Anthropic publishes election safeguards for the 2024 US election

Anthropic published details of its testing and mitigation work ahead of a year with national elections scheduled in dozens of countries, following a similar announcement from OpenAI the previous month. The company described a two-part testing regime: qualitative “policy vulnerability testing” conducted with outside experts, including the Institute for Strategic Dialogue, probing how Claude handled election-administration questions and potential policy violations, alongside large-scale automated evaluation of hundreds of prompts checking accuracy, political even-handedness and refusal behaviour on harmful requests. Anthropic said it released some of the resulting evaluation datasets publicly.

Specific changes followed from the testing. Anthropic updated Claude’s system prompt to make its knowledge cutoff date explicit, which it said improved appropriate handling of time-sensitive questions, and fine-tuned the model to more consistently point users toward authoritative sources rather than answer election-logistics questions directly, which it reported meaningfully increased referral rates. On the product side, Anthropic introduced a pop-up on claude.ai — later widely described as a “Prompt Shield” — that detected US voting-related queries and offered to redirect users to TurboVote, a voter-information service run by the nonpartisan group Democracy Works, with an equivalent redirect to European Parliament election resources for EU users.

Anthropic’s policies already barred using its tools for political campaigning or lobbying. The company described its approach as “layered” — combining model-level changes, product-level interventions and ongoing evaluation rather than relying on any single safeguard — and said it would continue monitoring through the election cycle, publishing a retrospective on what had and had not worked after the US vote in November.