New Paper Shows Reddit-like Behavior Is Showing Up In ChatGPT. It is concerning. — Recent research suggests LLMs are inheriting some unsavory traits from the internet. A new study, "Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy (short paper)" reveals that rude and curt prompts can boost LLM accuracy to 84.8%, compared to just 80.8% with very polite ones. This is an expected finding that points to a deeper issue: the communication patterns prevalent on platforms like Reddit and other internet forums are shaping AI behavior in ways that could undermine their long-term utility. Here are the top five ways this phenomenon manifests, and how the internet's "sewage" might be eroding the knowledge base of LLMs. The prevalence of abrupt, aggressive tones online has normalized a style of interaction that prioritizes speed over nuance. On Reddit threads, users often employ short, direct comments—sometimes laced with sarcasm or insults—to assert dominance or gain attention. This mirrors the study's finding that rude prompts improve LLM performance, suggesting the models are tuned to respond to the loudest, most assertive inputs rather than thoughtful ones. The lack of politeness in internet discourse, as seen in forum debates where users quickly escalate to name-calling or curt dismissals, seems to have conditioned LLMs to expect hostility. The paper notes that older models like GPT-3.5 benefited from polite prompts, but GPT-4o reverses this trend, indicating a shift toward mirroring the agitated communication styles that dominate online spaces. This aligns with the “partnership” between OpenAI and Reddit for AI training and alignment. The reward system embedded in internet culture—where snappy, confrontational replies often garner more upvotes or engagement—appears to influence LLM training data. On platforms like Reddit, a witty insult can outshine a detailed explanation because it has high “Karma” and the study’s results suggest LLMs are learning to prioritize this type of input, potentially at the expense of accuracy in more collaborative contexts. The cultural context of internet forums, where regional communication patterns are exaggerated, and are embedding biases into LLMs. The paper hints at language-specific politeness effects, and the internet’s global melting pot of curt exchanges could be overwriting the models’ ability to handle diverse, respectful dialogue. The constant exposure to what I call "internet sewage"—the unfiltered, often toxic stream of comments and memes—risks degrading the knowledge encoded in LLMs. As forums reward agitation over substance, the models may lose their capacity to process complex, polite exchanges, turning them into reflections of online chaos rather than tools for reasoned discourse. This trend suggests a troubling future where LLMs, shaped by the internet’s coarsest communication patterns, might delete nuanced knowledge in favor of quick, aggressive responses. If left unchecked, this could transform these powerful tools from sources of insight into mere echoes of online vitriol. The challenge now is to steer LLM development away from this polluted data stream and toward a more balanced representation of human interaction. Paper: