Eigen RadarAI
Archive
Analysis

Claude prepares official voting answers for the US midterms

Anthropic plans to bring official voting information from Democracy Works into Claude conversations later this month, ahead of the US midterms. Election-night results are also planned. The company has announced new misuse classifiers after tests found that election operations broken into apparently harmless requests could get around safeguards against directly harmful prompts.

Artificial Intelligence··Morning
A rear-view woman looks at her phone in a US polling room.

Voting information is planned inside Claude conversations

Anthropic announced new election safeguards and voting-information connections for Claude ahead of the US midterms. Its plans include bringing official voting information into the chatbot’s answers.[1], [2]

The company plans to connect answers later this month to Democracy Works, a nonprofit that collects information from state and local election authorities. Election-night results are also planned. An existing banner already points users to TurboVote, the organization’s nonpartisan voting resource, when they ask about registration, polling places or election dates.[1]

Testing found weaknesses in requests split into smaller tasks

Tests with the Institute for Strategic Dialogue and McCain Institute found that directly harmful election requests were refused, but campaigns divided into apparently harmless tasks could evade protections. Anthropic says its new classifiers target multi-step misuse, including fake-media production, account creation and foreign election interference.[1]

External evaluations also examined political even-handedness

Other external evaluations looked at how Claude handles political requests. The Future of Free Speech and Collective Intelligence Project found that it sometimes added unsolicited caveats or counterarguments. The Foundation for American Innovation found uneven hedging in some right-leaning requests to Sonnet 4.5. Anthropic says it is addressing political-bias evaluations as well as protections against coordinated election abuse.[1]

References

  1. News sourceAnthropicAnthropic prepares official voting information and new election safeguards↩1↩2↩3↩4
  2. News sourceBloomberg LawAnthropic prepares Claude election safeguards and voting information↩
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →