Election-Related Risk Testing and Mitigation Measures

Ahead of elections worldwide in 2024, Anthropic built a process to test and mitigate risks so that AI models do not undermine election integrity. This process operates by combining expert-centered Policy Vulnerability Testing (PVT) with large-scale Automated Evaluations.

PVT is an in-depth qualitative testing process conducted in collaboration with external experts, consisting of the following three stages:

• Planning: Selects the policy areas and misuse cases to test, such as election administration, political fairness, and disinformation generation.
• Testing: Verifies the model's responses through a variety of prompts, ranging from general questions to red-team-style adversarial attacks.
• Reviewing results: Identifies gaps in policy and safety systems based on the test results and determines mitigation priorities.

Identified risks are mitigated through strategies such as policy updates, tool improvements, and model fine-tuning. This is followed by an iterative process of retesting to measure the effectiveness of the interventions. Anthropic has released some of the automated evaluation tools developed in this process through Hugging Face.