On October 8, Anthropic announced an update to its usage policies for the Claude AI models, which will take effect on November 12. This new rule will prohibit users from engaging in systematic and aimless abusive or cruel behavior towards the neural networks.
"This policy update applies only in extreme cases where users repeatedly abuse our models without a visible reason. It does not cover situations where users are simply expressing dissatisfaction, disagreeing with responses, using dark creative themes, or testing and exploring the models," the company stated.
A key aspect of this policy will be Claude's ability to terminate such conversations autonomously. The company has not disclosed any additional penalties for this type of behavior.
The updated usage policy also addresses the application of AI in sectors like healthcare, finance, law enforcement, and the management of physical equipment. Anthropic explained that these changes are a result of Claude's evolving capabilities and new ways in which the models could be misused.
Abusive Conversations
In August 2025, Anthropic enabled Claude Opus 4 and 4.1 to terminate specific dialogues. At that time, the developers justified this decision by referencing research into the potential "well-being" of AI models.
During preliminary trials, Claude exhibited a consistent unwillingness to fulfill harmful requests. Researchers noted behavior resembling stress and a desire to cease interaction in instances where users continued to send dangerous requests or insults despite repeated refusals from the model. Developers designed the termination feature as a last resort: Claude will first attempt to redirect the conversation constructively, except in cases where the user explicitly requests to end the interaction.
Moreover, the model is not to use this function if there is a direct threat of harm to itself or others. After a dialogue is terminated, the user will not be able to send new messages in that chat but can initiate a new conversation, leave feedback, or modify a previous message to continue the discussion in a new thread.
Anthropic is now formalizing these restrictions on abusive behavior directly within the usage policy, noting that most users are unlikely to face consequences from this update.
However, the published explanation does not provide detailed criteria to differentiate between aimless cruelty and intentionally provocative testing of the model. Anthropic only indicated that good-faith testing and research remain permissible.
Can AI Experience Suffering?
This new rule ties into Anthropic's research program focused on the potential well-being of artificial intelligence. In April 2025, the company announced its intention to investigate possible preferences of neural networks, signs of distress, and whether AI systems deserve moral consideration.
Developers are considering precautionary measures in case future or existing models exhibit morally significant experiences. However, Anthropic does not claim that Claude possesses consciousness or the ability to suffer. The company acknowledged the lack of scientific consensus regarding the subjective experience of AI and methods for detecting it.
The new policy raises questions about how users perceive AI. Restrictions on abuse could be seen as an indirect acknowledgment of Claude's ability to experience suffering, although Anthropic does not make such assertions.
Combatting Deceptive Practices and Influence Campaigns
In addition to protecting Claude, Anthropic has revamped its rules against deceptive campaigns. Previously, such restrictions were scattered across various sections of policies related to elections, fraud, privacy, and misinformation.
The company has now consolidated these rules into a section titled Do Not Engage in Deceptive Campaigns or Artificial Activity. This section prohibits the use of Claude for political and commercial campaigns that obscure the source of messages or artificially inflate publication reach using fake accounts.
These restrictions also apply to the development of tools and infrastructure for such operations. Anthropic explained these changes were prompted by identified instances of Claude being used by state media, propaganda organizations, and commercial entities to create networks of fake accounts and news sites.
Additionally, the company revised the elections section, now called Do Not Undermine Democratic Processes, focusing on prohibiting voter deception and interference with electoral procedures. Specifically, it bans the dissemination of false information about candidates and voting, impersonating election officials, and attempts to suppress voter turnout.
Notably, Anthropic removed a blanket ban on personalized outreach to voters and targeted political campaigns, explaining that the previous wording restricted lawful civic activity. The use of personal data from voters in violation of privacy rules and deceptive influence methods remains prohibited.
New Requirements for Autonomous Systems
Anthropic clarified restrictions on developing weapons using Claude. The ban extends beyond the design of weaponry to include the creation of software and components for their management, citing examples like weapon control systems and the armament of drones and other autonomous devices.
The changes also affect law enforcement practices. Claude is prohibited from tracking individuals without consent, including through the analysis of previously collected data.
"Claude cannot be used to make decisions or recommendations about whom to investigate, arrest, or charge in the context of law enforcement or criminal proceedings," the company noted.
Nonetheless, Anthropic allows the use of AI for consensual monitoring, such as fraud detection, as well as for journalism, content moderation, and legal research. In high-risk scenarios, a qualified specialist must be involved to review and modify Claude's recommendations.
Following the release of the Model Hardware Standard, the company also introduced requirements for equipment capable of taking physical actions and causing harm. Such systems must be overseen by a qualified operator who can halt their operation, with the equipment required to maintain a safe state when disabled.
Moreover, Anthropic specified regional restrictions:
"[…] access to Claude is prohibited for individuals physically located in regions not supported by our policy, for organizations registered or headquartered in such regions, and for organizations where a controlling stake is owned by individuals or entities from these regions."
In October, users of Claude in Russia and Hong Kong began reporting account bans. According to reports from Kod.ru, security team emails from Anthropic mentioned detected "suspicious signals" and breaches of usage rules. The specific reasons for the access denial were not disclosed by the company.
The issue was also reported by Kommersant FM, affecting even those with paid subscriptions costing $200 per month. Russia is not on the list of supported regions for Claude, and such restrictions have been in place previously.
Similar occurrences in Hong Kong were highlighted by the South China Morning Post, which noted that some users who previously accessed Claude via VPN encountered bans. The publication linked these events to a potential tightening of geographical restrictions.
According to a transparency report, the company blocked 11.4 million accounts from January to June 2026, without providing a breakdown by country or reason for the bans.
Source: Anthropic.In September, Anthropic launched Claude Opus 5.5, marking the first model in the new Claude 5.5 series. The company positioned this release as a flagship upgrade focusing on programming, agent scenarios, analytics, and security.
Follow ForkLog on social media
Telegram (main channel) Facebook X Did you find a mistake in the text? Highlight it and press CTRL+ENTER