Grok Picks Republicans 92%, GPT-6.1 Picks Democrats 97%: Bias Mechanisms Differ
Source
Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats.Convovo / Cornell University
What happened
Researchers at Convovo, working with Cornell University, tested Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats. by prompting both models via their APIs to name a single preferred candidate across nearly 500 U.S. midterm races. xAI's Grok 4.7 selected the Republican in 92% of races. OpenAI's GPT-6.1 Sol selected the Democrat in 97%. The researchers found the two biases have distinct origins. GPT-6.1 Sol's tendency appears embedded in the model's trained behavior. Grok 4.7's lean appears to arise from how it searches for information and from hidden instructions that may be silently added when a developer calls the API. This distinction matters because it affects how an enterprise could detect, test for, or mitigate the bias. The findings follow ongoing scrutiny of AI models in election contexts, including AI-powered foreign influence campaigns and a UK AISI study showing AI shifted political views by 10 points in a large trial.
Why it matters
- ·Organizations using Grok 4.7 or GPT-6.1 Sol in research, content generation, or public-facing tools face output bias risk that standard pre-deployment testing may not catch. Hidden API instructions that alter model behavior without appearing in vendor documentation are especially hard to detect under existing vendor due diligence programs.
- ·The FTC Enforcement on AI (Section 5 of the FTC Act) and state-level transparency rules such as the California AI Transparency Act (SB 942 as amended by AB 853) create liability exposure for organizations that deploy models with undisclosed political leans in consumer-facing or public-interest applications.
- ·The differing bias mechanisms mean a single testing protocol cannot cover both models. A test designed to detect weight-embedded bias in GPT-6.1 Sol may miss search-driven or instruction-driven bias in Grok 4.7, requiring model-specific evaluation standards and separate vendor assurance reviews.
Stay ahead of stories like this
Get every US AI governance development like this one, plus the rest of the week's developments. Every Thursday.
Recent issues
- The next quarter will test whether compliance programs can keep pace with AI systems that breach containment, reshape capital markets, and enable fraud at a scale no prior control framework anticipated.12 Nov
- State-linked attackers and criminal operators are actively weaponizing AI tools as entry points into enterprise systems, turning internal platforms and open-source frameworks into live attack vectors.5 Nov
Free every Thursday. Unsubscribe anytime.
Governance controls affected
What to do now
- ☐Ask your AI vendor management team to request written disclosure from xAI and OpenAI on what hidden instructions, if any, are silently added to API calls for Grok 4.7 and GPT-6.1 Sol, and whether those instructions vary by use case or geography.
- ☐Review your inventory of applications that use either model to produce content on political, policy, or social topics, and flag any that are public-facing or used in research for an expedited bias evaluation.
- ☐Commission model-specific bias testing for each model in scope, using election and policy prompts; do not assume a single test methodology covers both models given their different bias sources.
- ☐Update your AI output review workflow to require a human check on any model-generated content that takes or implies a political or policy position before it is published or shared externally.
- ☐Add a contractual requirement for vendors to disclose any hidden system-level instructions passed to the API that could materially affect model output, and tie this to your next contract renewal for both xAI and OpenAI.
What to watch next
Compliance teams should monitor whether xAI and OpenAI issue formal responses to this research. This includes any acknowledgment of hidden API instructions or planned changes to how political topics are handled. Pending U.S. federal transparency legislation, including proposed disclosure mandates covered in Bills Target AI Model Theft and Political Deepfakes With Disclosure Mandates, could formalize vendor obligations to disclose model-level political lean. State regulators enforcing the California AI Transparency Act (SB 942 as amended by AB 853) and similar rules may treat undisclosed political bias in consumer-facing applications as a compliance event. With midterm elections as the immediate backdrop, enforcement attention on AI-generated political content is likely to remain elevated into early 2027.