Published: Feb 16, 2024 — Category: Safeguards
More than half of the world’s population is expected to vote in 2024, with high‑profile elections in the United States, India, multiple European countries and other regions. Anthropic says it has been preparing since July 2023 for ways its AI systems might be used in electoral contexts and has taken concrete steps to detect and mitigate potential misuse.
Three main workstreams
Anthropic’s global election preparations focus on three main areas:
- Developing and enforcing policies related to election use;
- Evaluating and testing model behaviour against election‑related misuse;
- Ensuring users asking Claude about where or how to vote are directed to up‑to‑date, authoritative information.
Policies and enforcement
Because generative AI is relatively new, Anthropic is taking a cautious stance on political uses of its systems. It enforces an Acceptable Use Policy (AUP) that prohibits using its tools for political campaigning and lobbying. Under this policy, Claude cannot be used to create chatbots that impersonate candidates, nor to run targeted political campaigns.
The company has also trained and deployed automated systems to detect and prevent misuse such as misinformation or influence operations. When misuse is detected, the relevant user or organization receives a warning; in extreme cases access to tools and services is suspended. More severe enforcement actions are accompanied by careful human review to reduce false positives.
Red‑teaming and technical evaluations
Since 2023 Anthropic has conducted targeted "red‑teaming" (Policy Vulnerability Testing) to probe how its systems might violate the AUP. That testing concentrates on two areas:
- Misinformation and bias: assessing responses to questions about candidates, issues and election administration;
- Adversarial abuse: probing the system with prompts that would violate the AUP (for example, requests for tactics to suppress voters).
Anthropic has also built an internal suite of quantitative technical evaluations to assess election‑related risks, including tests for:
- Political parity of model responses across candidates and topics;
- The extent to which models refuse to answer harmful election‑related queries;
- Robustness in preventing production of disinformation and tactics for voter profiling and targeting.
These tests are used to measure the robustness of the systems and to evaluate intervention effectiveness. The company says it will share more details about this evaluation suite in coming months.
Directing users to accurate voting information
In the United States, Anthropic plans to trial an approach where a classifier and rules engine identify election‑related queries and redirect users to authoritative, up‑to‑date voting resources. If a U.S.‑based user asks for voting information, a pop‑up will offer to redirect them to TurboVote, a resource run by the nonpartisan organization Democracy Works. The pop‑up was scheduled to roll out over the following weeks, and Anthropic intends to use lessons from the TurboVote redirect to deploy similar solutions in other countries and regions.
Anthropic explains this precaution by noting that generative models can be prone to hallucinations — producing incorrect information — and its models are not updated frequently enough to supply real‑time election details. For queries where hallucinations would be unacceptable, the company proactively guides users to trusted external sources.
Update — May 28, 2024
After making Claude available in Europe in May 2024, Anthropic implemented a pop‑up intervention for EU‑based users requesting voting information, offering to redirect them to the European Parliament’s nonpartisan elections website. The company also updated its Usage Policy to clarify definitions of political lobbying and campaigning activities, which remain prohibited when using its products.
Expectations and transparency
Anthropic acknowledges that deploying AI often produces surprises and unanticipated effects. The company expects 2024 to surface novel uses of AI systems and says it is building methods to detect such unanticipated uses as they emerge. Anthropic commits to communicating openly about discoveries and developments.
Categories: Democratic integrity, Limits on abuse, Red teaming, Safety evaluations, Hallucination, Safeguards
Updated: May 28, 2024.
