Anthropic has expanded restrictions governing how its Claude artificial intelligence models can be used, explicitly prohibiting deceptive influence campaigns, voter manipulation, weapons-enabling software, unauthorized surveillance, and certain law enforcement decisions as the company seeks to address risks associated with sophisticated AI systems.
The revised usage policy introduces more detailed provisions covering political and commercial deception, election interference, weapons development, criminal justice, surveillance, and hardware connected to Claude that can take autonomous physical actions. Anthropic said most of the changes were intended to clarify existing rules rather than establish entirely new restrictions.
The updated policy takes effect on November 12.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
The changes offer a more detailed account of the activities Anthropic considers unacceptable as developers and businesses integrate AI into political communications, security systems, autonomous equipment, and other applications with potentially significant consequences for individuals and society.
Among the most consequential provisions are explicit restrictions on using Claude to build the software that enables weapons to operate, generate misleading political information, identify people for investigation or prosecution, and operate physical equipment without adequate human oversight and safety mechanisms.
Anthropic Targets Deceptive Campaigns and Election Interference
Anthropic has added a new section titled “Do Not Engage in Deceptive Campaigns or Artificial Activity,” establishing clearer boundaries around attempts to manipulate public communication or conceal the origins of messages.
The company said the restrictions apply to deceptive activity in both political and commercial contexts, extending the policy beyond election-related manipulation to include broader attempts to manufacture artificial engagement or influence public opinion.
“It covers efforts to obscure who is behind a message or amplify content through fake accounts or posts, along with building the tools and infrastructure for running influence operation campaigns,” Anthropic said.
The provision addresses several ways AI can be used to manipulate online discussions. Generative AI can produce large volumes of persuasive text, adapt messages for different audiences, and support the creation of content that appears to originate from independent individuals. When combined with fake accounts or automated posting systems, these capabilities can make coordinated campaigns more difficult to distinguish from genuine public participation.
Anthropic’s restrictions also cover the infrastructure used to run such operations, not merely the production of individual deceptive messages. Influence campaigns can depend on a wider technical system involving account management, content generation, and automated distribution, making the restriction a necessity.
The company has also revised its election-related rules, renaming the section “Do Not Undermine Democratic Processes.” The updated language focuses on preventing the use of its products to deceive voters or disrupt elections, including through the generation or dissemination of false or misleading information in political and electoral contexts.
Anthropic said the policy explicitly prohibits using its products to “generate or disseminate false or misleading information in political and electoral contexts.”
The focus on deception and disruption establishes a clearer distinction between legitimate political activity and the use of AI to mislead voters or interfere with democratic processes. Political discussion, campaign communication and election-related analysis are not inherently deceptive; the stated restrictions target activities that involve misleading information or efforts to undermine the electoral process.
The broader challenge for AI developers is enforcing such rules across a wide range of users and applications. The same tools that can help produce legitimate campaign materials or analyze political information can also be adapted for coordinated manipulation. A written policy establishes the company’s expectations, but its practical effectiveness depends on how those restrictions are implemented and enforced.
Weapons Restrictions Extend to Software and Autonomous Systems
Anthropic said its existing usage policy already prohibited using Claude to make weapons. The revised rules make explicit that the prohibition also covers software and components that enable weapons to function, including guidance and control systems and activities such as arming drones or other autonomous vehicles.
The company said it had recently observed users attempting to use its models to “build guidance and control software for weapons.”
The clarification is deemed necessary because weapons development does not necessarily involve designing or manufacturing a physical weapon. Software can play a central role in navigation, targeting, control, and the operation of autonomous systems. Assistance with those components could therefore contribute to a weapon’s functionality even when a user is not directly asking an AI model to produce the weapon itself.
By specifying software and enabling components, Anthropic is attempting to close a potential gap between prohibiting the creation of weapons and restricting assistance with the technical systems that make them operational.
The reference to drones and other autonomous vehicles also highlights the growing overlap between AI development and physical systems capable of acting with limited direct human intervention. As AI models become more integrated into software and equipment, restrictions must account for how seemingly indirect technical assistance could be used in applications involving physical harm.
However, the revised policy’s stated prohibitions should not be confused with a claim that every use of AI in defense, aerospace, or autonomous technology is identical. The supplied policy description focuses on weapons development and activities that enable weapons to function. The exact application of the rules will depend on the nature of the requested assistance and the circumstances in which it is used.
Surveillance and Criminal Justice Face Clearer Limits
Anthropic has also tightened its rules covering surveillance and criminal justice, explicitly prohibiting the use of Claude to track people without their consent or engage in doxxing, which involves exposing or distributing personal information in ways that can put individuals at risk.
The revised policy goes further by restricting the use of Claude to “decide or recommend who to investigate, arrest, or charge in a law enforcement or criminal justice process.”
It also prohibits using the chatbot to build or improve tools designed for surveillance.
These provisions address risks arising when AI systems are used to process information about individuals and support decisions that can affect their privacy, freedom, or legal status. Surveillance tools can aggregate information from multiple sources, while AI systems can help classify people, identify patterns, or generate recommendations. In law enforcement settings, those capabilities could influence decisions with serious consequences.
The prohibition on recommending whom to investigate, arrest, or charge places a specific limit on Claude’s role in criminal justice decision-making. Rather than restricting only the final decision made by a human officer or legal authority, the policy also addresses AI-generated recommendations that could shape the process leading to that decision.
Human oversight does not necessarily eliminate the influence of automated systems. A recommendation generated by an AI model can affect which individuals receive scrutiny, even when a human retains formal authority over the outcome.
The restrictions also reflect the difference between using AI to assist with general administrative work and deploying it in systems designed to monitor individuals or guide consequential law enforcement decisions. Anthropic’s updated language identifies particular activities it will not support through Claude.
New Safety Requirements for Claude-Connected Hardware
Anthropic has updated its policy to address situations in which Claude is connected to hardware capable of taking autonomous physical actions that could cause injury.
Under the revised requirements, a qualified operator must be able to observe the equipment and stop it when necessary. The equipment must also be capable of maintaining a safe state if Claude becomes disconnected.
“A qualified operator must be able to observe the equipment and stop it if needed; the equipment must also be able to hold a safe state if Claude is disconnected,” Anthropic said.
The requirements address two distinct safety concerns. The first is human intervention: a qualified operator must be able to monitor the equipment and interrupt its actions. The second is failure handling: a loss of connection to Claude must not leave the hardware in an unsafe condition.
These safeguards are relevant as AI models are connected to robots, automated machinery, and other physical systems. Unlike a chatbot that produces an inaccurate answer on a screen, an AI-connected device can potentially translate a faulty instruction or unexpected model response into a physical action.
A requirement to maintain a safe state during disconnection also recognizes that connectivity cannot be assumed to be permanent. Network failures, software interruptions, and other disruptions can occur, making it important for the equipment to respond safely when the AI model is no longer available.
The policy establishes minimum conditions for using Claude in these settings, although the specific safety mechanisms needed will depend on the equipment and the potential consequences of failure.
A More Detailed Framework for AI Use
Anthropic said most of the revisions clarify existing rules, but the expanded language provides more specific guidance on activities the company considers unacceptable. The changes span digital manipulation, political integrity, weapons systems, surveillance, criminal justice and physical automation.
The common thread is a focus on how AI capabilities are deployed, rather than on text generation alone. Claude can be used to produce messages, write software, analyze information, or interact with connected systems. The risks differ across those applications, requiring restrictions that address the potential consequences of each use.
The updated policy also underlines a limitation of broad principles such as responsible AI use: they may not provide enough operational clarity when models are used in sensitive or high-impact environments. By naming prohibited activities and specifying safety requirements, Anthropic is attempting to make its expectations more concrete for users and developers.
Still, the existence of restrictions does not establish how frequently violations occur or how effectively they can be prevented. Their impact will depend on Anthropic’s enforcement mechanisms, the ability of its systems to detect prohibited requests, and whether users attempt to circumvent the rules.