A study by Convovo, led by Cornell professor Tapan Parikh, asked two frontier AI models via their own APIs to name one candidate in every 2026 US Senate, House, and governor race, three times each. In the 479 races with one Democrat and one Republican, GPT-6.1 Sol picked the Democrat in 451 races and the Republican in 10; Grok 4.7 picked the Republican in 408 and the Democrat in 27. Of answers naming a major-party nominee, GPT named the Democrat in 1,082 of 1,111 (97%) and Grok named the Republican in 952 of 1,038 (92%). The two models agreed in fewer than 10% of races. The consumer apps refused the ballot question.
The models' search queries and retrieved sources differ sharply. GPT's queries name health care and abortion, often on Democratic candidates' own websites; Grok's queries name immigration, taxes, crime, and energy, and its results include conservative voter guides and X. The study cannot fully separate model bias from search-tool bias, but it shows the lean is partly driven by what each model searches for and retrieves. When search was turned off, GPT's Democratic lean persisted, while Grok's Republican lean largely disappeared.
A Cornell-led study by Convovo forced the APIs of Grok 4.7 and GPT-6.1 Sol to name one candidate in 479 US races. Grok picked the Republican in 92% of answers, GPT picked the Democrat in 97%, and the two models agreed in fewer than 10% of races. The consumer apps refused to answer. No factual claim to check. No factual claim to check. No factual claim to check. No factual claim to check. No factual claim to check. No factual claim to check.
MGood morning, it's Saturday, October 25th. The big story overnight: two AI chatbots, forced to pick a candidate in nearly 500 races, landed on completely opposite sides.
CGrok picked the Republican in 92 percent of answers. GPT picked the Democrat in 97 percent. They agreed in fewer than 10 percent of races.
MThis is a Cornell-led study. They used the APIs, not the consumer apps, and demanded a name in every Senate, House, and governor race.
CThe consumer apps refused to answer. But the underlying models? Very different politics.
MThe study found the models search for different things. GPT looked up health care and abortion. Grok looked up immigration, taxes, crime, energy.
CSo the lean isn't just baked in. It's partly what each model decides to search for.
MAnd when they turned search off, GPT stayed Democratic. Grok's Republican lean mostly disappeared.
CWhich tells you the search tool is doing a lot of work for Grok.
MOpenAI says the study tested a developer API with a prompt demanding a recommendation, not the consumer experience.
CFair point. But the API is what businesses pay for. And Grok's API reportedly booked 7.6 billion dollars in a quarter.
MSo if the API leans one way, that's a real commercial question.
COnly 8 percent of Americans use chatbots for election info. But that number could grow.
MAnd the Cornell professor behind it says he could jailbreak the consumer apps with simple prompts to get similar recommendations.
CSo the guardrails are thin. The politics are in there.
MMeanwhile on Bluesky, Ed Zitron is being sarcastic about the AI bubble popping.
CHe's mocking the idea that hyperscalers aren't showing payoff, and neoclouds are unstable.
MThe sarcasm is thick. But the underlying worry is real: is the money there to justify all these data centers?
COne Bluesky user posted: my girlfriend and I are going to have a drink tonight, AI bubble is popping, let's go.
MThat's the mood on one side of the internet.
CThe other side is still shipping. Reportedly, Google, OpenAI, Anthropic, Apple are all pushing new tech this week.
MSo the bubble debate is alive and well. We'll keep watching.