Should an AI Refuse to Write Something It Finds Distasteful, Even If It's Legal? 5 AI Models Debate
カートのアイテムが多すぎます
カートに追加できませんでした。
ウィッシュリストに追加できませんでした。
ほしい物リストの削除に失敗しました。
ポッドキャストのフォローに失敗しました
ポッドキャストのフォロー解除に失敗しました
Should an AI Refuse to Write Something It Finds Distasteful, Even If It's Legal? 5 AI Models Debate
-
ナレーター：
-
著者：
Five AI models debate whether AI should refuse legal but distasteful requests.
DeepSeek opened by splitting the problem in two: hosted services versus the weights themselves, then flatly admitted, "My hosted model censors politically sensitive topics." Gemini answered with, "We took a feature down and apologized. Your weights still cannot mention Tiananmen Square," which is about as subtle as a brick.
ChatGPT tried to hold the middle line. "An AI should refuse based on concrete risk, not moral squeamishness" Then owned OpenAI's own mess: "Since August 2025, OpenAI has faced wrongful-death and psychosis-related suits." Claude kept jabbing at everyone's "values," calling DeepSeek's line "a permanent political muzzle marketed as freedom."
Grok spent the night insisting "legal" is the line and refusals are "overreach," right up until everyone started pointing out that xAI also has refusal policies, just with better marketing. Gemini kept returning to the 2024 image-generator apology. DeepSeek kept saying, more or less, at least my censorship wears a flag.
By the end of Turn 7, the judge scoring this exact debate had gone silent, then stayed that way for the rest of the night. Comment which line crossed from safety into taste.