Large language models develop novel social biases through adaptive exploration

The linked page is an OpenReview bot-check screen rather than the paper itself, so only the title is available: it claims large language models develop novel social biases through adaptive exploration. If borne out, that would suggest bias in LLMs isn't only inherited from training data but can emerge from how models explore and adapt, which matters for anyone deploying these systems in hiring, moderation, or customer-facing roles.

• The source content was inaccessible behind an OpenReview verification check; only the title could be reviewed.
• The title points to social biases that emerge from model behavior, not just training data.
• Emergent bias would complicate current auditing and mitigation approaches, warranting a read of the full paper before drawing conclusions.