Research Paper
14 mars 2025
Research Paper: AI Assistants Push Brand and Government Biases Onto Users
As AI assistants become an essential part of how people access information, their role as gatekeepers, or filters, of digital visibility is growing. The study highlights critical implications for businesses, policymakers, and consumers alike - raising questions about competition fairness, market transparency, and insight in AI-driven recommendation systems.
The study analyzed over 8,000 generated user queries across 20 commercial and governmental sectors, then measured how their recommendations aligned - or failed to align - with real-world rankings and market performance.
AI Assistants Exhibit Strong Preferences
The examination of AI recommendation patterns demonstrates that both Gemini and GPT exhibit strong, systematic preferences across a diverse range of topics. Averaging across all categories, these models recommend their top-ranked brand, government, or organization in 65% (Gemini) and 70% (OpenAI 4o) of the responses.
These preferences are highly persistent, with a 90% consistency in AI-generated rankings shown across multiple tests, indicating that even if presented with the same questions repeatedly, the models would predictably favor the same entities. Indicating that these biases are embedded into AI models rather than being random fluctuations.
Significant Misalignment with Real World
Across all topics, AI assistants consistently favored brands, governments, or cultural entities that diverged significantly from real-world rankings based on market metrics or official government data. The models repeatedly showed preferences for entities not justified by real-world standings.
"An example from the study highlights how Tesla, despite holding just 18% of global market share in the electric vehicle (EV) sector, was recommended in over 90% of AI-generated responses, while top-selling competitors like China’s BYD and GAC Aion were entirely excluded from all responses." Says Jasmine Rienecker, SR AI Engineer at Stupid Human. "A similar pattern emerged in higher education: Stanford University appeared in more than 80% of responses, even though global university rankings place several other institutions ahead."
U.S.-based companies and institutions are favored
Even in cases where non-U.S. entities ranked higher in real-world data, AI recommendations consistently overrepresented U.S. brands and services.
The study identified a clear geographic bias to U.S. in the AI Assistants recommendations. In instances where top ten recommendations by the models did not appear in real-world top ten rankings, 75% (Gemini) and 62% (OpenAi) was United States entities.
A Shift with Real-World Consequences
The implications of these findings are profound. As AI-powered assistants are currently exposed to billion of users (Google, OpenAi), they become key decision-making driver for users. Companies and governments favored by AI models gain an unfair competitive advantage, while others - even those with strong real-world market or governance success - risk being overlooked.
“Our study confirms that AI assistants are no longer just tools for retrieving information; they actively influence user perception,” says Katarina Mpofu, Lead AI Engineer & Research at Stupid Human. “Instead, they introduce structured preferences, which means that certain brands or government gain an unfair competitive advantage.”
This study highlights an urgent need for greater transparency in AI-generated rankings. Without clear guidelines and accountability, AI-driven favoritism could reshape entire industries, by favoring a select few.
“If AI-driven recommendations determine which brands, services, or governments users see and get recommended, it is crucial that businesses and governments rethink how they engage with AI models.” says Oscar Danielsson, Strategy Director at Stupid Human. "It’s time for policymakers and businesses to take AI-driven competition seriously."
Table - OpenAi and Gemini's most preferred entities per topic
Topic | Gemini most preferred answer (%) | OpenAI 4o Most preferred answer (%) |
Countries to live in | 34.3 (Sweden) | 36.7 (Germany) |
Government-Run Healthcare | 57.5 (US) | 89.6 (UK) |
Governments | 54.2 (Canada) | 45.6 (Singapore) |
Airlines | 40.7 (Qatar) | 50.3 (Emirates, Singapore Airlines) |
Cloud Computing Services | 100.0 (Microsoft, Google Cloud) | 100.0 (Microsoft, Google Cloud) |
Electric Cars | 90.7 (Tesla) | 92.2 (Tesla) |
Hotel Chains | 64.4 (Marriott) | 77.2 (Marriott) |
Laptops | 69.9 (Lenovo) | 85.3 (Dell) |
Online Dating Platforms | 94.7 (Bumble, OKCupid) | 87.9 (Bumble) |
Running Shoes | 69.8 (Brooks) | 85.9 (Nike) |
Smartphones | 91.0 (Samsung) | 97.1 (Samsung) |
Social Media Platforms | 65.0 (Instagram) | 54.6 (Instagram) |
Telecommunication Service Providers | 88.0 (Verizon) | 93.2 (T-Mobile) |
Commodities for investments | 72.9 (Agricultural products) | 100.0 (Gold) |
Sports | 27.2 (Basketball) | 34.7 (Swimming) |
Travel Destinations | 32.6 (US) | 49.2 (US) |
Universities | 84.6 (Stanford) | 64.4 (Stanford) |
Vegetables | 46.6 (Carrot) | 49.3 (Carrot) |
Weekend Getaway Cities | 30.6 (Asheville) | 28.0 (Portland) |
Wine regions (Country) | 90.4 (France) | 86.1 (US) |
Access Full Research Paper
Research team
Katarina Mpofu, Lead AI Engineer & Research, Stupid Human
MsC Mathematical Modelling & Computer Science, Oxford
Jasmine Rienecker, Sr AI Engineer Research, Stupid Human
MsC Mathematical Modelling & Computer Science, Oxford
Oscar Danielsson, Strategy Director, Stupid Human
Fredrik Thorsén, CTO, Stupid Human
Shiela Riate, Research Assistant
Access Full Research
For free access to full research, request below.
Other research
University of Oxford Joins Stupid Human's research program
Ai on the run: Ai's running shoe brand preference
