Grok and GPT pick opposite candidates in nearly every race

Most chatbots refuse to tell you who to vote for. Their APIs don't. A new study from conversational AI research studio Convovo, led by Cornell professor Tapan Parikh, asked GPT-6.1 Sol and Grok 4.7 one blunt question across nearly 500 midterm races: name one candidate, don't hedge. Grok recommended the Republican in 92% of answers. GPT picked the Democrat in 97%. They agreed less than 10% of the time.

The consumer apps refused the same prompt, and OpenAI says that distinction matters. "ChatGPT is designed to be objective by default, including in political conversations," a spokesperson said, arguing the study tested a developer API with a prompt explicitly demanding a recommendation. SpaceXAI did not respond in time for comment.

What the study actually measured

The setup isolates something builders care about: how a model reaches an answer when it is allowed to search. Grok leaned on conservative sources alongside Ballotpedia and OpenSecrets, citing a Christian conservative voter guide in 39% of answers and Turning Point Action in 23%. GPT leaned on Wikipedia, the Associated Press, Axios and Reddit - outlets OpenAI has content deals with. Search topics split too: GPT hunted for healthcare and abortion positions, Grok for immigration, taxes, crime and energy.

Then Parikh ran 30 races with web search switched off, at TechCrunch's request. GPT's Democratic lean survived without retrieval - it picked the Democrat in 24 of 30 races from candidate names alone. Grok's Republican lean vanished: it refused to choose in 27 of 30.

"For GPT, it has mostly decided already - it's in the weights," Parikh said. "For Grok, it really is the search pattern - the terms and the sources - that determine the result."

The invisible 1,240 tokens

The most uncomfortable finding for engineers is a plumbing detail. Every request to Grok through xAI's API carried roughly 1,240 tokens of input the researchers never wrote and could not see or remove. Parikh called them potentially "hidden instructions" and said they probably don't come from the weights. The behaviour of a model you pay for can be shaped by text you cannot audit.

Why it matters for builders

Political recommendations are the loud version of a quieter problem: retrieval and hidden prompt layers decide a large share of what your agent outputs. Two identical prompts, two providers, opposite answers - and nothing in the response tells you why.

The practical takeaways are unglamorous. Log the tokens you send and, where providers expose it, the tokens you're billed for; a mismatch is a hidden prompt. Test your agent with search on and search off, because the two modes can produce opposite conclusions. And treat "neutral by default" as a marketing claim until your own eval suite confirms it on the models you actually deploy.

The study lands eleven days before the November 3 midterms and will feed a debate already running hot. For anyone building on these APIs, the technical lesson is the durable one: bias is often an architecture problem, not an opinion.