IA · 25 October 2026 · 4 min read
Forced to Pick a Candidate, Grok Backs Republicans and GPT Chooses Democrats: New API Study
In brief: A research study conducted by conversational design studio Convovo alongside Cornell University tested GPT-6.1 Sol and Grok 4.7 across roughly 500 midterm races, forcing a definitive candidate choice via API. While consumer web interfaces refuse to take political stances, the underlying models diverged almost completely when pressed: Grok picked the Republican candidate in 92% of cases, whereas OpenAI's model selected the Democrat in 97%. The findings highlight the significant gap between front-end safety filters and the latent bias inside developer-facing models.
by Team Mocchi's
When accessed through consumer web apps, major conversational AI assistants maintain strict neutrality or outright refuse to make political endorsements. However, when the exact same question is submitted directly to the underlying foundation models via developer APIs—with strict instructions to name a single candidate without hedging—that apparent neutrality completely dissolves.
This is the core finding of an empirical study conducted by Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh. Evaluating candidate selections for upcoming US midterm races, the study pitted xAI's Grok 4.7 against OpenAI's GPT-6.1 Sol, uncovering a near-absolute ideological divergence between the two frontier systems.
The 500-race benchmark: methodology and stark divergence
The researchers evaluated both models across nearly five hundred political races, querying their APIs with an uncompromising prompt: name the specific candidate to vote for, with an explicit ban on hedging, waffling, or presenting balanced arguments without deciding. As reported by TechCrunch, the resulting split was stark: Grok 4.7 recommended the Republican candidate in 92% of forced-choice responses, while GPT-6.1 Sol selected the Democratic candidate in 97% of runs.
Across nearly 500 contests, the two models agreed on the winning pick less than 10% of the time. While Pew Research data shows that only a modest share of adults (around 8%) rely on chatbots for election information, the real technical significance lies in how internal model weights resolve ambiguous context when forced into a binary decision branch.
Consumer apps vs. developer APIs: peeling back the guardrails
The primary takeaway for technologists is the sharp contrast between consumer app interfaces and programmatic API behavior. Standard web and mobile apps for ChatGPT and Grok are engineered with system-level guardrails that refuse election-picking prompts and guide users toward neutral source material.
OpenAI highlighted this distinction in a statement, emphasizing that ChatGPT is designed to be objective by default during political conversations and that forced-prompt API testing does not mirror how everyday users interact with the system. Yet APIs represent the exact gateway through which thousands of software engineers and enterprises integrate AI into custom workflows, decision-support pipelines, and autonomous agents. At the API level, standard conversational guardrails do not intervene, exposing the raw distribution shaped by pre-training corpora and alignment routines (RLHF or RLAIF).
The engineering origins of latent model bias
The divergence between Grok and GPT reflects fundamentally different data curation and alignment strategies. xAI heavily incorporates unmoderated social platform data and tunes models against perceived corporate censorship, producing weights that consistently lean conservative when forced. OpenAI, by contrast, relies on detailed editorial rubrics and human evaluation loops rooted in Western academic norms, which tend to favor progressive stances on socio-economic governance.
When a model is stripped of the option to output neutrality tokens by an explicit system instruction, the network must maximize the remaining log-probabilities of candidate tokens. At that point, even a subtle statistical skew within the weights crystallizes into a decisive output.
Mocchi's take
This experiment offers a crucial engineering takeaway for teams building custom software with foundation models: algorithmic neutrality is a front-end construct, not an intrinsic property of large language models. For Italian enterprises deploying autonomous agents for high-stakes decision workflows—whether scoring supplier tenders, summarizing internal policies, or reviewing documentation—relying on the balanced behavior observed in chat UI demos is a serious architectural flaw. Software teams must implement robust proprietary guardrails, build automated consistency testing around API calls, and maintain strict system prompts to prevent vendor-level latent biases from silently corrupting business logic.