specificity. Real commenters leak. They mention their job, their city, a thing that happened to them, an opinion that costs them something. Generated comments are relentlessly agreeable and biographically empty, because there is no biography.

Style tells you how something was written. Behaviour tells you whether anybody was there. Only the second one is evidence.

One published analysis worth knowing about approaches this statistically rather than by intuition: dejan.ai's write-up on comment bots proposes combining sentiment distribution, vocabulary diversity and structural consistency with behavioural assessment — impossible output volume, absence of conversational memory, no personal experience, suspicious consistency across disparate topics. Its most useful methodological point is the one I would repeat: combine several indicators, never act on one. I am deliberately not reproducing its percentages here, because I have not verified its sample and method, and a number you cannot check is worse than no number.

The tells that mostly accuse innocent people

This section exists because the popular detection folklore has a serious false-positive problem, and the people it hits are not bots.

• Em dashes. The most-shared tell of the last two years, and it convicts editors, academics, and anyone who has read a book. Punctuation is a habit, not a fingerprint.
• Clean grammar and full sentences. This accuses careful writers and, disproportionately, people writing in a second language who learned formal register first and have no instinct for casual English.
• Enthusiasm. "Great post, really useful" is what a genuinely appreciative person with nothing further to add writes. It is also what a bot writes. On its own it separates nothing.
• Words like "delve", "robust", "leverage". These correlate with a training distribution, not with a bot, and they are ordinary vocabulary in several professions and several dialects.
• A new account with few followers. That is every new user. It is also, admittedly, every new bot — which is exactly why it has no discriminating power.

Notice what all five have in common: they can be evaluated from a single comment, in one second, without leaving the page. That convenience is precisely why they spread, and precisely why they are unreliable. Every tell in the previous section requires you to click something. The reliable signals cost effort by construction, because they are about behaviour over time and behaviour over time cannot be read off one sentence.

The two-minute check, and the part you genuinely cannot verify

Here is the procedure, and it is short enough to actually run.

1. Reply with a question only a reader could answer. Not "are you a bot" — reference a specific detail. "Which of the three did you mean?" A person answers or says they skimmed. A generator produces another paragraph about the topic.
2. Open the profile and read ten comments across different posts. You are looking for one thing: the same skeleton on unrelated subjects. This takes ninety seconds and settles most cases.
3. Check the clock. Timestamps across those ten comments. Volume that no person has time for is the one tell that needs no interpretation.

Now the part nobody will tell you, because it is unsatisfying: you cannot verify any of this, and the platforms have not given you a way to. There is no label. There is no public endpoint that tells you when an account was created, how many comments it posted today, or whether it authenticated through an automation tool. The report menus mostly offer "spam" or "harassment" and have no box for "this is generated". Even the outcome is invisible — when a comment disappears, you are not told whether a filter caught it, a moderator removed it, or the poster deleted it. We ran into the same opacity from the author's side when we mapped out the three different systems that can remove a Reddit post, only one of which tells you.

So the honest ceiling on this skill is: you can reach a confident private judgement in about two minutes, and you can almost never reach a demonstrable one. Which argues for a specific posture. Use the judgement to decide where to spend your own attention — whose replies to take seriously, whose comment section is worth joining — and not to prosecute anyone in public. The cost of being wrong about a real person is high, and the benefit of being right about a bot is that you scroll past.

If you are reading this from the other side — you want to automate replies and you would rather not become the thing described above — the tells in the first section are a specification in reverse. Respond to something specific. Do not produce volume a person could not. Remember the conversation. Keep a human reading the drafts before they go out. We wrote that version up separately in the honest operator's guide to AI comment automation, and the contrast with pure comment generation is in the three kinds of TikTok comment bot and in what Discord follower servers actually run on.

FAQ

Is there any tool that tells me whether a comment is AI?

Nothing you should act on. General-purpose AI text detectors perform poorly on short text, and comments are short text — there is simply not enough signal in twenty words for a classifier to be reliable, and their false-positive rate falls hardest on non-native speakers. More importantly, the question they answer is the wrong one. A human can paste a generated sentence and a bot can post a human's sentence; what you care about is whether an account is operating at machine scale without reading anything, and that is a behavioural question no text classifier is looking at.

Why do bots comment on small accounts at all?

Because for most of these operations the comment is not the point — the profile is. A comment is a cheap way to put a clickable account in front of an audience, and small posts are easier to appear near the top of. Some of it is warming a new account with activity before it is used for something else. Either way, the value is in the impression, not in the conversation, which is exactly why the text never engages with anything specific: engaging specifically would cost more and buy nothing.

Should I report or block them?

Block, generally, and report only when there is a category that fits — most report menus offer spam or harassment and have no option for "generated", so a report often does nothing. Blocking is immediate, costs you nothing, and does not require you to be right in public. The thing worth avoiding is the accusation in the replies: the stylistic tells that make people confident enough to accuse are the unreliable ones, and the person on the other end is often just someone writing carefully in their second language.