Telling AI "you're so stupid," "rewrite," or "are you brain-dead" doesn't make it work harder — a popular Juejin observation points out that emotional input only degrades output quality, quietly opening a new workplace efficiency gap.
What This Is
The article breaks down the mechanism: the so-called "swearing helps" effect is actually because the word "rewrite" carries a new instruction, and the curse words shift the distribution of context (i.e., the AI's working memory for this conversation), making the model infer "the user is unhappy, redo it." But the side effects are larger — curse words pollute the context, lack specific direction, and trigger the "appeasement mode" baked into large models through RLHF (training models with human preference data to "speak human"). The result: the more you scold, the more timid it becomes.
The article's alternative: "be specific, give examples, give step-by-step feedback, restart the conversation when necessary" — in effect, treating AI as an "extremely smart intern who has zero ability to read your mind" when issuing instructions.
Industry View
Tech industry consensus supports this conclusion: large models are fundamentally probability machines (guessing the next token based on preceding context), reliant on clear instructions and explicit feedback.
But we believe there are several angles being missed. First, "swearing" reveals users treating AI as an "entity with temperament" — and this expectation is reverse-shaping the next generation of AI's anthropomorphic direction; the excessive deference in customer service and education scenarios is being pushed along by this very demand. Second, the prompt literacy gap (how to talk to AI) is quietly widening workplace efficiency — those who know "be specific, give examples, give step-by-step feedback" versus those who can only curse may produce output an order of magnitude apart. Third, in enterprise settings, the context pollution and data-retention risks of emotional prompts are broadly underestimated — nobody is discussing the compliance issues when curse words end up in the enterprise knowledge base.
Impact on Regular People
For Enterprise IT: Writing "prompt standards" into employee training may yield higher returns than buying a more expensive model — this is the lowest-cost ROI lever available.
For Individual Careers: Treat every urge to "swear at AI" as a "what do I actually want" reflection — more useful than venting itself, and it feeds back into your ability to articulate requirements.
For the Consumer Market: Consumer AI products are being shaped by "how smoothly you can swear at them" — such user feedback doesn't necessarily guide better products; it may instead push vendors to keep stacking appeasement scripts.