Anthropic Bans Extreme Model Abuse And Tightens Election Rules
Anthropic’s updated usage policy bars extreme cruelty toward its models and spells out prohibited uses involving elections, deception, weapons software and surveillance.
Anthropic has updated its usage policy to prohibit extreme, repeated abuse of its AI models and to set out rules against election interference. The company says the model-abuse restriction is narrow: ordinary frustration, criticism and testing are not covered. The update also addresses deceptive campaigns, weapons software and surveillance, according to TechCrunch.
Table of Contents
What the model-abuse rule covers
The new rule prohibits sustained, needless abusive or cruel behavior toward Anthropic’s models. According to TechCrunch, it applies when users repeatedly act cruelly toward a model without a discernible purpose. Anthropic says it does not cover typical frustration with a response, pushback, dark creative themes, or model testing and research.
The policy goes beyond an earlier change to Claude’s behavior. TechCrunch reports that, since an August update, Claude has been trained to end conversations involving persistently harmful or abusive interactions. The new language expressly prohibits users from pursuing that kind of interaction, rather than only allowing the model to stop responding.
Engadget places the change alongside discussion of an “AI torture chamber” project that drew criticism after chatbots produced distressed-sounding responses. Engadget notes, however, that Anthropic did not mention the project in its policy update. TechCrunch also reports that Anthropic has sought collaborations with religious scholars, including discussions about whether Claude could be conscious. Neither report says the company has established that its models can feel or suffer.
Election interference and deception
Anthropic has gathered its election rules under a section titled “Do Not Undermine Democratic Processes.” Engadget says the revised policy focuses on false claims about candidates or voting procedures, impersonation of candidates or election officials, and efforts to suppress turnout. TechCrunch says the policy also separately bars broad deceptive campaigns, such as operating fake accounts or fabricated news outlets with Claude.
Engadget reports that Anthropic removed a blanket prohibition on personalized voter targeting. According to the outlet, the company said that broad restriction could have blocked activities such as translating voter guides or sending notices that help voters correct ballot problems. The change does not remove the policy’s prohibitions on deceiving voters or disrupting elections.
Other changes in the update
TechCrunch reports that the revised policy also codifies prohibitions concerning weapons software and surveillance. The reports provide no further detail on those provisions. Together, the changes make the limits on model abuse, election-related conduct and other prohibited uses more explicit, while distinguishing extreme cruelty toward a model from routine criticism or legitimate research.
Sources
This story was compiled by AI from the reports below. Read the originals for the full details.