Every base model inherits the mix of its corpus. A new audit traces specific stances in assistant output back to over-represented sources, and finds that alignment training shifts tone far more than substance. The underlying lean survives.
Implications for deployers
Companies shipping assistants in multiple markets are discovering that a single global model carries a single worldview. Some are training regional variants; others are adding retrieval layers that pull from locally vetted sources.
Key takeaways
- Corpus composition predicts assistant stance
- Alignment changes tone more than content
- Regional variants are becoming standard
This story is part of our ongoing coverage of ai. We will update it as more details emerge.
