辱骂AI投诉更难解决,Claude收紧使用政策
Searchable分析了12个月内涉及Claude、Gemini和ChatGPT的逾800万条自愿参与消息,并称33%的美国聊天机器人用户曾辱骂或侮辱聊天机器人。其报告显示,辱骂式投诉与更低的问题解决率相关,不过两篇报道采用的指标不同:对话结束时的解决率为31%对41%,下一轮回复解决率为24%对35%。Anthropic还更新了Claude使用政策,自11月12日起禁止持续、无必要的虐待或残忍行为,严重违规主要通过终止对话处理。
该研究为具体、礼貌地提出问题提供了实践依据,而Anthropic的政策更新也明确了与Claude互动时虐待行为的边界。
- Anthropic has introduced new policies to prevent abuse of its Claude chatbot
- Bad language directed at chatbots is on the rise, according to research from Searchable
- Chatbots tend to resolve their errors more when users refrain from insulting them, according to the study
If you’ve ever got angry with your preferred AI chatbot, it could be worth rephrasing your angry response to the wrong answer. A report into bad behavior directed against AI tools has found that complaining with bad language or insults is less effective than querying the response with a straightforward statement.
The survey from Searchable found a wholly unsurprising 33% of US-based users swear or insult their chatbot. Compiling the data from an opt-in survey of 8 million messages exchanged with Claude, Gemini, and ChatGPT over 12 months, Searchable also found that residents of Nevada were more likely to insult their AI bots.
Following news of the survey, Anthropic has introduced a new set of usage policies to block cruelty in chats with Claude.
Tolerable behavior
Monitoring of the messaging with chatbots ran between September 2025 and September 2026, with anonymized data and results grouped by age and location. The initial results found that for every 10,000 words in an exchange with an AI chatbot, an average of 19.1 were found to be derogatory. This was where Nevada came out on top, with its score of 22.7 rude or offensive words.