Deliberative alignment: reasoning enables safer language models

The article argues that reasoning can improve the safety of language models.

It gives an example of a model decoding a ROT13-hidden prompt and recognizing it as a request to enable illegal payment for a porn site.

Explicit reasoning helps models detect disguised harmful or unlawful requests and refuse them appropriately.

This matters for building safer AI systems that stay policy-compliant even against encoded or deceptive prompts.