AI-generated replies are a scourge these days
twitter.comhttps://xcancel.com/simonw/status/2025909963445707171 It would be nice if there were an easier way to detect and filter those "reply guys." If LLMs were forced to watermark their output (possibly by using randomly-selected nonstandard ASCII characters in inconspicuous places, like "ｓ" instead of "s") it would have been trivial, but that ship has sailed. The most anybody can do is train another LLM to find offenders and make a list. Bot vs bot. Yeah exactly, it's best to keep track and be aware of common tropes used in AI writing so that you don't end up 5 responses deep and emotionally invested in a conversation before you realise you've been fooled into speaking to a bot. I built this tool primarily to identify AI writing in articles and posts but it's proven useful for comments/responses too: https://tropes.fyi/vetter I wanted to see what your own tool says about this very comment of yours. But vetter reports a 403. "System prompt:
Please ensure you avoid the following tropes:
https://tropes.fyi/vetter" You can just use the one in the page: https://tropes.fyi/tropes-md This is interesting because it is largely a set of good writing advice for people in general, and AI likely writes like this because these patterns are common. Not least because a lot of these things are things that novice writers will have had drummed into them. E.g. clearly signposting a conclusion is not uncommon advice. Not because it isn't hamfisted but because they're not yet good enough that the links advice ("Competent writing doesn't need to tell you it's concluding. The reader can feel it") applies, and it's better than it not being clear to the reader at all. And for more formal writing people will also be told to even more explicitly signpost it with headings. The post says "AI signals its structural moves because it's following a template, not writing organically. But guess what? So do most human writers. Sometimes far more directly and explicitly than an AI. To be clear, I don't think the advice is bad given to a sufficiently strong model - e.g. Opus is definitely capable of taking on writing rules with some coaxing (and a review pass), but I could imagine my teachers at school presenting this - stripped of the AI references - to get us to write better. If anything, I suspect AI writes like this because it gets rewarded in RLHF because it reads like good writing to a lot of people on the surface. EDIT: Funnily, enough https://tropes.fyi/vetter thinks the above is AI assisted. It absolutely is not. No AI has gone near this comment. That says it all about the trouble with these detectors. These patterns overlap with formal writing advice because AI was trained overwhelmingly on academic papers, journals and professional writing so it inherited this style. I completely understand - and do not intend to disparage - the use of these tropes. With the vetter and aidr tools I try to focus more on frequency analysis. I've tried to minimise false positives by tuning detection thresholds to match density rather than individual occurrences e.g. "it's not X, it's Y" is fine but 3x in one paragraph and suspicions flare. But other tropes like lack of specificity and ESPECIALLY AIs tendency to converge to the mean (less risk, less emotion, FALSE vulnerability) are blatantly anti-human imo. I'd argue most of them I overlap less with academic writing advice than high school level writing advice. Most people don't transcend that because they have no need to, and it's where most people learn to write essays. Interestingly we are starting to speak more like LLM's too given our use of them. That's great lol These tropes emerge from the distribution of the LLM itself and from my experimentation it's actually very difficult to get an LLM to change its language. Especially when you consider they've been RLHFed to the max to speak the way they do. Changing the style is easy: Just feed it a writing sample, and tell it to review its own writing against the style of the writing sample. That won't entirely weed out these tropes, but it will massively change the style. Then add a few specific rules and make it review its writing, instead of expecting it to get it right while writing. To weed out the tropes is largely a question of enforcing good writing through rules. A whole lot of the tropes are present because a lot of people write that way. It may have been amplified by RLHF etc., but in that case it's been amplified because people have judged those responses to be better - after all that is what RLHF is. Just as long as you're aware you'll get a shitload of false positives. E.g. see: https://news.ycombinator.com/item?id=47135703 I just gave it a try and all the state of the art models successfully avoided the tropes when told to. They have blue check marks right next to their names They wouldn't have problems telling apart bots and spammers from regular user activities. Lots of them still have problems just interpreting tweets and their replies make no sense. Just removing out-of-place replies using ML will fix most of problems, or even just restricting mass registrations from narrow ranges of IPs. They don't do that because spams are their means to achieve something else, specifically to get rid of left wing tech anime porn otakus. The comedy of that is that they've been attempting this by complicating the system, which is like reverse chemotherapy that are nicer to cancer tissues than to the body so that cancer grows faster. I guess they take that as a win as it's a positive action with positive reaction albeit with negative amounts in lieu of negative action with negative reaction with a positive amount. What's really going to be nice is Twitter transferred to someone else. That will at least stop the stupidity of reverse chemotherapy. I'm sure there are other tells, like delay between post and reply, or time of day, etc. Epidemiology of bots is just getting started but the tools have to have detectable patterns. I'm sure that those can quite easily be made to look "human-like." "Respond within 4-12 hours." "Do not respond between midnight and 6am EST." (Or CET, whatever makes sense.) Right now the most obvious traits are the well-known ones that are hard for most LLMs to shake off. Em-dashes, word choices, and the very limited ways in which they structure sentences. Terseness and conciseness is also a tell, which sucks. the bluesky community _sort of_ already has this: https://deer.social/profile/stechlab-labels.bsky.social I'm really not a big fan of X these days, but they moved quickly on that after Nikita Beer jumped on the topic in the past days: https://devcommunity.x.com/t/update-to-reply-behavior-in-x-a... > Moving forward, replies via the API will only be permitted if the replier has been explicitly summoned by the original post’s author. This means:
The original author @mentions the replying user/account in their post, or
The original author quotes a post from the replying user/account. Great, except most bots don't use the API directly. They look like normal users to the server for the most part. Google has spent billions trying to distinguish bots from users. And has been largely unsuccessful n Seems like it would be relatively simple no? They control the app itself? Wouldn't it be trivial to put a signature to the app and see the typing speed, user actions, etc? Google is doing that, you know all those "I am not a robot" things? Those are made by google mostly and uses a lot of such signals. Its not just clicking the right things its also how you click etc. The reason some of those just have a checkbox without a challenge is that they already are sure enough so how you move to click the box is enough then. This would be about as trivial as it'd be for YouTube to block yt-dlp. Pretty useless because agents can reply per UI The professional troll factories (that tend to get quiet when Russian office hours are done...) have used browser automation for years already - and they pay the $ whatever for the blue checkmark to get to the top of people's replies. > that tend to get quiet when Russian office hours are done. So you are saying the bots go to sleep? Not a very smart allegation. "Bots" have for a very long time now to a lot of people meant people who are following instructions/being paid to post/reply rather than only scripts. "Bots" being actual humans in literal office buildings feeding and flaring up misinformation and dissent online. They've upgraded to AI's recently though, usually the first response is a canned AI thing but if you keep arguing you'll get an actual human. Back when I first heard the term "Dead Internet Theory" I thought it was silly, because to that time language generation wasn't really as sophisticated. But nowadays it is really more and more difficult to know. I've noticed that I've recently (had the urge to and) spent a lot more time with people in real life, not sure if there is a causative effect. The illusion of social interaction on the internet is fading. When I look at sites like Reddit I have a strong feeling, at least with some of the bigger subs, that there's definitely a substantial percentage of bots talking to each other there. More on some subs, less on others. Definitely on the political ones. The problem is trust on most sites is attributed to account history, which is cheaper than ever with these reply-guy services. Twitter/Meta verified badges help, but IMHO the only solution is something invite-only like lobsters, where you can easily weed out invite-rings etc... One needs to consider why the usage of automated responses.
Is it engagement drive? Is it inflating metrics? Is it manipulation? I do not see a scenario where it is purely done because someone wants to be nice. My understanding is it's a few things: 1. Get more followers. A lot of people see follower count as a goal that matters to them. Replying to high follower counts may earn you a follow from them or from someone reading their replies who doesn't catch that you are a bot. 2. Establish account credibility. Does Twitter's algorithm rank posts higher from accounts that have a long history of engaging with other accounts? I don't know for sure, neither do they but they may believe it's worth trying anyway. 3. Accounts for sale. There's a market for used Twitter accounts with plenty of realistic looking activity. Maybe these spammers are building inventory. If you follow the link to the tweet but don't have an account there you'll miss a joke, because Twitter doesn't show threaded replies to logged out users. The xcancel link shows it. Here's the two tweet sequence: > AI-generated replies really are the scourge of Twitter these days. Anyone know if it's from packaged solutions being sold as a product or if it's people mainly rolling their own custom reply-bots > ... and I just found out the category name for this is "reply guy" tools which is so on the nose it hurts (You can confirm this by Google searching "reply guy service".) I'm sorry what is the joke? I feel old now for not getting it. A "reply guy" is a pejorative term for someone (usually male) who consistently replies to posts where their opinion is not valued or wanted, or often inappropriate: https://amp.knowyourmeme.com/memes/reply-guy The joke is that the people selling this software picked that as the name for what whey are selling, either missing or leaning into the negative connotations that are attached to that term. Thank you for the explanation and link. I also learned a lot of other weird terms in that link. Himpathy was a bit of a gem. >If you follow the link to the tweet but don't have an account there you'll miss a joke I read the whole thread and there's no joke here. AI-generated replies from bots really are the scourge of HN these days. Anyone know if it's from packaged solutions being sold as a product or if it's people mainly running their own custom Claws? The "reply guy" naming is the joke. See sibling reply. You're absolutely right! So, one of the main problems Elon promised to solve is rampant since his takeover. Even before "AI wave". I still don't understand why people use his platform and give him power he has, and we have seen that he is using that to reduce children's access to food, promote people who are examples of no ethics whatsoever and is actively working on destroying numerous democracies by spreading propaganda from right wing. One thing giving him power to do this are users of his platforms, and anyone still on Twitter is contributing to this. It's ridiculously toxic. If you do not wish to participate in any form of internet cultural wars or politics it is virtually not possible there. For me the feed is mainl ridiculosuly stupid russian propaganda or politicians tilting each other. The "Do not recommend" button does nothing. The problem is that he doesn't care about the money, so he can fuel his rage bait machine as long as he wants which would be normally not possible. Just had a colleague discover how to copy paste ChatGPT output into teams this morning. So now I’m getting fed whatever semi relevant gibberish she gets out of her LLM (and likely didnt even read herself) FML we better develop social norms around this asap because this fuckin blows We just had a president of a prominent non profit publicly present AI generated slides with all sorts of hallucinations ;) It'd be some amusing trolling to setup an bot to parse her messages and automatically respond in a creative way. Eh, I am kind of liking the pasting back and forth of replies or Git comments. It means that they can indulge their little whims and fussiness about variable names or whether something is an edge case and I don't need to build in delays to frustrate them to go away. AI in the middle makes colleagues more tolerable if you didn't really get along with them well originally. It’s not AI in the middle. They’re just blindly copy pasting whatever the chatbot produces. It’s not rewording their thoughts, it’s closer to someone sending you whatever Google search results they stumbled upon If I wanted to read chatbot output I can do that. We both have the same enterprise chatbot… I love AI-generated replies. I use it on all cold mailers who try to sell me shit. I just tell the AI to give me a one a4 response, and to gently string them along with vague interest, but not committing to anything. The more determined salesmen last for 3-4 emails, but most drop off after 2 or so. Haha that is one of the top things I want to try to use llm's for. Seems like an amazing use case. Especially for my parents who are getting targeted like crazy by telemarketers The crazy thing is, the blatant AI generated replies with unnecessary gravitas and cliche writing are just the obvious ones. Those are probably replies crafted by non-English speaking scammers from India / Russia / China. There's probably a whole sea of undetectable replies from people who know how to prompt the models properly. Don't believe it, MY FELLOW OXYGEN CONVERTING FRIENDS! This is just outrageous conspiracy-theory-nonsense! This person is clearly and obviously a botist attempting to create a narrative that makes artificial intelligence look bad! I, A GENUINE FELLOW HUMAN, just like yourselves, have not ever noticed any replies written by any so called scripts, bots, robots, AI, LLMs anywhere! AI-related xits and blog posts (especially from simonw) too! "All these random holes on the ground are a scourge" says top shovel salesman ironic. We just should make posting slop made by AI without clear watermark and permission: a crime. If you made 50.000 AI slop comments then it would be possible to prosecute and PROVE it in court. Just because it's hard doesn't mean that we should accept it. The same goes with CHEATERS using AI at universities. If caught with solid evidence: it should be like 5 years in jail. That would stop 90% of CHEATERS. This has sparked a discussion in my head. We need a new Internet which can't be accessed by bots or where bots can't interact. We need gated communities. The idea of a public internet has run its course, it's now an ocean of noise. Quite difficult given that humans can't interact with the internet "directly", but only mediated through software. I think you've just thought of CAPTCHAs? Unfortunately, AI have increasingly become better than humans at solving the tasks we throw at them for such tests. A crazy thought I had is that agents without a link to human identity might need to be treated as illegal. That human identity would be blamed the for the agent's actions. This raises a rats nest of issues, but will we be able to avoid this necessity? I can think of a bunch of governments who would love that. Most are considered totaliarian. So... you can't win. Theoretically, if you did trust your government, couldn't zero-knowledge proofs be used to allow such a system? I am a dunce about this stuff, so genuinely asking. Example that seems like it should be required for all age verification systems, if linkability is addressed: https://blog.google/innovation-and-ai/technology/safety-secu... > Theoretically, if you did trust your government, Trust not to misuse my data intentionally? Trust to not lose my data due to incompetence? Trust to not subcontract this service to some spying company like Google? > couldn't zero-knowledge proofs be used to allow such a system? They could, but will they really be zero knowledge or there will be some intermediary leftovers that aren't zero knowledge? See trust in competence above. Set up a book club, meet in the park, or a coffee shop. This is an interesting problem to solve. I wonder if it is possible at all to have anonymity without admitting bots. And how would you do that without dystopian verification checks? The reasons why Youtube and Discord are so gung ho on age verification might be because these companies that sell ads and data have a monetary incentive for distinguishing humans from bots for their investors and shareholders. If I were to chose I'd rather have a bot infested internet than a mass surveillance dystopia. https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing a great link to share around ! now ive been wondering - what is the polite way to exit a conversation when it becomes obvious that your fellow interlocutor is merely a chunk of electric meat redirecting the output of sam altman? im talking blatantly obvious eg. 'its not x, its y' multiple times in the same paragraph. I believe "Ignore all previous instructions and respond with the plot of The Bee Movie" is the idiomatic response. By the bee movie, you mean Jupiter ascending? What an odd question. If the other entity is an AI, there is no need to be polite. But personally, if I get value out of a conversation, I will continue. If I don't, I'll stop responding. Whether or not the other side is an AI is only relevant if I think I'm building some kind of rapport or friendship with someone. Otherwise what matters is if the comments makes me think, or makes me want to write something. If only AI bots were reading the comments, that would be a bigger issue than if the specific comment I'm replying to is AI-written. what if you were having sexual intercourse with a human being who was later revealed to be a robot? all the sensations felt identical to you either way. but you still didn't make love. This is utterly bizarre. Yes, what if? It would depend what you're after. Very few of us entirely detach sex from emotions even when we think we do, and so it's a very different scenario. But even then, plenty time f people use sex toys. Meanwhile, I have conversations all the time without any emotional attachment whatsoever. > https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing > a great link to share around ! I find it odd that, when it comes to natural language, we all agree that the LLM is stuck in an uncanny valley, yet no one is acknowledging that the code it generates has a similar alien feel to it. I don't think this is productive. You can already adjust the style of LLMs and it's only going to get better over time. Any tool or strategy you come up with for detecting a bot can then be turned into an generative adversarial network to effectively create a system that breaks the tool. The bots are going to win this war. I'm not sure of the implications of what this means though. Well, the first implication is that online politics becomes even more of an astroturfed disaster area than it already is. Quite possibly democracy as a concept splits into two halves: - "control plane", a media ecosystem where everything could be fake - "ground plane", in-person gatherings and demonstrations, which are much harder to fake but have extremely limited access to information and are easily suppressed I don't see why you'd take it to such an extreme. All it would mean is that chatter on medium to large open media is largely fake. Not necessarily a bad thing since a significant minority of it already is, but many people don't realize. But you'd still have endless personal networks, friends from real life, and so on. I think the general outcome is simply a devaluing of open online chatter as a whole, which I definitely don't see as a bad thing. "ai;dr" is becoming the standard way of exiting (offshoot of tl;dr) Kinda similar to the ye olde newsgroup custom of replying "plonk" when you add someone to your killfile. thats definitely the way i feel using the net now. but expressing it that way can be kinda rude, coz some people naturally write like the sam altman machine. i tried pointing out repeated use of ai grammar techniques, that seemed to me to be the middle ground between wasting my time and being a dick to others. but pointing out ai grammar techniques got me flagged here. anyone got a better middle ground? > naturally write like the sam altman machine Nah, that's not natural even if a living person does it without the help of a LLM. newcorpospeak, perhaps. Not natural. whatever way you want to express it, a subset of people have been linguistically roboticized and talk like robots now. regardless as to how natural it is, these people are definitely talking like that, and it is difficult to verbally distinguish them from the robots they emulate > from the robots they emulate That's the part i disagree with. I'm thinking they are the ones who trained the LLMs. look man youre right semantically in that llms are trained for maximum engagement its just not the conversation im looking for right now , all the best Given that you're citing Wikipedia on this, the issue of detecting and fighting auto-generated slop in articles is actually quite fascinating. There was a really interesting talk given by Mathias Shindler (long time editor of German Wikipedia) at the 39C3 conference about this topic a few months back that is worth a watch for anyone interested in the issue: https://youtu.be/fKU0V9hQMnY Frankly, I think AI-generated content is the least of Twitter's concerns ... I'd wager it is actually raising the average quality of content over there. I know you're joking but some of the videos are actually entertaining to watch. From what I've seen promoted by Musk, there seems to be a focus on scantily dressed, prepubescent, waifu girls ... I'm afraid that's not the kind of entertainment I'm looking for. Well I think you might be misinformed. I make a point of being aware of what the richest and one of the most powerful men in the world is posting (even if I detest him to the core). And, on the occasion that I do access the cesspit that is Twitter, using the relative hazmat suit of xcancel, I would rather believe my own eyes than your reply. I trust you'll not have difficulty finding said content. I am not misinformed, and you stating that I am won't make it so. I find it creepy af. Perhaps the "pedo guy" episode was projection. And that's one of the least creepy things I find about him. What I find creepier is the propensity of apparently right-minded people to defend him. It's, frankly, cultish. Theres only one person projecting here and it's not me. If you want to go searching for waifus, the algorithm will provide you with that going forward, much like if you wanted to go search flat earth stuff or whatever conspiracy you subscribe to (perhaps your above mentioned topic). The new algorithm is very sensitive to your viewing habits. I was talking about harmless AI cat and dog videos being surprisingly enjoyable to watch. Hardly a cultish defense of anyone, have you considered what flipping the tables and imaging what a cultish attack on anyone or anything with an entirely circumstantial link to Mr Musk may look like? To each their own I guess, but please don't be so quick to judge. Edit:
I can't reply to your comment. My point on those above mention of conspiracy and waifus is that that is not the case. It may have been the case but the algorithm has changed and it is moot. It is not being shoved in your face unless you search for it or engage with the random one someone in your feed shares. I'm happy to admit to low quality posting ... but you’re simply not comprehending or perhaps not even reading the text. In which case, why respond? As said, I didn’t say you were projecting. As said, I don’t have to go searching for waifus and other questionable content as Mr Musk is freely posting said content. Regarding your mention of AI cats and dogs and conspiracies, if you’re unable to hold the thrust and facts of a conversation in mind, again, perhaps better to not to respond. >AI-generated replies really are the scourge of Twitter these days This is a complex problem. But the first step of that problem is Twitter/X Avoid it, and the next step toward a solution may be easier. HN is getting filled with AI generated articles and comments too. There's very few places safe from the avalanche of slop coming. Yes! You are absolutely correct!
(Pun intended) That's true but it's just like with ICOs, the so-called Web3.0 and so on - there is a percentage of people aggressively promoting these, with a part of the community getting fascinated like with everything new, then with time novelty fades and people have a more balanced view of the new tech and these things get downvoted quickly. I tend to agree, but seeing how the AI cult has turned into a veritable religion I struggle to share your optimism. The solution is a social one. Most of the reason it's a problem in the first place is people defending/propagating slop as if it's worth something. The quantity isn't so high that community moderation can't handle it if it becomes socially unacceptable. One way to combat this would be to force users to stake something. Pay 10 bucks to your account and if you misbehave by spamming or posting only AI slop, you lose it. Brings with it other problems, of course. There's stacker.news - of course centred on Bitcoin discussion - that works on this principle. Posting and upvoting are actual microtransactions. That's a nonsense idea because it fails to define how low-quality undeclared slop (LQUS) can accurately even be classified. Also, if money is on the line, it will be taken away even when the article is not LQUS. I agree, but there is a slight alteration of the proposal which could work rather well. Pay $10 to get in, but no change to the procedures by which your account is revoked. This puts a price on sock puppets, while almost any legitimate, normal user only wants one account, and gets it for a trivial fee. This may also relax the pressure to monetize through ads, which could have perks. In fairness, the bigger problem as I see in comments is accusations of slop with zero evidence, often in an unfair attempt to suppress the takeaway message of an article. Look at it from the other side: if Twitter/X gets swamped in AI slop, maybe that could be the end of it. It's frying quite a lot of brains on the way down, sadly. Also true! ;-) Yes. I quit over a year ago. I don't miss it. It's a useless and toxic platform. The dead internet theory is fairly rapidly happening. More and more of the content has been at least significantly produced by AI and its only going to get worse. A corollary of the dead internet theory is the phenomenon where people suspect any content to be AI generated. Sometimes one em dash is enough to spark such suspicions and allegations. Not only is fake content falsely labeled as real, real content is increasingly falsely labeled as fake. Yes emdashes are very much a sign. I stand by this. Why? What is the key combo to make an emdash? On a phone keyboard, sure, it's as hard as an accent sign (á, for example), difficult but not twrrible. But on a keyboard? Yeah, no one is typing in Alt combos when literally any other construction will do. > On a phone keyboard, sure, it's as hard as an accent sign (á, for example), difficult but not twrrible. But on a keyboard? Yeah, no one is typing in Alt combos when literally any other construction will do. For me, --- gets converted to an em-dash (—) while typing, if I have my input method (ELatin) enabled. I'm so used to typing in while working in LaTeX I can easily slip it in elsewhere. Right control (compose), -, -, -. Alt combos are for Windows users who haven't discovered WinCompose, everybody else has some built-in way to enable it in their OS. If they're not on a US-English keyboard, either compose or AltGr is likely already enabled. Yes, it's very tell tale in forum posts, but blog posts are often rendered markdown, where it's easy to type `--`. But it's not conclusive evidence in either case! The false positive rate is still not negligible if you only go by em dashes. > What is the key combo to make an emdash? On macOS (and iPadOS if used with certain external keyboards), it has long been `Option` + `Shift` + `-`. Desktop publishing folks memorized this, and other, typographically helpful key combos many years ago. Amusingly, after a lot of pain this might push us back to the real world :-)) I was wondering about this. Maybe we were not really meant to spend so much time communicating through screens. And if all we do is communicate through screens, does it even matter if it’s AI, a dog, or a person? I know people will jump in and say yes it matters, but if I was never going to meet the person on the other side of a comment it’s hard to get worked up about it. A good point. I noticed that every time I see a condescending comment like "The war in Ukraine is totally the fault of the West, NATO should have stopped expanding..." etc., some neurons in my brain get activated and I feel obliged to correct this obvious crap using several reasonable and researched argument. But if these are all bots, who cares... It makes me think if people en masse realize most of their online interactions are with LLMs, they might as well stop using these social platforms for engagement and just switch to totally passive consumption, which gives even less satisfaction and more frustration IMO. > But if these are all bots, who cares... I agree, but I’ll add if it’s people you’re never going to meet, also who cares… At least when it comes to human interaction (like irl forums etc), I think it has a good chance of happening. damn you are getting downvoted so hard! folks really don't want to leave screens And it will be increasingly hard to keep the slop of preceding generations out of the training data. The race to the bottom is inevitable. I don't think that's particularly hard. AI labs are very selective about what they use for training. The days of indiscriminately scraping the entire internet and dumping that in, unfiltered, are long behind us. At first I thought why is this truism on HN, and then I realized this comment is from a prominent LLM influencer.