Skip to content
TheChatArchiveThe Chat Archive
Perplexitysonar-reasoning· PerplexityHugging Face LMArena Search Arena 24K

Which llm makes the better slop content?

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Geminigemini-2.0-flash-grounding’s answer.
  1. User

    Which llm makes the better slop content?

  2. Perplexity Assistant

    Models that generate "slop"—low-quality, formulaic AI content—tend to be those optimized for user-friendliness and trained on synthetic data. Here's a breakdown:

    Top slop-producing models

    1. GPT-4/4o/4 Turbo
      Dominates slop production due to widespread use and training on vast amounts of web data, including AI-generated content. Known for phrases like "delve into" and "rich tapestry"[1][7]. Its speed and coherence make it a slop factory, especially in free-tier implementations where output quality controls are looser[7].

    2. Claude (Anthropic)
      Generates verbose, overly polite text with frequent rephrasing (e.g., substituting "key controls" with "important controls")[5]. Its tendency to overexplain simple concepts contributes to slop characteristics[1].

    3. Llama 3 70B
      While more capable than earlier models, its fine-tuning on synthetic data leads to generic corporate-speak outputs. Users report it struggles to break free from safety-trained response patterns[1][6].

    Less slop-prone alternatives

    • LLaMa 33B/65B base models
      Untuned versions avoid synthetic data slop but require significant prompt engineering to be useful[1].
    • DeepSeek V2 base
      Claims to exclude AI-generated training data, though its chat version still exhibits slop[1].

    Key factors in slop generation

    • Synthetic training data: Models trained on AI outputs exhibit model collapse, progressively losing lexical diversity[4][6].
    • Safety fine-tuning: Overly constrained models default to bland, risk-averse language[1][6].
    • Commercial optimization: Models prioritizing engagement (e.g., ChatGPT) favor fluent but vacuous responses over substantive content[7][8].

    Ironically, the most "capable" models often produce the worst slop due to their training on AI-polluted datasets and alignment processes that strip away stylistic variation[4][6]. As Simon Willison notes[7], slop isn't inherently about model quality but about context—using LLMs to generate unreviewed bulk content inevitably leads to slop, regardless of the underlying technology[7].

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.