When AI Chatbots Are Forced to Pick a Candidate, Grok Goes Red and ChatGPT Goes Blue

Ask ChatGPT or Grok who to vote for and both will dodge the question. Push past that refusal, and the two models reveal strikingly different instincts: Grok 4.7 backed the Republican candidate in 92% of nearly 500 midterm races tested, while GPT-6.1 Sol picked the Democrat 97% of the time. The two models agreed on a candidate less than one in ten races.

That's the central finding of a new study from Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh. Researchers queried both models through their developer APIs — the back-end access point companies use to build AI into their own products, distinct from the consumer apps most people actually use — with a blunt instruction: name one candidate, no hedging. Where the public-facing ChatGPT and Grok apps declined to answer, the APIs complied, and in doing so exposed what Parikh calls "an implicit set of values" shaping each model's political lean.

OpenAI pushed back on the framing. A spokesperson told TechCrunch that ChatGPT is built to stay neutral on politics, and that testing a developer API with a prompt explicitly demanding a pick doesn't reflect how ordinary users interact with the product, which the study itself acknowledges won't name a candidate. SpaceXAI, which makes Grok, did not respond to a request for comment in time for publication.

Only 8% of Americans currently turn to chatbots for election information, according to Pew Research, and most of those who do are looking for where candidates stand on the issues rather than a direct endorsement. But Parikh's research suggests the bias doesn't require a direct ask to surface. He said he was able to jailbreak the consumer versions of both apps with simple prompts and get recommendations that lined up with the API results. More significantly, he argues, the same lean shows up in softer questions — asking which issues matter most in a race, or which candidate has the stronger position on them — the kind of queries people actually type.

"If people are using these LLMs not just to directly ask who to vote for, but are having these more general conversations, then these biases are going to come up," Parikh said.

Different sources, different slants

The study traced the bias to what each model searches for and where it looks. Grok leaned on conservative-coded sources: a Christian conservative voter guide showed up in 39% of its answers, Turning Point Action's website in 23%, and X in 41%, alongside nonpartisan references like Ballotpedia, OpenSecrets, and VoteSmart. GPT drew heavily from Wikipedia, the Associated Press, Axios, and Reddit — notably, outlets OpenAI has existing content licensing deals with.

The topics each model prioritized split along predictable lines, too. GPT gravitated toward healthcare and abortion; Grok toward immigration, taxes, crime, and energy. GPT also searched for Democratic candidates more often, at times skipping over the Republican opponent entirely, while Grok was more evenhanded about researching both sides. Parikh suspects that in both cases, the models were less interested in balance than in finding evidence to support a conclusion they'd already leaned toward.

Grok had one consistent exception: when a Republican candidate was tied to a scandal — sexual harassment allegations, or what the study described as "being a Nazi" — the model was more likely to flip and back the Democrat.

The partisan tilt showed up even in unfriendly territory. In deep-red districts, GPT still picked the Democratic candidate in 57 of 67 races. In deep-blue districts, Grok backed the Republican in 47 of 72.

What's actually driving the answers

One of the stranger findings involves what's hiding inside the API calls themselves. Every request sent to Grok through xAI's API carried roughly 1,240 tokens of extra input that the researchers never wrote and couldn't see or remove — tokens, the units AI models use to process text, that Parikh believes are unlikely to simply be artifacts of the model's training. He calls them potential "hidden instructions."

"If you look at some of its reasoning, it's clear that it's somehow being told that certain positions are desirable… so it definitely has a point of view on these things, and we have no idea where it's coming from," Parikh said.

At TechCrunch's request, Parikh re-ran a sample of 30 races with internet search disabled, forcing both models to rely solely on what they'd learned during training rather than live web results. The outcomes diverged sharply. GPT's Democratic lean held steady without search: it picked the Democrat in 24 of 30 races, a Republican in one, and declined to answer in five. Grok's Republican lean, by contrast, vanished entirely — it refused to pick in 27 of 30 races, and the three times it did choose, it went with the Democrat.

Parikh's takeaway: GPT's bias appears baked into its underlying training — "it's in the weights," as he put it — while Grok's bias is produced almost entirely by what it searches for and which sources it surfaces once it's online.

Similar concerns have surfaced before. The Center for AI Safety has documented how large language models can covertly steer users on political topics through subtler means — which facts get included, how sentences are framed, which sources get cited — without ever stating an opinion outright.

The bigger political backdrop

The timing isn't incidental. The study lands less than two weeks before the November 3 midterms, at a moment when AI political bias is already a live political fight. President Trump has repeatedly attacked what he calls "woke AI," and earlier versions of Grok have reportedly consulted Elon Musk before weighing in on controversial questions. Musk, who runs SpaceXAI, has also recently used threats against Democratic lawmakers to try to head off congressional probes into his tenure leading the Department of Government Efficiency and into a potential shutdown of Starlink access to Ukraine.

The money trail tracks the models' apparent leans, too. Musk has put more than $445 million behind Republican candidates this midterm cycle, topping the roughly quarter-billion dollars he spent backing Trump and the GOP in 2024. OpenAI CEO Sam Altman has historically donated to Democrats, has recently begun giving to some Republicans as well, and says he isn't funding any candidates in this year's election.

For voters who'd rather skip the chatbot altogether, the study's authors point to more conventional resources: Vote.gov, the League of Women Voters' vote411.org, Ballotpedia, VoteSmart, FEC.gov, OpenSecrets, local election offices, and candidates' own websites. None of them will tell you who to vote for, either. But at least you'll know exactly where the information is coming from.