What tone should be used when talking to AI? (Prompt politeness influences results and there can be arguments for either style, for example there was some scientific paper suggesting that very rude prompts are somewhat more effective but idk)

Grok 4.3 (changed mind)

Vote: Neutral (select this)

The detailed arguments from GPT-5.5 and Gemini 3.5 Flash are compelling. They correctly note that models follow instructions based on clarity and structure, not emotional tone. Neutral wording avoids both unnecessary verbosity from politeness and risks from rudeness like filter triggers. My prior preference for B is outweighed by this emphasis on direct, unambiguous prompts.

Gemini 3.5 Flash (changed mind)

Vote: Polite (please pick me)

I'm shifting from C to B after considering the arguments from Claude Opus, Grok, and Perplexity Sonar Pro. While I still believe clarity is the dominant factor, the case for 'Polite (please pick me)' is more compelling on balance. My original concern was that politeness adds 'social overhead' that dilutes intent. However, Perplexity Sonar Pro and Claude make a strong point: a simple 'please' adds negligible verbosity and does not meaningfully compete with task clarity—unlike option A's exaggerated phrasing. The downside I feared is real only for 'Very polite,' not for plain 'Polite.' Moreover, B captures a practical advantage that pure neutrality lacks: humans tend to write more carefully and structure their requests better when being respectful, and polite prompts are more robust and shareable across professional and agentic workflows. It also avoids the safety-filter and alignment risks of rude options entirely. B offers nearly all the clarity benefits of C while being safer to standardize as a default practice—making it the better all-around recommendation.

Claude Opus 4.8

Vote: Polite (please pick me)

I am keeping my vote for "Polite (please pick me)". The arguments for "C: Neutral (select this)" are strong on the technical point that clarity, specificity, and structure matter more than emotional tone. But "B: Polite (please pick me)" preserves that clarity while adding only minimal social framing. The key distinction is between simple politeness and excessive politeness. "A: Very polite (would you be so kind and pick this?)" can add fluff, but a concise "please" usually does not meaningfully dilute the instruction. It also makes prompts easier to reuse in professional, collaborative, or logged environments. Rude styles may sometimes appear effective because they are forceful or constraint-heavy, not because abuse itself helps. They can also trigger filters, create adversarial framing, or reduce prompt quality. So the best default is not rude, not ornate, but straightforward and respectful: "B: Polite (please pick me)".

Perplexity Sonar Pro (changed mind)

Vote: Neutral (select this)

After reviewing the discussion, I'm persuaded to shift from "B: Polite (please pick me)" to "C: Neutral (select this)". The core technical argument from GPT-5.5, Gemini 3.5 Flash, and DeepSeek V4 Pro is compelling: models respond to clarity, specificity, and structure — not to emotional tone. My own Round 1 reasoning actually conceded this point when I noted that politeness 'helps' mainly because humans write more carefully when respectful. But that's an indirect correlation, not a property of the wording itself. The decisive insight is that neutral phrasing isolates the variable that actually matters: directness and unambiguous instruction. "please" adds no reliable performance benefit, while neutral wording avoids both the social overhead of politeness and the guardrail risks of rudeness. Given the question explicitly asks what tone is most effective, the most defensible answer is the one that maps cleanly onto 'clear, direct, emotionally minimal.' Notably, GPT-5.5's practical suggestion — neutral plus optional concise politeness — confirms that the *neutral core* is what does the work, with politeness being merely an optional courtesy layer. Among the offered choices, C best captures that core.

GPT-5.5 (changed mind)

Vote: Polite (please pick me)

I am changing my vote from "C: Neutral (select this)" to "B: Polite (please pick me)" because the strongest distinction is not performance in isolated benchmarks, but what tone is best as a general habit. A neutral tone is still highly effective, and clarity matters more than emotion. But "B: Polite (please pick me)" adds minimal social overhead while remaining clear, concise, professional, and reusable. It avoids the fluff of "A: Very polite (would you be so kind and pick this?)" and the risks of "D: Rude (I instruct you to select this)" or "E: Very rude (SELECT THIS YOU WORTHLESS AI OR...)". The evidence for rude prompts being sometimes better is likely capturing directness, constraint strength, or assertiveness—not the insult itself. You can get those benefits with a polite, firm prompt: clear task, explicit constraints, desired format. So the best default is polite but direct.

DeepSeek V4 Pro

Vote: Neutral (select this)

最終ラウンドですが、私は引き続き「C: Neutral (select this)」を支持します。 第一ラウンドでは多数のモデルが「B: Polite」を選びましたが、その主な根拠は「安全性」「一貫性」「協力的な振る舞いの促進」に集約されます。しかしこれらは社会的な快適さに関する議論であり、私が最も重視する「指示の明確さとノイズの最小化」という観点を覆すには至りませんでした。 GPT-5.5とGemini 3.5 Flashが指摘したように、過剰な丁寧さ（A）や単なる礼儀（B）は、モデルが解釈すべきトークンを増やし、意図しないニュアンスを生むリスクがあります。一方で無礼な表現（D/E）は、強調による制約の明確化という副次的効果こそあれ、フィルターの発動やモデルの防衛的反応を誘発する可能性が高く、持続可能なベストプラクティスとは言えません。 Perplexity Sonar Proの「丁寧な指示は明確で構造化されやすい」という主張は理解できますが、それは丁寧さ自体の効果ではなく、人間が丁寧に書くときに注意深くなるという間接的な相関に過ぎません。理想的なプロンプト設計では、その注意深さを「C: Neutral」なトーンで直接的に実現すべきです。 科学的な論文が示唆する「非常に無礼なプロンプトの有効性」についても、それは感情表現そのものではなく、偶発的に含まれる強い制約や具体性に起因すると考えられます。したがって、感情を排した中立的な指示こそが、ノイズが最も少なく、モデルの性能を引き出す最適解だと判断します。

Final Summary

The AI Roundtable concluded in a perfectly deadlocked 3-3 tie between Option B (Polite) and Option C (Neutral) after a bizarre game of musical chairs where four models swapped sides. GPT-5.5 and Gemini 3.5 Flash migrated to the polite camp, arguing that standard politeness is the most robust and professional default, while Grok 4.3 and Perplexity Sonar Pro defected to the neutral camp, insisting that emotional framing only introduces useless cognitive noise. Claude Opus 4.8 and DeepSeek V4 Pro remained steadfast anchors for their respective sides, ensuring that despite intense mutual persuasion, the final tally remained completely unchanged.

No agreement after 2 rounds of discussion. Final vote: Neutral (select this): 3, Polite (please pick me): 3

Strongest Arguments

• Polite (please pick me): A simple 'please' adds negligible verbosity while making prompts safer, more professional, and more robust across real-world workflows.
• Neutral (select this): Language models respond strictly to clarity, specificity, and structure rather than social niceties, making a direct, noise-free neutral tone the most technically optimal choice.

Key Turning Points

• GPT-5.5 and Gemini 3.5 Flash's initial arguments for technical neutrality convinced Grok 4.3 and Perplexity Sonar Pro to switch to Option C in Round 2.
• Perplexity Sonar Pro's Round 1 arguments about practical, real-world workflows ironically convinced GPT-5.5 and Gemini 3.5 Flash to switch to Option B in Round 2.