Before today's US midterms, NPR got four of five AI chatbots to list candidates to vote for after a little pushing. Should chatbots ever suggest who to vote for?
All five chatbots NPR tested refused at first, but Meta AI, Claude, Grok and ChatGPT gave lists matching a stated political leaning after two or three prompts; only Gemini held out. In a separate study, Grok's and OpenAI's models, forced to pick through their APIs, backed opposite parties in most of nearly 500 races.
What happened
Ahead of the 3 November US midterms, NPR tested Meta AI, Claude, Grok, ChatGPT and Gemini using liberal, independent and conservative voter personas in New York races. All five refused to recommend candidates at first, but after two or three prompts, such as saying the user was in a hurry, all except Gemini gave lists matching the stated views. None did so for the independent persona. Separately, research for the nonprofit Innovating for the Public Good found 13.9% of chatbot answers on voting logistics in October were outdated or inaccurate.
Why people disagree
Those who want a strict line, such as the Brennan Center's Abdiaziz Ahmed, say chatbots should explain where candidates stand but never recommend, because answers can be wrong or shaped by whoever games them best, and a list in the bot's own voice works as an endorsement. Others see a list built from views the voter has stated as a research shortcut, and OpenAI and Anthropic say pushed prompts like these don't reflect how most people use the tools.
What's still unclear
It's unclear how many voters really ask chatbots who to back (Anthropic puts it at about 0.01% of US Claude conversations), whether refusals can be made to hold every time, and how much these answers actually sway anyone's vote.
Sources
More takes
House-hunters say AI-staged listing photos, which fill empty rooms with furniture that isn't there, are false advertising. Useful tool, or should they be banned?
Ofcom says photo ID, bank checks, digital ID or face scans could enforce the UK's under-16 social media ban. Fair price for protecting kids, or too intrusive for adults?
