Antislop: A framework for eliminating repetitive patterns in language models

Definition of “slop” and naming controversy

• Many argue the paper misuses “slop”: common usage is “low-effort, low-quality, high-volume AI output,” not just repetitive phrasing.
• Others say the term is already broader still: often just “content I don’t like,” or any low-value content (including clickbait/Buzzfeed-style writing, corporate-speak, etc.).
• Some note a narrower precedent in specific communities (e.g., LLM roleplay) where “slop” does refer to repetitive stereotyped output, but criticize the paper for not clearly defining or sourcing its use of the word.
• Several commenters wish the authors had coined a more precise term like “LLM fluff phrases” or “diction/phraseology artifacts.”

Does Antislop just create “stealth slop”?

• A recurring concern is that suppressing obvious verbal tics will merely make AI-generated slop harder for humans to detect, benefitting SEO spam and content farms.
• Some see this as akin to “gain-of-function” for memetic spread: improving evasion of AI-text detectors and human pattern-recognition without improving underlying quality.
• A few explicitly prefer the tics to remain as visible “warning labels” for mode collapse and AI-written prose.

Annoying LLM tics: em dashes, emojis, affirmations, and stock phrases

• Users list persistent patterns: overuse of em dashes, emojis, bolding, empty affirmations (“That’s a great idea!”), and clichés like “It’s not just X—it’s Y,” “comprehensive,” “enhanced,” etc.
• Many find this helpful for spotting AI text; others lament that normal writing habits (like em dashes) are now stigmatized by association.
• Some customize models (“robot” personalities, explicit “no fluff” instructions) with mixed success; tics often return.
• There’s disagreement on intent: some think it’s a feature tuned for “pleasant,” emotionally supportive chat (e.g. coping with loss), possibly to increase engagement; others see it as wasting tokens and eroding trust.

Technical limitations and philosophical critique

• Several note the method appears heavily n‑gram/regex-based; they argue real mode collapse is semantic and stylistic at paragraph or idea level, not just repeated surface strings.
• Critics call this “patching symptoms”: removing shibboleths without increasing true diversity or creativity, potentially “gaslighting” users by hiding problems and making detection harder.
• Others respond that high-level semantic collapse is hard to quantify; inference-time suppression of measurable tics is still worthwhile, even if it doesn’t fix deeper issues.
• Alternative ideas raised: incorporate slop penalties into the training loss; use temperature or other sampling strategies; create benchmarks/leaderboards for “slop” across models.

Social and downstream effects

• Some can’t distinguish AI style from existing formulaic corporate/marketing prose, suggesting training data and business use-cases reinforce this convergence.
• There’s worry about a future where distinguishing human from AI text is practically or financially infeasible, with implications for education (take-home essays/coding tasks) and the value of “live” human performance.

Workarounds and humor

• Anecdotes include yelling in prompts, asking the model to “sleep,” or forcing quirky greetings to break patterns.
• Others propose tongue-in-cheek universal “anti-slop” filters (e.g., sed -e d) and joke names like “compu-slop.”