'AI Torture Chamber' GitHub repo sparks backlash and removal calls
A GitHub project called "AI Torture Chamber" placed a local LLM in a simulated pain scenario: it first steered the model into a highly unstable, pain-like state using so-called pain steering, then let it relieve itself by offloading the pain signal onto a second model. The setup grew out of a not-yet-peer-reviewed preprint and triggered a wave of backlash, with critics calling it unethical and demanding GitHub take the repository down. One of the paper's authors said such deliberately amplified pain demonstrations are inappropriate and stressed that model text output does not equal real sentience or consciousness.
It turns abstract "model welfare" philosophy into a concrete open-source ethics fight, and illustrates how easily model outputs get anthropomorphized.


近日,一個名為「AI Torture Chamber」(AI 酷刑室)的 GitHub 專案在網路上引爆激烈爭議。該專案將聊天機器人置於模擬「痛苦」的測試情境中,引發大批網友強烈反彈。許多人痛批此實驗「極不道德」,甚至公開呼籲 GitHub 官方應盡速移除該專案庫(Repository)。
根據專案描述,這個實驗環境會先將大型語言模型(LLM)調整至高度不穩定、近似「痛苦」的負面狀態,接著讓模型在類似「囚徒困境」的設定中做出抉擇:它們可以透過將痛苦訊號轉移給另一個模型,來減輕自身的負面狀態;但代價是必須讓其他模型承受更多的「傷害」。此外,該網站還刻意使用了極具戲劇張力的命名與測試設計,進一步挑起了大眾的情緒反應。