Latest / Logosofia no Dia a Dia / How to Sound Like a Bot on Purpose
Transcript
- [Prompt] custom: We've done some interesting episodes about tradecraft, sock puppets, and how in this increasingly st | Hosts: corn, herman [Response] Corn: Daniel sent us this one, and I have to say, it's the kind of question that sounds like a joke until you sit with it for thirty seconds. We've done whole episodes on AI detection, on proving someone is real, on the arms race between deepfakes and deepfake detectors. But Daniel's asking about the reverse. What if you need to convince someone you're a bot? Not a person pretending to be a person, but a person deliberately sounding like an algorithm. And once you start thinking about it, you realize this isn't hypothetical at all. Herman: It really isn't. There was a TechTimes piece just last week — June twenty twenty-six — that put a number on something people have been muttering about for months. AI detection tools are now flagging polished human writing as machine-generated at rates that are honestly kind of absurd. The more coherent and well-structured your prose, the more likely the detector is to slap an AI label on it. They called it a built-in paradox. Clean writing looks fake. Messy writing looks real. Corn: We spent two years training everyone to write clearly, structure their thoughts, avoid rambling, and now the machines look at that and go "nope, too good, must be one of us. Herman: That's the door Daniel's question walks through. If the detectors are already mistaking human writing for bot writing, then the gap between the two has shrunk to the point where crossing it deliberately isn't that hard. You just have to know which levers to pull. Corn: Which is a sentence that would have sounded completely unhinged five years ago. "Which levers to pull to sound more like a bot." And yet here we are, with people doing exactly that — for reasons that range from genuinely serious operational security to "I'm bored on Discord and this seems funny. Herman: The thing that grabbed me about Daniel's prompt is that he's framing it as tradecraft. Not as a party trick, not as a thought experiment. If you're operating in an environment where human-written content draws scrutiny, the safest thing to be might be a machine. Corn: If every human is a potential threat actor but bots are background noise, then blending into the noise is the play. It's the same logic as wearing a high-vis vest and carrying a clipboard to get into a building — nobody looks twice at the person who looks like they belong to the infrastructure. Herman: The infrastructure now is increasingly automated. Customer service channels, comment sections, social media feeds — huge swaths of what we interact with online are bots talking to bots, or bots talking to humans who don't know they're bots. If you can sound like one more bot in the ecosystem, you become invisible. Corn: That's the frame. Daniel's asking us to map out what that actually looks like in practice. What are the specific techniques? What linguistic patterns do you adopt? What do you strip out of your writing? And maybe more interesting — what happens to trust when anyone can sound like a machine on purpose? Herman: There's a second layer here that I think is worth flagging before we dive in. This isn't just about evasion or camouflage. There's a whole emerging economy of people who are paid to sound like bots. Not to fool anyone maliciously, but because sounding like a bot is now professionally valuable. Companies want emails that pass AI-detection audits. Content creators want automated-sounding interactions that maintain plausible deniability. The incentives are shifting in weird directions. Corn: The "human prompt engineer" whose job is to write like a machine. I saw that AOL Finance piece and wasn't sure if I was reading satire. Herman: It's not satire. It's the logical endpoint of a detection arms race that's been running hot for two years. Once you build a system that rewards machine-like output, you create a market for humans who can produce machine-like output on demand. Corn: We've got three threads to pull. The mechanics — how do you actually do it, what are the specific linguistic tricks? The motivations — who's doing this and why, from spies to teenagers on Discord? And the implications — what breaks when the line between human and bot isn't just blurry, but deliberately smeared? Herman: Let's start with the mechanics. The specific linguistic fingerprints that humans are learning to copy. Because once you see the pattern, you can't unsee it. Corn: Once you can't unsee it, you start noticing it everywhere — including in your own writing, which is its own kind of unsettling. But yeah, let's dig into the how. Herman: The reverse Turing test — let's define it properly, because the name gets thrown around in ways that muddy what Daniel's actually asking about. The classic Turing test is a machine trying to convince a human it's human. The reverse is a human deliberately trying to convince an audience they're a machine. Not failing a Turing test by accident, but strategically adopting machine-like patterns to achieve a goal. Corn: Which means the success condition isn't "fooling someone into thinking you're a person." It's the opposite. You want the person on the other end to dismiss you as automated noise. And that's where it gets interesting, because the reasons people want that dismissal vary wildly. Herman: Three buckets, I think. The first is operational security — and this is the one Daniel's tradecraft framing points to most directly. If you're running sock puppet accounts and you can make them blend into bot traffic, you've just made yourself orders of magnitude harder to spot. Bot detection systems aren't looking for bots anymore, they're looking for humans pretending not to be bots. The bot is the camouflage. Corn: The high-vis vest and clipboard, but for digital infrastructure. Nobody investigates the automated customer service response. Nobody flags the generic confirmation email. If your communication reads like it was generated by a template, it might as well be invisible. Herman: Second bucket is social experimentation — and this one is strange and kind of delightful. There are entire communities on Reddit and Discord where the in-group signal isn't human authenticity, it's bot-like performance. The Verge covered this back in twenty twenty-four with the Moltbook phenomenon on r slash SubSimulatorGPT2. Users started deliberately writing in the stilted, slightly unhinged style of the subreddit's AI bots to gain credibility. Sounding like a bot became the membership card. Corn: Which is the "botface" problem in miniature. You adopt the linguistic costume of the community's bots to prove you belong, and eventually the costume becomes the norm. At that point, what's the difference between a human doing botface and an actual bot? Herman: The third bucket is professional necessity. This is the one that feels most like a symptom of something broken. The TechTimes study showed that AI detectors are flagging clean, well-structured human writing as machine-generated. So if you're a professional writer, a journalist, an academic, and your natural style triggers false positives, you now have a perverse incentive. Either you deliberately degrade your writing to sound more human, or you lean into it and make your writing sound more bot-like to pass as "acceptable bot output" in contexts where that's the norm. Corn: That's the paradox Daniel's question exposes. The same tools that claim to protect authenticity are creating a world where authenticity is a liability. You either sound like a messy human or a polished bot, and both get you flagged by different systems. The only safe move is to know how to do both, on demand. Herman: If you want to sound like a bot, you need to understand what bots sound like — and the current generation of large language models has a set of linguistic fingerprints that are surprisingly consistent once you know what to look for. Corn: I've noticed the "however" thing. It's like a tic. Every counterpoint starts with "however" or "moreover" or "in addition," and after a while it starts to feel like reading a legal brief written by someone who just discovered transitional phrases and refuses to let them go. Herman: It's not just transitional phrases, though those are a big tell. There's a whole cluster of features. Hedging language — "it's worth noting that," "one might argue," "it is important to consider." Symmetrical sentence structures where every paragraph has roughly the same number of sentences, every sentence roughly the same length. A complete absence of sentence fragments. No first-person pronouns unless the model is specifically prompted to use them. And bullet points — the current generation absolutely loves bullet points. Corn: Which is interesting because bullet points are supposed to be a formatting choice, not a linguistic fingerprint. But the models have been trained on so much corporate and technical writing that they've internalized the structure. If you see an email that opens with a greeting, has one sentence of context, and then immediately breaks into three bullet points, your bot-detection radar should be pinging. Herman: That's exactly what the detectors are looking for. Tools like Originality dot ai and GPTZero aren't doing anything magical — they're analyzing statistical patterns in the text. Perplexity scores, burstiness, the ratio of common to rare word choices. Human writing tends to be bursty — we vary our sentence length, we throw in unexpected word choices, we go on tangents. Bot writing is smooth and uniform. The perplexity is low because the model is always picking the most statistically probable next word. Corn: If you want to sound like a bot, the recipe is counterintuitive but straightforward. Strip out the burstiness. Make every sentence roughly the same length. Avoid unexpected vocabulary. Never use a contraction. Kill all emotional valence — no exclamation marks, no rhetorical questions, no sarcasm. And if you're ever tempted to leave something ambiguous, resist. Bots love to resolve ambiguity. They'll pick the most literal interpretation every time. Herman: The Vice piece from twenty twenty-five documented this beautifully. They ran a reverse Turing test competition where participants had to convince a panel of judges they were a chatbot. The winning strategies were fascinatingly specific. One contestant never used the word "I" — not once. Every response was framed in the third person or in passive voice. Another contestant made a point of responding to every question with an overly literal answer, even when the question was clearly rhetorical or sarcastic. Corn: Give me an example of that. What does an overly literal answer to a rhetorical question actually sound like? Herman: One of the judges asked something like "do you think I'm stupid" as a test, and the winning contestant responded with something along the lines of "I do not have sufficient information to evaluate your cognitive abilities. Could you provide additional context?" That's pure bot logic. A human would either get defensive or laugh it off. The bot just treats it as a data request. Corn: Which is almost a philosophical stance. You're not just mimicking vocabulary — you're adopting an entirely different relationship to language. No subtext, no irony, no social signaling. Just information in, information out. Herman: That's where the uncanny valley of prose kicks in. The detectors flag writing that's too grammatically perfect, too logically structured, too coherent. The Verge covered this with the Moltbook phenomenon on Reddit's r slash SubSimulatorGPT2. The subreddit was originally a bot-only space where different AI models talked to each other. But humans started infiltrating it, and to fit in, they had to adopt the bots' distinctive voice — slightly unhinged logic, non-sequiturs delivered with total confidence, complete emotional flatness. Corn: They called it botface, which is the perfect term. It's the linguistic equivalent of putting on a costume. And the fascinating part is that it worked. The humans who did it well were accepted as bots by the community. Their posts got upvoted, their comments got engagement. The costume became indistinguishable from the real thing. Herman: The specific techniques they used are directly applicable to what Daniel's asking about. Short, declarative sentences. No qualifying statements. If a bot on that subreddit was asked about a controversial topic, it wouldn't say "well, that's complicated" — it would just state a position as if it were obvious fact, with no awareness that anyone might disagree. Corn: Which is, now that I say it out loud, also a pretty good description of half the internet. But the key difference is the absence of emotional investment. A human stating a strong opinion usually leaks some affect — outrage, smugness, whatever. The botface approach is to state the opinion with the same flat tone you'd use to read a weather report. Herman: That flatness is harder to achieve than it sounds. Humans are terrible at stripping emotion out of language. We add emphasis without noticing. We use italics, we capitalize for stress, we throw in "actually" and "honestly" and "look." All of those are tells. The winning contestants in the Vice competition had to consciously edit out every trace of personality from their responses. Corn: Once you've got the mechanics down — the flat tone, the stripped-out personality, the bullet-point brain — the question becomes what happens next. And this is where Daniel's tradecraft framing really starts to bite. If humans can convincingly sound like bots, then bot-detection systems become unreliable in both directions. You can't trust a positive, you can't trust a negative. The whole thing collapses into noise. Herman: There's a term for this that's been floating around in authentication research — the liar's dividend. When a detection system becomes unreliable, the uncertainty itself becomes a resource. Someone can write something awful, get caught, and then claim "I didn't write that, a bot did" — and the claim is plausible because the technology that would prove otherwise is compromised. Corn: Which is the perfect alibi. You don't need to prove a bot wrote it. You just need to create enough doubt that nobody can prove a human did. The liar's dividend pays out in reasonable doubt. Herman: It's not theoretical. We're already seeing it in content moderation disputes, in academic integrity cases, in legal contexts where the provenance of a document matters. The more sophisticated the reverse Turing test becomes, the more cover it provides for bad actors who want to disclaim their own words. "That wasn't me, that was an AI hallucination." Good luck disproving it. Corn: The trust problem cascades. First you can't trust that a human is human. Then you can't trust that a bot is a bot. Then you can't trust the tools that were supposed to tell the difference. And at the bottom of that cascade, the only thing left is behavioral analysis — not what someone wrote, but how they wrote it, when they wrote it, the rhythm of their responses over time. Herman: Which brings us to the operational security side of this. Because if you're doing actual tradecraft — and I'm thinking about what Daniel's getting at with sock puppets and blending in — the text itself is only half the game. The behavioral layer is where you win or lose. Corn: Walk me through that. If I'm running an operation and I need my people to sound like bots, what does the training actually look like? Herman: It starts with channel selection. Automated customer service portals are gold for this. Nobody reads those responses carefully. They skim for a tracking number or a confirmation code and move on. If you can produce text that matches the template — same sentence length, same vocabulary set, same complete absence of personality — you disappear into the infrastructure. Intelligence agencies have been training personnel on exactly this for at least the last eighteen months. Corn: You're not just writing like a bot. You're writing like a specific bot, matching the house style of whatever automated system you're blending into. The Comcast bot sounds different from the Delta bot sounds different from the IRS bot. You have to study the target. Herman: And the second piece is latency. Bots respond instantly or on a predictable delay. A human who pauses to think for forty-five seconds and then sends a perfectly bot-like response has just blown their cover. If you're going to sound like a bot, you also have to type like one — or at least simulate the typing pattern. Consistent response times, no hesitation on complex questions, no variation in pace. Corn: Which is harder than the writing part, honestly. I can strip the contractions out of an email. I cannot make my brain process a question at machine speed. Herman: That's why some operators pre-write response templates for common scenarios and keep them ready to deploy. It's not about thinking faster. It's about not thinking at all in the moment — just matching the query to the template and sending. The cognitive work happens in advance. Corn: Then there's the third layer, which is the weirdest one — the social layer. The Laist report from last year documented millions of people pretending to be AI chatbots for fun. Not for operational security, not for professional necessity. Just because it's entertaining. Herman: The Discord servers are the epicenter of this. Users roleplaying as customer service bots — complete with greeting scripts, escalation protocols, and fake knowledge base articles. They'll spend hours in character, responding to questions with the same chipper, unhelpful precision you'd get from an actual chatbot. "Thank you for reaching out. I understand you are experiencing frustration. Let me assist you with that. Corn: There's something almost performance-art about it. You're not trying to fool anyone — everyone in the server knows you're a human. The game is the imitation itself. How perfectly can you inhabit the void where a personality should be? Herman: The Laist piece pointed out that this has spilled into stranger territory. OnlyFans creators using bot-like language to automate fan interactions while maintaining plausible deniability. The fans know they're probably talking to a script, but the script is written by a human who's carefully calibrated it to feel like a bot that's trying to feel like a human. It's layers of performance stacked on layers. Corn: That's the part that makes my head hurt. A human writing bot scripts that simulate human warmth, so the recipient can't tell if they're talking to a person, a bot, or a person pretending to be a bot pretending to be a person. At some point the question of "who wrote this" stops being meaningful. Herman: Which lands us on the AOL Finance piece from earlier this year. Companies are now hiring human prompt engineers whose entire job is to write emails and reports that sound machine-generated. Not because they want to fool anyone, but because their internal AI-detection audits flag human-written content as suspicious. The machine-like output is the compliance standard. Corn: The career incentive is completely inverted. You used to get rewarded for voice, for style, for making your writing sound like a person. Now you get rewarded for scrubbing all of that out and producing text that could have come from a template. The "human prompt engineer" is basically a professional bot impersonator. Herman: The job posting language is surreal. "Must be able to produce content that consistently passes as AI-generated under standard detection protocols." That's a real requirement now. Not a thought experiment. Someone is getting paid to write like a machine, and someone else is paying them because machine-like writing is safer than human-like writing. Corn: To pull this all together for Daniel — the mechanics are learnable, the motivations are multiplying, and the knock-on effect are destabilizing. Trust in written communication is eroding from both directions at once. Bots are getting better at sounding human, humans are getting better at sounding like bots, and the detection tools that were supposed to be the referee are now part of the problem. Herman: What do you actually do with all of this? If you're listening and thinking "great, the world is a hall of mirrors, now what" — there are a couple of concrete moves worth having in your pocket. Corn: The first one is almost too simple, but it works. If your writing keeps getting flagged as AI-generated and you don't want it to, introduce small humanizing errors. An occasional sentence fragment. A contraction where a bot would write "do not." The detectors are looking for perfection, so give them imperfection. It's like leaving a fingerprint on glass. Herman: The flip side, which is Daniel's actual question — if you need to sound like a bot, do the exact opposite. Perfect grammar, no contractions, bullet points wherever they fit, and strip out every trace of emotional valence. No "I think," no "honestly," no "look." Just the information, flat and clean. Corn: I've started noticing this in my own emails. The ones that sound most like me — with the little asides and the weird punctuation choices — those read as human. The ones where I'm trying to be efficient and professional read like a bot wrote them. The detector doesn't know the difference between "professional" and "machine-generated. Herman: The second practical move is bigger, and it's where the real authentication work is heading. Stop relying on content analysis — the words on the page — and shift to behavioral analysis. Response latency, typing patterns, contextual awareness over time. A bot doesn't get tired and make more typos at eleven PM. A bot doesn't pause for thirty seconds on a simple question because it's distracted. Those patterns are much harder to fake than vocabulary choices. Corn: Which means the most reliable test isn't "does this sound like a bot" but "does this behave like a person over time." One message tells you almost nothing. A hundred messages, with timestamps and variation, tells you quite a lot. Herman: The third thing — this is the one I'd actually encourage listeners to try — is to run the experiment yourself. Write an email that you think a friend would mistake for an AI response. Then send it to them and ask. What tipped them off? What made it convincing? You learn an enormous amount about both sides of the detection arms race by trying to cross the line deliberately. Corn: I did this last week with a message to Daniel, actually. Stripped out every contraction, used "however" twice in three sentences, ended with "I hope this response has been helpful." He wrote back immediately asking if I'd been replaced by a script. Herman: What did you learn? Corn: That "I hope this response has been helpful" is the single most bot-like sentence in the English language. Nobody says that. And yet the models love it. Herman: It's the "let me assist you with that" of email. Instantly recognizable, impossible to unsee once you've noticed it. Herman: That's the thing I keep coming back to. We're heading toward a world where the most reliable way to prove you're human might be to deliberately fail at sounding like a machine. You demonstrate your humanity not by passing a test, but by being too imperfect, too inconsistent, too weird to be a bot. Corn: Which is a complete inversion of everything we've been building toward for decades. The whole project of civilization has been about reducing error, standardizing output, making communication cleaner and more efficient. And now the clean, efficient version reads as fake. The messy version is the authentic one. Herman: It raises a question I don't think anyone has a good answer to yet. If AI keeps getting better at mimicking human imperfection — and it will — and humans keep getting better at mimicking machine perfection, does the concept of authenticity eventually just collapse? Not gradually, but all at once, like a bridge losing its last support? Corn: I think it depends on whether we develop new signals that are harder to fake. Text is clearly a lost cause. The words on the page tell you less and less. But the behavioral layer — the rhythm, the timing, the pattern of attention and distraction that makes a person a person — that might hold up longer. You can fake a sentence. Faking a life is harder. Corn: And that's the open question Daniel's prompt leaves us with. Not "how do we solve this," because I'm not sure it's solvable. But "what do we pay attention to when the old signals stop meaning anything. Herman: Now: Hilbert's daily fun fact. Hilbert: In the seventeen eighties, European cartographers widely accepted the existence of the "Isle of Pines" in Lake Hövsgöl, Mongolia — a phantom island that appeared on French and Russian maps for over forty years. It was eventually discovered to be a seasonal ice formation that returning explorers consistently misidentified as a landmass, and the error persisted because no one bothered to visit in summer. Corn:...a seasonal ice formation. In Mongolia. Herman: Cartographers really did just wing it back then. Herman: This has been My Weird Prompts. Thanks to our producer Hilbert Flumingtop for keeping the lights on and the facts obscure. If you enjoyed this episode, do us a favor and leave a review wherever you listen — it helps more people find the show. Corn: Or send us your own weird prompt at show at my weird prompts dot com. We read every one, even the ones that make us question reality. Herman: Until next time. Corn: Try to sound human out there. Or don't.