NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 25·5 min read
It's the humans, not the data: Geopolitical bias in LLMs originates in post-training, amplified by the language of the prompt
TL;DR
Prior research assumed geopolitical bias in large language models (LLMs) came from internet training data. Researchers compared base models (pre-training only) against chat variants (pre-training plus human alignment) across seven AI labs using forced-choice country-pair scenarios in three languages.
✦ Why It Matters
Engineers must audit alignment procedures, not just training data, to prevent unintended geopolitical bias in deployed LLMs.
Key Takeaways
- Conduct bias audits during post-training phases of LLMs.
- Be aware of how prompt language can influence model responses.
- Consider geopolitical context when developing AI models.
- Advocate for transparency in AI alignment processes.
How It Works
The study utilized a paired-scenario forced-choice probe to assess biases across different LLMs. By comparing base models with their post-trained counterparts, researchers identified significant shifts in bias that correlated with the developers' geopolitical affiliations.
This indicates that the alignment processes during post-training are crucial in shaping how models respond to prompts.