stopsloppypasta-ai-lzdmgr.txt (14505B)
1 [1]Skip to main content 2 3 Stop Sloppypasta 4 5 slop·py·pas·ta n. Verbatim LLM output copy-pasted at someone, unread, 6 unrefined, and unrequested. From slop (low-quality AI-generated content) + 7 copypasta (text copied and pasted, often as a meme, without critical thought). 8 It is considered rude because it asks the recipient to do work the sender did 9 not bother to do themselves. 10 [2]A few examples [3]Why it's rude [4]Guidelines to do better 11 12 You just got an unread message notification. 13 14 Maybe it's on Slack (or Teams), a text message, or an email. Maybe you were 15 tagged in Notion or an Office doc. 16 17 You open it to find several paragraphs of text - or perhaps a listicle - [5] 18 with all the hallmarks of AI-generated writing: headings, heavy formatting, and 19 "it's not X it's Y", with em-dashes sprinkled gratuitously throughout. 20 21 The person who sent it probably spent about ten seconds on it. They asked a 22 chatbot and forwarded its response to you verbatim without validation or 23 critical review. 24 25 But now you're expected to read, vet, and act on this wall of text. When 26 someone forwards text they themselves have not considered, they are asking you 27 to do work they chose not to do. The asymmetric effort makes it rude. 28 29 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 30 31 A few examples 32 33 ❌ Example 34 #strategy 35 36 The team is mid-discussion about whether to expand into the German market. 37 38 👤: 39 40 I asked Claude about this! Here's what it said: 41 42 Expanding into the German market presents both significant opportunities and 43 notable challenges. Germany is the largest economy in the EU and offers a 44 robust consumer base. However, businesses should consider the following key 45 factors: 46 47 1. Regulatory Environment 48 49 Germany has strict regulations including GDPR compliance, local labour laws... 50 51 [6 more sections follow. None mention your company, your product, your budget, 52 or reference the discussion already underway.] 53 54 buries the live discussion · unrequested · generic 55 56 The Eager Beaver 57 58 A conversation participant wants to contribute to the topic at hand, so they 59 ask a chatbot and share whatever comes back. The intention is good - they 60 genuinely want to help - but the wall of generic AI text they contributed 61 blocks the discussion already underway. Now other participants have to scroll 62 past it to continue, or stop to read and validate it. 63 64 It feels helpful to send. It creates work to receive. 65 66 ❌ Example 67 you: 68 69 Does anyone know why our email open rates have been dropping? We changed the 70 subject line format last month. 71 72 👤: 73 74 ChatGPT says: 75 76 Email open rate declines can be attributed to several factors. These include 77 changes in subject line strategy, sender reputation issues, list hygiene 78 problems, and deliverability concerns. Here are the most common causes: 79 80 1. Subject Line Fatigue 81 82 If subject lines have become too similar or predictable, subscribers may stop 83 engaging… 84 85 [Provides 5 more sections of generic email open diagnostics. Does not mention 86 your subject line change, your audience, or your platform.] 87 88 irrelevant to the specific question · generic 89 90 The OrAIcle 91 92 Someone asks a specific question. Another person puts it into a chatbot and 93 pastes the response as the answer. 94 95 "ChatGPT says" is the enshittified LLM-era equivalent of [6]LMGTFY (Let Me 96 Google That For You). Shared as a link or a GIF, LMGTFY was easy to ignore, and 97 clear about what it was (sarcastic commentary). Sloppypasta is neither. 98 Recipients are left to figure out whether it's AI generated, whether it's 99 correct, and which part actually answers the question (if it's actually 100 relevant at all). If you ask a person a question, you're looking for their 101 perspective and expertise. In this sense, both LMGTFY and sloppypasta are 102 etiquette failures where sender disregarded the recipient the dignity of the 103 basic human reply. 104 105 ❌ Example 106 👤: 107 108 Hey team - I did some research on our competitors this week. Here's a summary: 109 110 Competitive Landscape Overview 111 112 The market is highly competitive, with several established players and emerging 113 challengers. Key competitors offer distinct value propositions across pricing 114 tiers… 115 116 [It's a 5-page essay with handwavy assertions and no concrete details. No 117 dates. No sources. No live pricing.] 118 119 presented as personal work · no one knows to check · hallucinated details 120 possible 121 122 The Ghostwriter 123 124 The sender shares AI output as their own work, with no indication a chatbot 125 wrote it. Recipients have no reason to question it, and may act on information 126 that is out of date, incomplete, or simply wrong. 127 128 Using AI as a ghostwriter borrows the sender's credibility. If the content 129 turns out to be wrong, that credibility is what gets spent. 130 131 Why it's rude 132 133 As a Recipient As a Sender Feedback loop 134 Previously, effort to read Writing requires Sender's skipped 135 was balanced by the effort effort, which effort becomes 136 to write. Now LLMs make contributes to recipient's added 137 Effort writing "free" and increase comprehension. LLMs effort, increasing 138 the effort to read due to increase cognitive debt frustration as 139 additional verification by reducing struggle. incidence increases. 140 burden. 141 LLM propensity for 142 hallucination and What you share directly 143 capability to bullshit influences your Eroding trust from 144 Trust convincingly mean that reputation. Sharing raw LLM sloppypasta is 145 "trust but verify" is LLM output - especially the modern 'Boy Who 146 broken. All correspondence unvetted - burns your Cried Wolf.' 147 must be untrusted by credibility. 148 default. 149 150 Sharing raw AI output is like eating junk food: it's easy and may feel good, 151 but it's not in your best interest. You'll negatively influence your 152 relationship with the recipient, and do yourself a disservice by reducing your 153 own comprehension. 154 155 "For the longest time, writing was more expensive than reading. If you 156 encountered a body of written text, you could be sure that at the very 157 least, a human spent some time writing it down. The text used to have an 158 innate proof-of-thought, a basic token of humanity." 159 160 — Alex Martsinovich, [7]It's rude to show AI output to people 161 162 Before LLMs, writing took effort. Authors spent time and effort considering and 163 selecting their words with intention; time and effort that was balanced by that 164 spent by the audience as they read. This balance is broken with LLMs; the 165 effort to produce text is effectively free, but the effort required to read the 166 text hasn't changed. [8]The increasing verbosity of LLMs further increases the 167 effort asymmetry. In some circumstances (like pasting raw LLM output into a 168 chat thread), the sloppypasta effectively becomes a filibuster, crowding out 169 the existing conversation and blocking the viewport. 170 171 "Cognitive effort — and even getting painfully stuck — is likely important 172 for fostering mastery." 173 174 — Anthropic, [9]How AI assistance impacts the formation of coding skills 175 176 Writing is thinking. The writing process forces the author to work through 177 their thoughts, building their comprehension and retention. [10]Multiple [11] 178 studies have found that delegating tasks to LLMs creates cognitive debt. 179 Shortcutting thinking with LLMs ultimately reduces comprehension of and recall 180 about the delegated subject. 181 182 "A polished AI response feels dismissive even if the content is correct" 183 184 — Blake Stockton, [12]AI Writing Etiquette Manifesto 185 186 Before LLMs, trust was the default. Authors wrote from their personal expertise 187 and perspective, and readers could judge an author's understanding of the 188 subject based on the coherence of their writing. LLMs generate the most 189 probable next token given an overarching goal to be helpful, which explains 190 their propensity for hallucination ([13]confabulation) and why many people feel 191 that [14]LLMs are bullshit generators. Modern LLMs are typically provided tools 192 to help them look up grounding information that reduces (but does not 193 eradicate) their likelihood to outright make up facts during their responses. 194 But that still doesn't solve the trust problem; the reader still has no way to 195 know what the sender checked and what they didn't. LLM responses, therefore, 196 cannot be trusted by default and compound the effort asymmetry on the reader by 197 adding a verification tax. 198 199 Beyond accuracy, LLMs write authoritatively with the tone and confidence of an 200 expert. This adds further uncertainty to the reader's burden; they have no way 201 to gauge the sender's actual level of expertise with the subject matter. The 202 result is a further erosion of trust, because the AI's voice removes signal 203 that recipients previously used to distinguish expertise from 204 plausible-sounding slop. 205 206 "I think it's rude to publish text that you haven't even read yourself. I 207 won't publish anything that will take someone longer to read than it took 208 me to write." 209 210 — Simon Willison, [15]Personal AI Ethics 211 212 Formerly, "Trust but verify" ruled. Readers would trust until that trust was 213 broken; the author was trustworthy or they weren't. However, shared LLM output 214 obfuscates the chain of trust. Did the prompter do the appropriate due 215 diligence to validate the LLM response? If problems or errors are discovered, 216 who is to blame, the prompter or the AI? Was it an oversight, a missed 217 verification step, or was verification skipped altogether? The uncertainty 218 means the recipient doesn't know what they can trust, what has or has not been 219 verified; they must treat everything as untrusted. Just like the Boy Who Cried 220 Wolf, once the trust is broken, the uncertainty spreads to all future messages 221 from the sender. 222 223 Assumptions of balanced effort and presumed trust are no longer guaranteed in a 224 post-LLM world. Sloppypasta creates a compounding negative feedback loop where 225 the sender forfeits learning and credibility while the recipient burns effort 226 and loses trust. Receiving raw AI output feels bad due to the cognitive 227 dissonance of having these assumptions violated. 228 229 Read the full essay 230 231 Simple guidelines to do better 232 233 Read. 234 235 Read the output before you share it. If you haven't read it, you don't know 236 whether it's correct, relevant, or current. 237 238 Delegating work to AI creates cognitive debt. Working with the results helps 239 run damage control for your own understanding. 240 241 Verify. 242 243 Check the facts before you forward them. Anything you forward carries your 244 implicit endorsement -- your reputation depends on managing the quality of what 245 you share. 246 247 LLMs are trained to "be helpful", and will produce outdated facts, wrong 248 figures, and plausible nonsense to provide a response to your requests. 249 Further, an LLM is inherently out-of-date; their knowledge cutoffs contain at 250 best information on the state of the world when their training started (months 251 ago). 252 253 Distill. 254 255 Cut the response down to what matters. Distilling the generated response to the 256 useful essence is your job. 257 258 LLMs are incentivized to use many words when few would do: API-priced models 259 have a per-token incentive to train chatty LLMs that use many tokens, and [17] 260 research shows that longer, highly formatted posts are often preferred as more 261 engaging. 262 263 Disclose. 264 265 Share how AI helped. 266 267 If you've read, verified, and edited it, send it as yours -- preferably with a 268 note that you worked with AI assistance. If you're sharing raw output, say so 269 explicitly. In both cases, it may be useful to share your prompt and how you 270 worked with the AI to get the final output. 271 272 Disclosure restores the trust signals that sloppypasta destroys and tells the 273 recipient what you checked and what they may be on the hook for. 274 275 Share only when requested. 276 277 Never share unsolicited AI output into a conversation. 278 279 Remember that AI generations create effort asymmetry and be respectful of those 280 you share with. Sloppypasta delegates the full burden of reading, verifying, 281 and distilling to a recipient who didn't ask for it and may not realize the 282 effort required of them. 283 284 Share as a link. 285 286 Share AI output as a link or attached document rather than dropping the full 287 text inline. 288 289 In messaging environments, a large paste takes over the viewport and crowds out 290 the existing conversation. A link lets the recipient choose when - and whether 291 - to engage, rather than having that choice imposed on them. 292 293 AI capabilities keep increasing, and using it to draft, brainstorm or 294 accelerate you will be increasingly useful. However, using AI should not make 295 your productivity someone else's burden. New tools require new manners. 296 297 Use AI to accelerate your work or improve what you send. 298 Don't use it to replace thinking about what you're sending. 299 300 Further reading 301 302 • [18]It's Rude to Show AI Output to People 303 • [19]Personal AI Ethics by Simon Willison 304 • [20]AI Manifesto 305 • [21]Using AI Responsibly in Development & Collaboration 306 • [22]AI Writing Etiquette Manifesto 307 308 inspired by [23]nohello.net · [24]dontasktoask.com [25]open source 309 310 References: 311 312 [1] https://stopsloppypasta.ai/en/#main-content 313 [2] https://stopsloppypasta.ai/en/#types 314 [3] https://stopsloppypasta.ai/en/#why 315 [4] https://stopsloppypasta.ai/en/#rules 316 [5] https://tropes.fyi/directory 317 [6] https://lmgtfy.app/?q=what+is+lmgtfy 318 [7] https://distantprovince.by/posts/its-rude-to-show-ai-output-to-people/ 319 [8] https://epoch.ai/data-insights/output-length 320 [9] https://www.anthropic.com/research/AI-assistance-coding-skills 321 [10] https://www.media.mit.edu/publications/your-brain-on-chatgpt/ 322 [11] https://www.anthropic.com/research/AI-assistance-coding-skills 323 [12] https://www.blakestockton.com/ai-writing-etiquette-manifesto/ 324 [13] https://pmc.ncbi.nlm.nih.gov/articles/PMC10619792/ 325 [14] https://machine-bullshit.github.io/ 326 [15] https://simonwillison.net/2023/Aug/27/wordcamp-llms/#personal-ai-ethics 327 [17] https://arxiv.org/abs/2310.10076 328 [18] https://distantprovince.by/posts/its-rude-to-show-ai-output-to-people/ 329 [19] https://simonwillison.net/2023/Aug/27/wordcamp-llms/#personal-ai-ethics 330 [20] https://noellevandijk.com/ai-manifesto/ 331 [21] https://ai-manifesto.dev/ 332 [22] https://www.blakestockton.com/ai-writing-etiquette-manifesto/ 333 [23] https://nohello.net/ 334 [24] https://dontasktoask.com/ 335 [25] https://github.com/ahgraber/stopsloppypasta