What Is AI Slop? How to Spot and Block AI Comments on X, Reddit & YouTube
What AI slop is, the phrases and formatting that give away AI-written comments, and how the Universal AI Slop Finder detects and flags them on X, Reddit, and YouTube.
Some comments give themselves away. They agree enthusiastically with the post above them, praise it in vague terms, and read like a customer-service email. Nothing in them responds to anything specific, and after a paragraph or two they have said nothing at all.
Comments like these are usually machine-generated, and there is now a common name for them: AI slop.
The best tool we have in Tweeks for dealing with it is the Universal AI Slop Finder, a tweek built by a user who goes by pizzaman. It is one of the most installed tweeks in the public library. This post is a walkthrough: what AI slop is, the writing patterns that identify it, how the detector's scoring works, and where it fails. Anyone selling you a perfect AI detector is lying.
What Is AI Slop?
AI slop is AI-generated content produced in bulk with little or no human review: images, articles, videos, and comments. A model writing the text is not what makes it slop. If nobody reviewed it, nobody asked for it, and it exists mainly to collect clicks or engagement, it is slop.
The word started as internet slang around 2022, when image generators made cheap AI content common. Simon Willison began promoting it as a standard term in May 2024, arguing that unwanted AI content needed a name the way unwanted email needed "spam". It caught on, and Merriam-Webster chose "slop" as its 2025 Word of the Year, which is a decent outcome for a word that started on imageboards.
Comment slop has different incentives than article slop. An article wants your click. A comment wants replies, upvotes, and an account history that looks human, because the account itself is valuable. That is why slop comments are so agreeable and so empty. Outrage farms engagement too, but agreement is easier to automate.
Why So Many Comments Are AI Now
Because it costs nothing and it pays.
A generated reply costs a fraction of a cent. On X, impressions convert into creator payouts, so bots camp under anything viral. Reddit accounts with a few years of age and decent karma sell for good money, and the fastest way to build karma is pasting model answers into advice threads. On YouTube, fresh channels use bot comments to look alive before they start pushing scams. None of this takes talent, just an API key and a hundred lines of Python for a thousand accounts' worth of activity.
The Tell-Tale Signs of an AI Comment
A caveat first: no single phrase proves anything. I sign off with "hope this helps" myself. What gives slop away is pile-up, three or four patterns landing in one short reply.
The best public list of these patterns is, funnily enough, maintained by Wikipedia's editors, who got tired of AI-written article submissions and started documenting the tells in a guide called Signs of AI writing. I think pizzaman's detector borrows heavily from it.
Giveaway phrases
Chat models are tuned to be helpful and polite, and the customer-service reflex leaks into everything they write. So you get "Great question!", "Hope this helps!", "Let me know if you have any questions", "It's worth noting that", "Feel free to explore". Nobody talks like that under a meme. The classic "As an AI language model" has mostly disappeared, but the detector still checks for it, and when it hits, case closed.
Inflated significance
AI text inflates the importance of whatever it describes. Everything "plays a pivotal role", "marks a key turning point", or is "a testament to" someone's dedication. A comment claiming a speedrun video "reflects broader shifts in the creator economy" was generated. No, AI, it's literally just a guy playing Mario fast.
The vocabulary
Some words appear far more often in model output than in casual human writing: delve, tapestry, intricate, multifaceted, garnered, seamlessly. A user using one of them means nothing, but if a two-sentence reply includes three of them, a model probably wrote it.
Essay formatting in a reply box
Numbered lists with bold headers, title-case section headings, and closing "In conclusion" paragraphs are normal in documents and strange in comments. People do not structure a throwaway reply like an English paper, but chat models do it by default for everything.
Off-key punctuation
Watch for em dashes in casual replies (models love them), curly quotes (most phone keyboards produce straight ones; models emit typographic ones), "not just X, but Y" constructions, and lists of exactly three items. All of this is ordinary in edited prose. In a dashed-off comment it suggests the text was generated.
Vague authority
"Experts say". "Studies show". Which experts? Which studies? Models reach for these fillers to sound grounded when they have no source.
How the AI Slop Detector Works
The Universal AI Slop Finder turns this checklist into an automatic score, and there is no AI involved in doing it. It is a userscript, and the full source, 989 lines, is visible on its share page before you install anything. The pattern list is worth reading on its own. Detection is pattern matching that runs locally in your browser: no server, no model, and nothing you read is sent anywhere.
As you scroll, the script scans new comments and replies, skips anything too short to judge, and tests the rest against roughly 120 phrase patterns and a dozen structural checks. Each match adds to the comment's score:
| Signal | Weight |
|---|---|
| "As an AI" / "As a language model" | 4.0 |
| Knowledge-cutoff phrasing ("as of my last update...") | 3.8 |
| "Hope this helps" | 3.0 |
| Formatted list inside a casual comment | 2.1 |
| "Plays a pivotal role" | 1.7 |
| Three or more em dashes | 1.3 |
| Each AI-vocabulary word (delve, tapestry, ...) | 0.65 |
A raw score is not enough on its own. pizzaman gated the scoring so a comment needs signals that cluster: one strong giveaway, a couple of medium signals, or several weak vocabulary hits together. A single "crucial" in an otherwise normal comment scores too low to matter. When a comment crosses the bar, it gets a badge next to the author's name reading "AI", "Likely AI", or "AI Slop", with the score on it. Hovering over the badge lists exactly which patterns fired, so there is no black box to argue with.
Three sensitivity levels (Cautious, Balanced, Aggressive) control how much evidence it wants, switchable from the Tweeks extension menu. Aggressive is the default.
What It Gets Wrong
Plenty, and you should know it up front.
False positives: some people write formal English because that is how they learned it, and some people genuinely love an em dash. Precise humans occasionally get side-eyed, which is why the script badges comments instead of deleting them.
False negatives: prompt a model with "reply like a tired shitposter, all lowercase" and it walks past every regex on the list. Pattern matching catches lazy slop. Fortunately, lazy is the whole business model. If a slop farm has to hand-tune prompts to dodge a browser extension, its costs just went up.
So treat the badge as a reason to check the account, not a verdict. A fresh profile that posts every forty seconds and replies to hundreds of viral posts a day settles the question in a way no phrase list can.
AI Slop on X
X has the most industrialized comment slop, because reply visibility converts directly into creator payouts. Bot accounts follow viral posts and reply with restatements and compliments. A reliable test: if a reply would fit under any post on your timeline without changing a word, it was probably generated.
On X, the Slop Finder scans posts and replies as they render and skips anything under 36 characters, since there is no point scanning "lmao". It also applies a slightly stricter threshold here, because terse writing is normal on X and brevity should not be penalized. Badges appear next to the display name.
X is one of the most modified sites in the Tweeks library; browse everything published for x.com.
AI Comments on Reddit
Reddit slop runs long. The classic specimen is a full ChatGPT answer pasted into an advice thread: bold headers, numbered steps, a safety disclaimer, and a closing offer to elaborate further. There is also a species the detector cannot catch: karma bots reposting old top comments word for word. That text was human the first time around, so no writing heuristic will flag it, but account history will.
The script scans Reddit comments of 48 characters and up, and it runs a bit more sensitive than on X because longer text gives the patterns more surface. Essay formatting inside a casual thread is its single loudest signal, and Reddit is where that signal appears most.
You can also see what else people run on reddit.com.
AI Slop in YouTube Comments
YouTube slop is mostly generic praise: "This video is a testament to your dedication and hard work!" posted minutes after upload by an account with no history and a default avatar. Summary bots restate the video description, and pods of small channels compliment each other on schedule.
The detector works on watch pages, follows YouTube's in-page navigation, and keeps rescanning as threads load, so badges appear as you expand replies. Inflated praise is what it catches most there.
Browse youtube.com tweeks for the rest of that shelf, including the widely installed Shorts removers.
Can You Block AI Comments Instead of Flagging Them?
Partly, and you can go further if you want.
Built in, there is a menu toggle that dims strong matches so they fade but stay readable. pizzaman built it so it never deletes or hides comments outright, because false positives happen, and hiding a person's comment by mistake is worse than making you glance past a bot's.
But this is a tweek, so that choice does not have to be yours. Open it in Tweeks and describe the change:
hide comments the slop finder marks as "AI Slop" instead of dimming them
Tweeks rewrites the userscript, and now you have a blocker. You could also hide "Likely AI" matches or auto-collapse flagged Reddit threads; the detection levels sit right in the DOM, so filters like these are one-sentence requests. If editing a script by describing the change is new to you, the Tweeks vs. Tampermonkey post covers how that workflow works.
Try It
Setup takes about a minute:
- Install the Tweeks extension for Chrome or Firefox.
- Open the Universal AI Slop Finder share page, read the source if you want, and click install.
- Open X, Reddit, or YouTube and scroll.
Sensitivity, dimming, and on/off all live in the extension menu on supported sites. If the badges feel too eager, switch to Cautious. Aggressive is the default and catches the most.
Comment sections were never entirely human. Now the machine-written parts are at least labeled.