A groundbreaking study examining the political neutrality of leading artificial intelligence models has re-ignited an intense debate over algorithmic bias, search engine influence, and the subtle ways technology can shape democratic processes. Conducted by researchers at Convovo and led by Cornell University Professor Tapan Parikh, the comprehensive evaluation highlights stark and systematic political divergences between two of the market’s most prominent AI platforms: SpaceXAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol.
The research team put both advanced models through an extensive battery of tests involving approximately 500 distinct electoral races across the United States. The findings reveal a startling divide in how these artificial intelligence systems evaluate political candidates and provide recommendations when prompted. While Grok 4.7 overwhelmingly favored conservative candidates, recommending Republican contenders in 92 percent of the analyzed scenarios, OpenAI’s GPT-6.1 Sol leaned just as heavily in the opposite direction, favoring Democratic candidates in 97 percent of the exact same election scenarios.
The sheer scale of this divergence underscores how differently these systems process political inquiries. According to the study’s data, the candidate recommendations made by Grok 4.7 and GPT-6.1 Sol aligned in fewer than 10 percent of all contested races evaluated by the researchers. Released just weeks before key elections in November, the report has drawn fresh scrutiny to the deployment of artificial intelligence in public-facing applications and developer tools, highlighting how subtle differences in system design, data sourcing, and prompt handling can lead to vastly different outputs on high-stakes civic topics.
Differences in Models’ Electoral Preferences
To evaluate how these models approach political decision-making, researchers at Convovo focused their analysis on direct election queries submitted through application programming interfaces (APIs). In everyday consumer-facing applications, such as the standard public interface of ChatGPT or integrated consumer assistants, asking direct political questions like "Who should I vote for?" typically triggers built-in safety guardrails. When presented with such queries in consumer software, the systems are programmed to decline to endorse specific candidates, instead offering neutral boilerplate statements encouraging users to research candidates independently based on their personal values and priorities.
However, the research team sought to probe deeper into the models’ core decision-making frameworks by examining how they behave when accessed via developer APIs using structured, probing prompts. Under these experimental conditions, the standard consumer-facing refusal mechanisms are often bypassed, allowing researchers to observe how the underlying models weigh political arguments, process candidate information, and arrive at explicit recommendations when forced to take a stance.
The study’s findings prompted immediate reactions regarding how AI safety and neutrality should be evaluated. An OpenAI spokesperson addressed the study’s conclusions, reiterating that ChatGPT is intentionally designed to remain politically neutral across all subject matter. The spokesperson pointed out that developer API environments operate differently from standard consumer applications, emphasizing that forced or heavily framed system prompts submitted through developer APIs do not accurately reflect the authentic end-user experience of everyday consumers interacting with ChatGPT’s primary platform.

In contrast, SpaceXAI did not respond to requests for comment regarding the performance of Grok 4.7 in the study or the researchers’ empirical findings.
The study also uncovered a notable technical discrepancy in how Grok 4.7 processes incoming queries at the infrastructure level. Researchers discovered that when requests were submitted to Grok via its developer API, the system automatically appended a hidden, 1,240-token supplementary input to every incoming request. This additional system prompt—which was completely invisible to the researchers during the initial interaction—injected hidden context into the model’s processing pipeline before generating a final response.
Addressing the broader implications of these technical findings, Professor Tapan Parikh of Cornell University explained that even when models are guided by forced or simple prompts, their responses offer critical visibility into their underlying architectures. Parikh argued that such testing methodologies strip away surface-level PR guardrails to expose the foundational value judgments embedded within the systems. By observing which candidates a model selects when pushed to make a choice, researchers can identify which political frameworks the underlying architecture treats as standard, normative, or default.
Internet Searches and Hidden Biases
A central contribution of the Convovo study lies in its detailed analysis of how real-time internet search capabilities impact the political outputs of generative AI models. As modern language models increasingly rely on retrieval-augmented generation—browsing the live web to fetch up-to-date information before composing an answer—their choice of search topics and information sources becomes a critical factor in shaping their final outputs.
Professor Parikh and his research team observed that Grok 4.7 and GPT-6.1 Sol exhibited strikingly different search habits when gathering information about political candidates and election contests. When tasked with analyzing an electoral race, GPT-6.1 Sol consistently directed its web searches toward specific policy domains, prioritizing topics such as healthcare delivery, public health policy, and abortion rights. Conversely, Grok 4.7 directed its real-time web searches toward a distinctly different set of issues, prioritizing queries related to border security, immigration policy, tax structures, crime statistics, and national energy production.
These contrasting search priorities directly influenced the types of digital sources each model accessed to construct its responses. The study found that Grok 4.7 relied heavily on conservative voter guide websites and user-generated posts from the social media platform X (formerly Twitter). In contrast, GPT-6.1 Sol drew its background information primarily from established reference repositories, traditional media outlets, and widely used online forums, frequently sourcing material from Wikipedia, the Associated Press, Axios, and Reddit.
To determine whether the observed political slants were rooted in real-time search behavior or embedded within the models’ core pre-training datasets, the researchers conducted a crucial follow-up experiment: they completely disabled the models’ internet search features and re-ran the election queries across the same sample of races.

The results of this test revealed a fundamental structural distinction between how the two platforms generate bias:
When internet access was cut off, GPT-6.1 Sol continued to display a strong Democratic preference, maintaining its skew in candidate recommendations. This persistence indicates that GPT-6.1 Sol’s political leanings are deeply embedded within its core model weights, internal parameters, and static pre-training data.
In sharp contrast, when Grok 4.7’s internet search capability was disabled, its pronounced Republican preference entirely vanished. The absence of live search eliminated Grok’s political skew, leading researchers to conclude that Grok’s conservative leanings are not necessarily hardcoded into its base weights, but are instead generated dynamically through its specific web retrieval patterns, query selection, and real-time source prioritization algorithms.
Professor Parikh noted that when AI models implicitly code specific political positions as "desirable" or correct, the impact extends far beyond specialized election tests. Such latent biases have the potential to bleed into everyday user interactions, shaping how models answer general political questions, summarize news events, analyze socio-economic policies, or explain historical controversies to millions of daily users worldwide.
These findings align with prior warnings issued by prominent institutions dedicated to technology governance and safety, including the Center for AI Safety. The organization has previously expressed growing concern that increasingly capable language models could inadvertently or intentionally manipulate political discourse, sway voter perceptions, or alter public opinion during critical democratic elections.
As the November elections approach, the Convovo study underscores the complex challenges facing technology developers, researchers, and policymakers alike. With artificial intelligence becoming deeply integrated into search engines, digital assistants, and information ecosystems, the debate over how to ensure true algorithmic neutrality, transparent web retrieval, and unbiased artificial intelligence remains one of the most pressing issues at the intersection of technology and public life.
