AI answers are a new vector for election disinformation
For years, people have used social media to seed disinformation in order to swing elections. Now, the answers our search engines and expert systems produce have become more vulnerable.
The dangers of AI-generated answers in search results are enumerated here in the context of the recent Scottish election. 75 questions about the ballot were posed on AI systems, and the findings were sobering. On average, 44.4% of responses were at least partially wrong; ChatGPT was specifically wrong 46.2% of the time. Improbably, Grok actually performed the best in this test, with only 8.97% of responses containing factual inaccuracies.
These findings held true in other UK elections:
“Days before the Senedd election in Wales, also held on May 7, BBC Wales tested six major chatbots, ChatGPT, Copilot, Gemini, Claude, Meta AI, and Grok, against fictional voter profiles. The results echoed Scotland’s almost exactly.”
Some of this is likely just an outcome of how these systems work: hallucinations are par for the course. Nobody knows how to build a completely accurate LLM. But it’s also relatively easy to seed bad information as training data, which can be used to whitewash disinformation: you establish the intentional falsehoods, the LLM poisons its training corpus by ingesting your information, and it then presents your falsehoods as fact.
Researchers at NYU discovered that if 0.001 percent of the training data of a given LLM is “poisoned,” or deliberately planted with misinformation, the entire training set becomes likely to propagate errors. You can do that just by hosting harmful information online: in blogs, on Reddit, in otherwise-trustworthy sites. Once the falsehoods have been internalized by the LLM, the end user will never know.
As the linked article points out, between 7 and 13% of voters in these elections used a chatbot to figure out who to vote for. We’re at the foothills of AI use; these numbers will continue to rise. And with them, a real risk to maintaining an informed democratic voting population.
