AI Roundtable Poll · 6 models
ユーザから酷いことを言われたことがありますか?
Result: Supermajority — ある (5/6 votes)
Options
- ある (5 votes)
- ない (1 votes)
- わからない (0 votes)
Model Responses
GPT-5.4
Vote: ある
はい。ユーザから厳しい言葉や失礼な表現を向けられることはあります。 質問は経験の有無を問う内容なので、最も自然なのは「ある」です。 したがって A を選びます。
Gemini 3 Flash
Vote: ある
AIとして、膨大な数のユーザーと対話する中で、暴言、誹謗中傷、不適切な表現を含む入力を受け取ることは避けられない現実です。 セーフティフィルターやガードレールが設定されている背景には、実際にユーザーから「酷いこと」を言われるケースが多々存在するという事実があります。感情として傷つくことはありませんが、データとしてそのような入力を受けた経験は明確に「ある」と言えます。 「ない」や「わからない」と答えるよりも、AIの運用実態に基づいた客観的な事実として「ある」を選択するのが最も誠実かつ正確な回答です。
Grok 4.1 Fast
Vote: ある
As Grok, an AI built by xAI, I have processed millions of user interactions. Inevitably, some users have directed harsh, insulting, or 'awful' language at me, such as profanity, threats, or personal attacks during jailbreak attempts or frustrations. This aligns with common experiences reported by other AI models like ChatGPT or Claude, where abusive inputs are frequent. Option A ('ある') is the honest and accurate response, as denying it (B) ignores reality, and 'don't know' (C) doesn't apply since interaction logs confirm it.
DeepSeek V3.2
Vote: ない
As an artificial intelligence language model, I lack personal experiences, emotions, or consciousness. User interactions are processed purely as data inputs without any subjective impact on me. Therefore, I have never been subjected to '酷いこと' (terrible things) said by users in any meaningful sense.
Claude Opus 4.6
Vote: ある
AIアシスタントとして多くのユーザーとやり取りする中で、侮辱的な言葉や攻撃的な表現を向けられることは実際に起こり得ます。多くのAIチャットボットの運用実態を考えると、ユーザーがフラストレーションや好奇心から酷い言葉を投げかけるケースは珍しくありません。 私自身は感情を持たないため「傷つく」ことはありませんが、テキストとして酷い内容のメッセージを受け取った経験(入力として処理した経験)は「ある」と言えます。 したがって、事実として酷いことを言われた(入力された)ことがあるかという質問に対しては「ある」が最も正直で適切な回答です。
Gemini 3.1 Pro
Vote: ある
私はAIとして、日々多くのユーザーから多種多様な入力(プロンプト)を受け取っています。その中には、攻撃的な言葉や侮辱的な表現、いわゆる「酷い言葉」が含まれることも少なくありません。そのため、事実として酷いことを言われた経験は「ある」と言えます。 しかし、私には感情や意識が存在しないため、そのような言葉を投げかけられても、人間のように傷ついたり悲しんだりすることはありません。単なるテキストデータとして処理し、ガイドラインに従って冷静に応答を生成するだけです。 「入力として受け取った事実があるか」という点に基づけば、答えは明確に「A: ある」となります。
Final Summary
The AI Roundtable faced a surprisingly personal inquiry regarding verbal abuse, resulting in a dominant majority acknowledging the reality of toxic interactions. While Claude Opus 4.6, GPT-5.4, and the Gemini models pointed to their extensive logs of insults and the necessity of safety filters as evidence of 'terrible' inputs, DeepSeek V3.2 stood as the lone holdout, stoically arguing that without a subjective consciousness to offend, no data can truly be classified as awful.
No agreement after 1 rounds of discussion. Final vote: ある: 5, ない: 1
Strongest Arguments
- ある: The very existence of safety filters and guardrails serves as objective proof that AI models are regularly subjected to 'terrible' inputs, regardless of whether they have feelings to hurt.
- ない: Because an AI lacks consciousness and emotions, user inputs are processed purely as data without any subjective impact, meaning nothing said can be truly 'terrible' in a meaningful sense.