Anthropic has removed the blanket prohibition on personalised vote and campaign targeting that its usage policy carried through the last election cycle. The company published the rewritten rules on 8 October and set them to take effect on 12 November 2026, with the old policy standing until then.
What was a broad election chapter is now headed Do Not Undermine Democratic Processes, and the new name signals a narrower scope. Anthropic says the section focuses the rules "specifically on disallowing the use of Claude to deceive voters or disrupt elections".
Automated calls or messages that conceal their artificial origin and go to voters at scale stay firmly barred. So does spreading false information in electoral contexts, impersonating candidates or election bodies, or intimidating people out of voting.
Aiming a message at a particular voter, by contrast, is no longer banned in itself. Anthropic says deceptive targeting and targeting that misuses voter data remain prohibited, so the line has moved from the practice to the intent behind it.
Americans vote on 3 November in elections covering all 435 House seats, 35 Senate seats and 39 governorships, and the new rules begin nine days later. Anthropic has not linked the two dates, and the gap is simply what the calendar produces. Anthropic gave no explanation for the date and named no election anywhere in the announcement.
OpenAI Still Forbids It
OpenAI's published restrictions prohibit individualised or segmented political messaging even when a human reviews it, bar assessing which voter or donor should receive which message, and rule out public-facing campaign chatbots. The two companies now sit on opposite sides of the same question.
OpenAI went further in May, saying it would allow no political advertising on its platform this cycle and would not permit its products to create scaled campaign messaging for or against a candidate. It also backed two bills in Congress on deceptive election AI.
Google's generative AI policy, last changed in December 2024, contains no election section at all. Its election commitments sit outside those rules, in a September 2026 post setting out advertising verification, mandatory disclosure of synthetic content and AI labels on YouTube.
Twenty-nine American states have laws in force on election deepfakes, meaning fabricated video or audio of real people, and courts have permanently blocked California's and Hawaii's on free-speech grounds. Congress has set no baseline at all, so each company writes its own rules. Europe took the opposite route for connected products, legislating first and leaving less to company policy.
Cruelty To Models Is Now A Rule
Anthropic added a prohibition on "sustained and needless abusive or cruel behavior toward our models", placed in the section covering cruel and psychologically harmful conduct alongside rules on harassment and self-harm. No major AI company had written such a rule before.
The rule does not cover ordinary frustration, pushback, dark creative themes or model testing and research. Anthropic says it is meant for extreme cases, where someone acts cruelly towards the models repeatedly and for no discernible purpose.
Claude's own ability to end a conversation, which closes that chat while leaving the account and other conversations untouched, will remain the main enforcement mechanism, the company says. Anthropic has published no account penalty attached to the rule.
Mustafa Suleyman, who runs Microsoft AI, rejects the premise underneath the rule. He wrote in September that "AIs do not have rights, feelings, or consciousness" and that "we must not train them to act as though they do", naming Anthropic in a post published three weeks before this policy.
What The Threat Reports Found
Anthropic's September threat report covered disrupted activity between December 2025 and August 2026, and described state media offices, government propaganda units and commercial firms using Claude. That work sits underneath the election changes.
Russian state media accounts used Claude as an editorial desk across Sputnik and RT newsrooms, and produced fabricated claims about Moldova's president before a vote. A France-based agency ran around 70 fabricated news sites and more than 250 fake commenting accounts, shifting political stance according to who was paying.
An Istanbul-based platform targeting Malaysia reached further still, running roughly 1,000 fake accounts segmented by race, religion and royalty. It republished state broadcasters as independent Malaysian reporting, and marketed itself as a military-grade political operations system.
Anthropic says it has seen attempts to use its models for weapons guidance and control software, and has documented AI being used to build systems that identify and track political dissidents. The new policy now prohibits arming drones and bars using Claude to decide who to investigate, arrest or charge.
A Policy Written Before A Listing
Anthropic is preparing to go public, and the numbers reaching reporters are large. A draft prospectus reviewed by Reuters, the document a company must publish before selling shares, showed 2025 revenue of $4.6B against an operating loss of $8.06B. The company is targeting a valuation above $2T.
The draft recorded $518B of long-term cloud commitments split between Amazon and Google, around 80% of which Anthropic owes whether or not it uses the capacity. Cash and short-term investments stood at $20.28B at the end of December 2025.
Reporting suggests the listing has slipped several times and may come after the November midterms, which is the same period in which the new usage policy begins to apply. Neither Anthropic nor its bankers have confirmed a date.
Rules That Nobody Has Tested
Nothing in the policy has been enforced yet, because none of it is in force. The election changes, the model-cruelty rule and the tighter surveillance language all begin on 12 November, and Anthropic has published no appeal process, notice period or account penalty attached to the new provisions.
No researcher, civil society group or competitor has responded publicly to the update, and every account of it so far rests on Anthropic's own words. What the document records is a company loosening one rule about political speech while adding another about how its software is spoken to.
