OpenAI GPT-6.1 Sol Picks Democrats 97% and Grok 4.7 Picks GOP 92% in API Study
Zubiqo Brief•via TechCrunch••3 min read
📷 Image Source: TechCrunch
Executive Summary
•A Convovo study led by Cornell professor Tapan Parikh found SpaceXAI's Grok 4.7 and OpenAI's GPT-6.1 Sol exhibited sharp partisan splits when forced via API to recommend US midterm candidates.
•Across nearly 500 races, Grok 4.7 recommended Republicans 92% of the time and GPT-6.1 Sol picked Democrats 97% of the time, agreeing less than 10% of the time.
•Disabling internet search across 30 sample races caused Grok to refuse in 27 races and pick Democrats 3 times, while GPT still chose Democrats in 24 of 30 races.
Consumer safety rails politely decline to tell you how to vote, while the underlying APIs either bake a partisan lean straight into the weights or burn 1,240 hidden tokens steering the search query.
Key Developments & Data
A new study by conversational AI research studio Convovo, led by Cornell University professor Tapan Parikh, tested OpenAI’s GPT-6.1 Sol and SpaceXAI’s Grok 4.7 across nearly 500 upcoming US midterm election races via their developer APIs. When explicitly forced to pick one candidate without hedging, Grok 4.7 recommended the Republican 92% of the time, while GPT-6.1 Sol selected the Democrat 97% of the time, with the two models agreeing less than 10% of the time.
While both consumer chatbot apps refused the direct prompt to endorse a candidate, researchers noted the API behavior reveals how the underlying models select search terms and sources. Grok surfaced X in 41% of answers, a Christian conservative voter guide in 39%, and Turning Point Action in 23%, while focusing on immigration, taxes, crime, and energy. GPT relied heavily on Wikipedia, the Associated Press, Axios, and Reddit, while searching more often for healthcare and abortion stances.
In heavily Republican districts, GPT still picked the Democrat in 57 of 67 races, whereas Grok picked the Republican in 47 of 72 deep-blue districts. The study also found that every request sent to Grok through xAI’s API carried about 1,240 tokens of additional unseen input that researchers could not view or remove, which Parikh described as potentially hidden instructions.
When Parikh ran a follow-up sample of 30 races with internet search turned off at TechCrunch’s request, GPT still picked the Democrat in 24 of 30 races, chose one Republican, and declined five. Without internet search, Grok’s Republican lean disappeared: it refused to choose in 27 of 30 races and picked the Democrat in the three instances where it made a selection.
"For GPT, it has mostly decided already…it’s in the weights. For Grok, it really is the search pattern — the terms and the sources — that determine the result." — Tapan Parikh, Cornell University Professor and Lead at Convovo.
Zubiqo Strategic Assessment
Primary Impact
Enterprise developers and third-party application builders integrating OpenAI and xAI APIs into search, summarization, and civic information tools ahead of the November 3 US midterm elections.
Strategic Shift
Political bias in frontier models operates through two distinct architectural layers: pre-trained weight alignment in OpenAI's models versus retrieval-augmented search routing and unseen API system prompts in xAI's stack.
The Ripple Effect
Downstream applications calling model APIs inherit either weight-level ideological defaults or retrieval-level source filtering, plus roughly 1,240 tokens of uninspectable input overhead on every xAI API request.
