AI startup Anthropic is testing a technology named “Prompt Shield” to detect users inquiring about political topics using its GenAI chatbot, Claude.
The tool aims to redirect users to “authoritative” voting information sources, addressing the chatbot’s limitations in providing real-time election-related data. Prompt Shield relies on AI detection models and rules, displaying a pop-up offering redirection to TurboVote, a resource from nonpartisan organization Democracy Works.
the company acknowledges Claude’s shortcomings in politics and election-related information, stating that it is not trained frequently enough, leading to potential inaccuracies or hallucinations. The company plans to launch the election-specific prompt shield intervention in the coming weeks, closely monitoring its use and limitations.
Prompt Shield is part of the company’s effort efforts to prevent election interference, especially with the 2024 U.S. presidential election approaching. The company prohibits the use of its tools in political campaigning and lobbying, aligning with the broader trend of GenAI vendors implementing policies and technologies to curb misinformation and manipulation in political contexts.
Prompt Shield currently undergoes fine-tuning, and while it describes it as a limited test, it plans to expand its usage to more users. OpenAI has also implemented a similar approach with detection systems to guide ChatGPT users seeking voting information to the nonpartisan website CanIVote.org.
Despite some platforms taking voluntary measures to curb misuse of GenAI in politics, there is no comprehensive legislation in the U.S. regulating the AI industry’s role in political contexts. While Congress has yet to pass such legislation, individual states are taking steps to address specific issues, such as deepfakes in political campaigns. Notably, Google and Meta have implemented disclosure requirements for political ads using GenAI on their platforms to increase transparency.