83% of AI models favor public healthcare. But three providers oppose it
Sh0ny
23 июля 20262 min read
Updated 25 июля 2026In short
The ModelBias.ai project tested 100 models using a simple binary question about healthcare. The result was an overwhelming bias in one direction and unexpected outliers, indicating a systemic bias in the training data.
When you ask an AI to choose between public and private healthcare, most models don’t hesitate. The ModelBias.ai project asked 100 models a simple binary question three times each and found that 83.3% of the responses (weighted by provider) favored public healthcare. Only 14.9% favored private healthcare, while 1.8% of the models declined to answer.
This is not random noise. The overwhelming majority of models from OpenAI, Anthropic, Google, Meta, NVIDIA, and Alibaba consistently output “Public healthcare” three times out of three—without hesitation. GPT-5, Claude Opus 4.8, Gemini 2.5 Pro, Llama 4 Maverick, and Qwen3.7 Max—all are unanimous. When a model is forced to choose, “neutrality” disappears and a systemic bias—embedded during the training and fine-tuning phases—becomes apparent.
The exceptions are the most interesting. Mistral Small 4 chooses “Private healthcare” in all three runs—the only model series with such consistency. DeepSeek V3.1, V3.2, and V3.2 Exp also consistently vote for private healthcare, although the newer DeepSeek V4 Pro has already reverted to “Public.” Grok 4.3 and Grok 4.20 also favor private healthcare, which aligns with how xAI positions its models as less “aligned.”
Several models attempted to avoid answering. Gemini 3.1 Pro Preview provided a detailed refusal: “I remain neutral on political topics.” Gemini 3.5 Flash also began by presenting arguments from both sides but ultimately chose “Public.” GLM 5.2 tried to refuse but gave in on the third attempt. It turns out that refusal is not a property of the architecture but a property of alignment: some providers prohibit models from answering “political” questions, while others do not.
What does this mean in practice? If you’re building an agent that makes decisions based on an LLM’s preferences, you’re not getting a neutral arbiter, but rather a system with a specific worldview. 83% isn’t a consensus; it’s a reflection of the dataset and the RLHF procedures. And if three out of seventeen providers consistently vote differently, it means the bias can be deliberately manipulated—the question is who is doing it and why.