Safeguards against manipulation

On 14 October, Anthropic detailed a new set of safeguards for Claude ahead of the US midterm elections. The company says it added classifiers intended to recognize multi-step schemes, including networks of fake accounts and sites, financially motivated manipulation and activity by foreign state actors. Later in October, Claude.ai for US users is expected to bring official Democracy Works information about registration, polling places and identification requirements directly into answers. On election night and afterward, the service also plans to display current official results.

Data on political queries

In Anthropic’s own sample of about 250,000 US conversations over a week, 5.75% concerned politics and public issues, 0.51% elections, and roughly 0.01% directly asked for candidate advice. External evaluators also found unsolicited caveats and uneven treatment of political arguments. Splitting a harmful campaign into apparently innocuous requests could evade some safeguards.

Sources