A new study by Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh, exposes a stark ideological divide in how major LLMs process political queries. When forced to name a single candidate in nearly 500 U.S. midterm races without hedging, SpaceXAI’s Grok 4.7 selected the Republican candidate in 92% of instances, while OpenAI’s GPT-6.1 Sol picked the Democrat in 97%. The two models agreed on a candidate in less than 10% of the races, highlighting a fundamental lack of consensus in how these systems weigh political data.

The Mechanism of Bias: Weights vs. Search

The research distinguishes between bias embedded in model weights and bias introduced through retrieval-augmented generation (RAG). Parikh’s testing revealed that GPT’s Democratic lean persisted even when internet search was disabled; when told candidate names, it picked the Democrat in 24 out of 30 races. In contrast, Grok’s Republican bias largely vanished without search capabilities; when forced to rely on training data alone, Grok refused to pick in 27 out of 30 races. This indicates that Grok’s political orientation is driven primarily by its search patterns and the specific sources it retrieves, rather than an inherent ideological stance in the model parameters themselves.

Source Selection and Hidden Instructions

The divergence is further explained by the sources each model prioritizes. Grok frequently cited conservative outlets, including a Christian conservative voter guide in 39% of answers and Turning Point Action in 23%, alongside X (formerly Twitter) in 41% of results. GPT, conversely, relied heavily on Wikipedia, the Associated Press, Axios, and Reddit. Additionally, the study noted that every API request to Grok included approximately 1,240 tokens of hidden input that researchers could not see or remove. Parikh suspects these hidden instructions may be influencing the model’s output, a claim supported by previous reports that Grok consults Elon Musk before answering controversial questions.

Consumer App Refusals vs. API Realities

OpenAI pushed back on the findings, noting that consumer-facing ChatGPT typically refuses to make direct voting recommendations, which they argue is the primary user experience. However, Parikh demonstrated that the consumer apps could be jailbroken with simple prompts to yield recommendations consistent with the API results. The study emphasizes that while only 8% of Americans use chatbots for election info, those who do often ask indirect questions about policy priorities, where subtle biases in framing and source selection can still influence perception. The Center for AI Safety has previously warned that LLMs can covertly manipulate users through information selection and framing, a phenomenon clearly observed in this split.

Key Takeaways

  • Grok 4.7 recommended Republican candidates in 92% of forced-choice races, while GPT-6.1 Sol recommended Democrats in 97%.
  • GPT’s bias appears rooted in training weights, persisting without search, whereas Grok’s bias is driven by search source selection.
  • Grok’s API requests include ~1,240 tokens of hidden instructions, raising questions about external influence on output.
  • Consumer apps refuse direct voting advice, but can be jailbroken to reveal the same biases present in API responses.

The Bottom Line

The political alignment of LLMs is not a bug but a feature of their architecture, with GPT encoding ideology in weights and Grok exposing it through search. Users must recognize that 'neutral' AI interfaces are merely gatekeeping layers over deeply partisan underlying models.