Anthropic Issues New Guidelines to Prevent AI Misuse in Elections
Dario and Daniela Amodei’s AI startup Anthropic has issued new measures to address the potential misuse of its AI systems in political contexts ahead of the 2024 elections.
The company's election preparations involve three key components:
Developing and enforcing policies: The Google-backed startup has established an Acceptable Use Policy (AUP) that strictly prohibits the use of its AI tools in political campaigning and lobbying. This includes preventing the creation of chatbots impersonating candidates and disallowing the use of its tools for targeted political campaigns. Automated systems have been implemented to detect and prevent misuse, with violators receiving warnings and, in extreme cases, facing suspension after human review.
Evaluating and testing model performance: Since 2023, Anthropic has been engaged in targeted "red-teaming" exercises to assess potential violations of its AUP. These Policy Vulnerability Tests focus on misinformation, bias, and adversarial abuse, evaluating how the AI system responds to election-related queries and inappropriate prompts. Quantitative tests assess factors such as political parity in responses, resistance to harmful queries, and effectiveness in preventing disinformation and voter profiling.
Providing accurate information: In the U.S., Anthropic is trialing an approach where its classifier and rules engine identify election-related queries and redirect users to accurate, up-to-date voting information from a nonpartisan organization, TurboVote. Recognizing the limitations of real-time training, especially in critical topics like elections, Anthropic guides users away from its systems when hallucinations or incorrect information could be generated.
A pop-up will offer users the option to be redirected to TurboVote when seeking voting information. This redirection strategy is planned for expansion to other countries and regions based on insights gained from the U.S. trial.
Major tech companies, including Adobe, Amazon, Google, IBM, Meta, Microsoft, OpenAI, and TikTok, have recently signed a voluntary pact to adopt "reasonable precautions" against the use of AI in disrupting global democratic elections. The pact was announced at the Munich Security Conference, with twelve other companies including Elon Musk's X, also joining the initiative.
The accord focuses on combating increasingly realistic AI-generated deepfakes, which deceitfully alter the appearance, voice, or actions of political figures and provide false information to voters.