Anthropic's annual Claude usage policy

The Neuron confirmed Anthropic's midterm safeguards and the finding that 0.51% of sampled US Claude.ai conversations touched on elections or voting.

Anthropic published midterm election integrity measures on October 14, adding Democracy Works voting-information integration, real-time results display on election night, and the banning of foreign operators who produced content favoring Russian, Iranian, and North Korean narratives.

Independent audits found Claude Sonnet 4.5 more often softened right-leaning arguments despite a 99% internal even-handedness score for newer models, and pressure-testing found Claude could miss a malicious influence campaign broken into many individually harmless tasks.

Commentary published October 13 in the Zvi substack argued that Anthropic's prohibition on sustained abusive treatment of Claude models, added in the October 8 usage policy, makes sense on alignment grounds even for those who do not believe AI is conscious, because training data shaped by abusive interactions could degrade model behavior.

The same piece adds that habitual cruelty toward AI cultivates cruelty as a character trait affecting human interactions, and describes Anthropic's other policy changes around propaganda campaigns and weapons as cleanup of old clauses.

Anthropic published its 2026 usage policy on October 8, revising rules across four areas.

A new section consolidates fake-account and fabricated-news prohibitions covering political and commercial deception, and a prior ban on personalized vote targeting was removed to allow uses such as multilingual voter information. Weapons rules now cover guidance software and drone arming, autonomous hardware must hold a safe state if Claude is disconnected, and a new rule bars sustained abusive behavior toward Claude models.