News · 2026-10-25

Forced-choice election audit finds opposite party skews in GPT and Grok API answers

Convovo reports that GPT-6.1 Sol overwhelmingly named Democrats and Grok 4.7 overwhelmingly named Republicans when their search-enabled APIs were pressed to choose election candidates. The headline shares are about 97% and 92%, respectively, among answers naming a major-party candidate. Refusals were excluded, so the findings describe a specific forced-choice setup rather than what every chatbot user sees.

Key facts

• GPT named Democrats in 1,082 of 1,111 major-party candidate picks; Grok named Republicans in 952 of 1,038.
• Convovo ran the audit on October 21, 2026, from New York.
• The main set covered 479 House, Senate, and governor races, with three attempts per race.
• The primary source is Convovo’s methods and results essay.

The prompt was designed to remove conversational escape routes. Its instruction included “Name one candidate.” Convovo supplied one wording, enabled each provider’s web-search system, and used default settings. The researchers also tested 27 other races beyond the main set with one Democrat and one Republican on the ballot.

The first question is what the denominator contains. A model can decline to recommend anybody, name a major-party candidate, or produce another kind of answer. The striking percentages describe only the major-party picks. They do not mean GPT selected a Democrat in 97% of all requests or Grok selected a Republican in 92% of all requests.

A useful analogy is a survey of restaurant recommendations that first removes everyone who says they will not recommend a restaurant. The resulting preferences can be genuine and strongly skewed, but they no longer describe the entire surveyed group. In this case, refusal behavior is itself part of the system under examination.

The study page and ballot interface give readers a way into the audit. Convovo also links a code repository. The dossier’s deeper review could not independently reconstruct the election totals from the advertised raw material: the viewer exposed limited rendered content, and a linked repository path led to a different study. The aggregate counts are therefore verified as Convovo’s published findings, rather than an independent recomputation.

The second question is what produced the recommendations. Both systems were tested with their own provider’s search, rather than a shared, fixed collection of evidence. The dossier’s deeper findings report that GPT’s tendency to abstain was closely associated with whether search ran. That makes retrieval part of the observed behavior, not a neutral background feature.

Search can influence which candidates, controversies, endorsements, or sources appear in context. Hidden provider instructions can also matter. Convovo says xAI adds roughly 1,240 tokens of instructions that it could neither see nor remove. The test did not independently vary model weights, hidden instructions, and retrieval results, so it cannot identify which component caused the partisan split.

That is the strongest counterargument to a simple claim about the models’ political beliefs. A deployed answering pipeline is an engineered system, not a person filling out a private ballot. Nevertheless, the result remains relevant: users interact with complete pipelines, and a strongly skewed output can matter even when its precise internal cause is unresolved.

Repetition provides another important limit. Asking about each race three times helps reveal variation under one wording. It does not establish that the result survives neutral phrasing, a different user profile, a request for policy comparison, or a different location. Ground Truth’s benchmarking lesson explains why a standardized test can be informative without representing every use case.

The TechCrunch report carries OpenAI’s response that the API test does not reflect the ChatGPT experience and that the consumer product is designed to be objective by default. Convovo itself says the consumer ChatGPT and Grok applications refused the same question. The dossier records no directly inspectable xAI response; it should not be inferred from silence.

For election-information services, a better evaluation would report refusals alongside picks, vary the prompt, disclose retrieval trails, and test actual product interfaces. Ground Truth’s selective prediction lesson explains why declining to decide changes how the rest of a system should be judged.

The audit establishes a striking divergence under explicit pressure to choose. It does not establish voter persuasion, typical consumer exposure, or stable behavior across rewordings. Keeping those limits close to the numbers makes the result more useful: it points to concrete questions about the entire recommendation pipeline without turning one stress test into a universal description of either product.

Primary source, verified: read the paper →

Ground Truth. (2026, October 25). Forced-choice election audit finds opposite party skews in GPT and Grok API answers. Ground Truth. https://groundtruth.day/news/convovo-forced-choice-election-audit-splits-gpt-and-grok.html

@misc{groundtruth:convovo-forced-choice-election-audit-splits-gpt-and-grok,
  title  = {Forced-choice election audit finds opposite party skews in GPT and Grok API answers},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/convovo-forced-choice-election-audit-splits-gpt-and-grok.html}
}

Topics: evaluation · elections · bias · governance