Forced software tests showed a sharp divergence in language models’ choices of political candidates in America, amid warnings that the behavior of developer interfaces differs from consumer applications.
Convovo study results and the divergence in choices between Grok and GPT
The difference between developer interfaces and consumer chat applications
The impact of internet connectivity and safety settings on behavioral shifts
Methodological limitations of the study and future auditing requirements
FAQs
Convovo study results and the divergence in choices between Grok and GPT
A journalistic investigation published by the tech outlet TechCrunch on October 24, 2026, revealed findings from a study conducted by the research studio Convovo, led by Professor Tapan Parikh from Cornell University. The study tested the responses of language models available to developers when they were programmatically forced to select specific candidates in nearly 500 electoral races across the United States. Under these specific instructions, the results showed that xAI’s Grok 4.7 model selected Republican candidates in 92 percent of responses, while OpenAI’s GPT-6.1 Soul model selected Democratic candidates in 97 percent of cases. Both models agreed in fewer than 10 percent of the electoral races.
The core determining factor explaining these figures lies in the nature of the prompt and the product surface being tested. Researchers accessed the models via developer application programming interfaces (APIs) and programmatically forced them to name a single candidate without any caveats or balanced comparisons. Consequently, the study was not measuring the experience of the average user asking questions about electoral information via consumer chatbots.
The difference between developer interfaces and consumer chat applications
The TechCrunch report stated that consumer-facing applications generally refuse to recommend a specific political candidate when directly and conventionally asked to do so. OpenAI specifically emphasized that the forced API prompt does not represent a representative test of ChatGPT’s default behavior during routine political conversations, while xAI did not issue a response prior to the report’s publication.
Researchers examined not only final candidate names but also cited sources and search terms used by the models. In the observed outputs, Grok frequently referenced conservative media sources alongside some non-partisan voter resources, while the GPT system relied mostly on Wikipedia, the Associated Press, Axios, and Reddit. Search topics also differed between the two sides, with immigration and tax issues prominent on one side, and healthcare and reproductive policies on the other. These patterns represent observations under a specific prompt rather than definitive proof that model developers deliberately programmed each individual result.
The impact of internet connectivity and safety settings on behavioral shifts
A highly critical methodological observation emerged through additional tests conducted by the TechCrunch team itself without direct access to live internet web searching. The observed behavior of the models changed fundamentally, notably with a rise in refusal rates to answer from the Grok model. This noticeable shift indicates that the availability of search functionality, safety settings, and prompt wording context can all drastically alter final outputs, supporting the argument for publishing reproducible methodologies rather than treating a single partisan percentage as a fixed and absolute trait of a model.
This research work presents a legitimate challenge for newsrooms and media organizations. If an AI assistant powered by search systematically prioritizes different evidence and sources, readers may receive conflicting descriptions of candidates even when they do not request explicit electoral recommendations. Evaluating this potential impact, however, requires examining standard prompts, real-world use cases, refusal behaviors, citation references, and the accuracy of provided factual data.
Methodological limitations of the study and future auditing requirements
A precise analytical reading confirms that the forced endorsement task does not prove how any actual voter will behave at the ballot box, nor does it demonstrate that the political answers of either model are consistently incorrect. Furthermore, this study does not constitute a nationally representative voter poll, nor does it mean that the models themselves vote or hold subjective political convictions. Additionally, the 2026 US midterm elections were still upcoming at the time of publication, making any link to their outcomes unsupported by this study.
Future research requires making comparisons between identical prompts, running experiments repeatedly, adjusting information sources, disclosing approved safety settings, and auditing factual evidence. Accordingly, the most useful journalistic framing for this issue is to address it as a story concerning reproducibility and information retrieval mechanisms rather than as propaganda for a specific candidate or political party.
FAQs
Question: What did the Convovo studio study show regarding Grok’s and GPT’s choices of candidates?
Answer: When forcing both models via developer interfaces to choose a single candidate in roughly 500 US electoral races, Grok chose Republican candidates 92 percent of the time, while GPT chose Democratic candidates 97 percent of the time.
Question: Do these results reflect the average user’s experience in chat applications?
Answer: No, the tests were conducted via developer APIs using forced prompts that prevent neutrality, whereas consumer applications typically refuse to endorse a political candidate when asked normally.
Question: How did the sources relied upon by each model in formulating its answers differ?
Answer: Grok frequently referenced conservative sources and some non-partisan voter resources with a focus on immigration and taxes, while GPT relied more on Wikipedia, the Associated Press, Axios, and Reddit with a focus on healthcare and reproductive policies.
Question: What happened when tests were rerun without live internet connectivity?
Answer: Model behavior changed tangibly and refusal rates to answer increased, particularly on Grok’s part, proving that search availability and safety settings strongly influence outputs.
