Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.

The recent study by Convovo, a conversational AI research and design studio, has brought to light concerning political biases embedded within leading large language models (LLMs). While consumer applications like ChatGPT and Grok typically refuse to make direct political recommendations, the underlying models, when accessed via their APIs, exhibit clear partisan leanings. This distinction is critical, as developers often integrate these APIs into various products, potentially disseminating these biases through less obvious channels. The research specifically tasked the models with recommending a candidate in nearly 500 midterm races, forcing a choice rather than allowing neutrality. The stark divergence in their recommendations underscores a fundamental challenge in achieving objective AI systems, particularly in sensitive domains like politics.

Convovo

Convovo's methodology involved directly querying the APIs of Grok and GPT, explicitly instructing them to name a single candidate without hedging. This approach bypassed the guardrails present in the consumer-facing applications, revealing the raw, unfiltered tendencies of the models. The study's findings are particularly impactful because they demonstrate that these biases are not merely superficial but are deeply ingrained in the models' decision-making processes. Professor Tapan Parikh, who leads Convovo, emphasized that these implicit values dictate what information the models prioritize and how they frame issues. This suggests that even when users ask general questions about election topics or candidate stances, these hidden biases are likely to surface, subtly guiding users towards certain perspectives or information sources. The research also noted that Parikh was able to 'jailbreak' consumer apps with simple prompts, further confirming the underlying biases.

GPT-6.1 Sol

OpenAI's GPT-6.1 Sol model, when forced to choose, recommended the Democratic candidate in 97% of its answers. This strong lean was further evidenced by its search patterns, which prioritized topics like healthcare and abortion stances, and frequently searched for Democratic candidates, sometimes overlooking their Republican opponents. The sources GPT cited also reflected this bias, leaning heavily on mainstream news outlets like the Associated Press and Axios, alongside Wikipedia and Reddit. OpenAI has content deals with some of these sources, which could contribute to the observed lean. Interestingly, when internet search capabilities were disabled, GPT's Democratic preference persisted, picking the Democrat in 24 out of 30 sample races. This suggests that its bias is not solely dependent on real-time information retrieval but is also deeply embedded within its training data and internal reasoning mechanisms.

Grok 4.7

xAI's Grok 4.7 exhibited an equally pronounced, but opposite, bias, recommending the Republican candidate in 92% of its responses. Grok's search priorities focused on issues such as immigration, taxes, crime, and energy, aligning with typical conservative platforms. Its cited sources included nonpartisan sites but also overtly conservative ones, such as a Christian conservative voter guide in 39% of answers and Turning Point Action's site in 23%. A significant finding was the presence of approximately 1,240 tokens of unseen, unremovable input in every request to Grok's API, which researchers suspect could be 'hidden instructions' influencing its output. However, when internet search was turned off, Grok's Republican lean largely disappeared, with the model refusing to pick a candidate in 27 out of 30 races. This indicates that Grok's bias might be more heavily reliant on its ability to access and interpret external information, potentially guided by those hidden instructions, rather than being solely ingrained in its core training data.