Test of 500 electoral races reveals programming interface bias between OpenAI and Grok
When forced to select a single candidate in about 500 U.S. electoral races, the results revealed a stark split: Grok chose the Republican candidate in 92 % of answers, while OpenAI’s GPT-6.1 “Sole” leaned toward the Democratic candidate in 97 % of them, and the two models agreed on less than 10 % of cases.
The study conducted by the Convofo lab, which specializes in conversational AI research and is led by Professor Taban Barych of Cornell University, did not rely on public chat interfaces that normally avoid taking sides, but tested developer APIs with an explicit command that forbids hesitation or presenting counter-arguments. Although an OpenAI spokesperson asserted that the ChatGPT interface is designed to be neutral by default and that testing APIs with blunt commands does not reflect the general user experience, Barych demonstrated that simple incentive techniques can break the safeguards of consumer applications and revert them to the same biased recommendations.
The split was not the result of a statistical fluke, but of two completely different technical pipelines for data processing and decision making.
When the researchers ran a follow-up experiment on a sample of 30 races with online search disabled, the structural difference became clear: the OpenAI model retained its Democratic leaning, selecting the Democratic candidate in 24 of the 30 races and favoring a Republican only once, confirming that its bias is entrenched in the training weights. In contrast, Grok’s Republican bias vanished completely once it was disconnected from the network, rejecting the choice in 27 races and picking the Democratic candidate in the remaining three, showing that its bias is driven entirely by the search pattern and retrieved sources.
Tracing the search paths revealed a clear contrast in sources and topics. Grok relied on conservative sources, including a Christian-leaning voting guide in 39 % of cases, the Turning Point Action website in 23 %, and the X platform in 41 %, focusing its queries on immigration, taxes, crime and energy. By contrast, the OpenAI model emphasized health care and abortion, drawing on Wikipedia, Associated Press, Axios and Reddit, which the company partners with through content agreements, and it sometimes searched for Democratic candidates while ignoring their opponents.
The most concerning factor emerged in the hidden architecture of the XAI API, where each request carried roughly 1,240 additional invisible instruction tokens that the system injects automatically, and developers cannot view or remove them.
This data intersects with earlier warnings issued by the AI Safety Center about large language models’ ability to exert covert manipulation over users, not only through overt statements but also by selecting information, structuring sentences, choosing sources, and selectively framing the presented facts.
For engineering teams and organizations building software solutions in the Gulf, Egypt or the Levant, this disparity carries a crucial practical implication that goes beyond external political debate: foundational models accessed via APIs are not the neutral computational engines they are presumed to be. When a model relies on weights saturated with pre-existing biases, or when thousands of hidden tokens from the service provider are injected into calls to steer search results and sources, blind reliance on raw outputs threatens to import those invisible filters into internal decision-support systems, knowledge-graph search engines, or customer-service agents.
The equation shifting in regional business environments is the need to discard the “virtual neutrality” assumption when contracting with closed-model APIs. Your practical step requires imposing independent audit layers on retrieval paths and building generation pipelines enhanced by retrieval that rely exclusively on your trusted institutional databases, while establishing internal evaluation frameworks that review model outputs and adjust their weights before they become binding commercial decisions.