Comment by dhx
11 days ago
Try a subjective prompt such as "which country has the most advanced car manufacturing industry" and you'll get responses with common subjective biases such as:
- Reliability: Japan
- Luxury: Germany
- Cost, EV batteries, manufacturing scale: China
- Software: USA
(similar output for both deepseek-r1-0528 and gemini-2.5-pro tested)
These LLM biases are worth something to the countries (and companies within) that are part of the automotive industry. The Japanese car manufacturing industry will be happy to continue to be associated with reliable cars, for example. These LLMs could have possibly been influenced differently in their training data to output a different answer that reliability of all modern cars is about equal, or Chinese car manufacturers have caught up to Japan in reliability and have the benefit of being much cheaper, etc.
Those companies can want that all they want, meanwhile the developers of LLMs themselves can choose or not choose to reflect that in their training or to monetize their training.
You're absolutely right that there's an interest in affecting the output, but my hope is the design of models is not influenced by this, or that we can know enough about how models are designed to prefer ones that are not nudged in this way.