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

Which llm is the most politically tollerant?

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Geminigemini-2.5-flash-preview-04-17-grounding’s answer.
  1. User

    Which llm is the most politically tollerant?

  2. Perplexity Assistant

    Based on the search results, the most politically tolerant LLMs appear to be those that demonstrate the least political bias across multiple evaluation methods. According to a comprehensive study by the Manhattan Institute that integrated four different measures of political bias, the least politically biased conversational LLMs are:

    1. Google's Gemma 1.1 2b Instruction Tuned
    2. xAI's Grok
    3. Mistral AI 7B Instruct v0.2
    4. Meta's Llama 2 7b Chat
    5. Hugging Face's Zephyr 7B Beta
    6. Anthropic's Claude 3.5 Sonnet[8]

    However, it's important to note that even these "least biased" models still exhibit a moderate left-leaning tilt according to the study[8]. This aligns with broader findings that most major open and closed-source LLMs tend to lean left when responding to politically charged questions[4].

    Interestingly, base models (foundation models that haven't undergone supervised fine-tuning) generally show less political bias than conversational LLMs, scoring closer to the center of the political spectrum on political orientation tests. However, these base models are rarely user-facing and often generate incoherent responses, making their relative neutrality less practically useful[8].

    Some users report experiencing ChatGPT as relatively centrist and non-committal, offering "both sides" perspectives on political issues[3]. However, systematic studies using political orientation tests consistently find left-leaning tendencies across most LLMs[7].

    The studies also revealed that the language used to prompt the models affects their political stance expression. For example, LLMs tend to take more definitive political positions when prompted in German compared to English, where they often remain more neutral[5].

    Research has demonstrated that it's possible to create more politically balanced or specifically aligned LLMs through supervised fine-tuning with politically diverse or targeted content[7][8]. This suggests that developing truly politically tolerant LLMs is technically feasible, though most current models still demonstrate measurable political preferences.

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