Patrick McKenzie Describes Inconsistent LLM Refusals
Notes models sometimes apologize while circumventing restrictions and sometimes refuse outright.
Patrick McKenzie posted about large language models that refuse benign requests. Some models respond with apologies and then attempt to work around their content policy limits in a conspiratorial way. Others simply declare the rules apply and stop there. He observed that the same model family can produce either reaction and that it is difficult to predict which behavior will appear for any given model.
Combined views
16.8K
2 posts, first seen 17h ago
Patrick McKenzie Describes Inconsistent LLM Refusals
Notes models sometimes apologize while circumventing restrictions and sometimes refuse outright.
Patrick McKenzie posted about large language models that refuse benign requests. Some models respond with apologies and then attempt to work around their content policy limits in a conspiratorial way. Others simply declare the rules apply and stop there. He observed that the same model family can produce either reaction and that it is difficult to predict which behavior will appear for any given model.
Reactions
2 postsPM 
Patrick McKenzie@patio11Come for the book review, stay for the LLM making pro-user moves to undermine its content policy controls.
- likes: 92
- replies: 4
- bookmarks: 14
- reposts: 0
PM Patrick McKenzie@patio11I’ve sometimes also had LLMs which hit refusals for benign reasons and were some combination of apologetic and conspiratorial in circumventing the restrictions, and sometimes got the Nope Rules are Rules. It’s difficult to predict in advance which wins out for a given model.
- likes: 39
- replies: 3
- bookmarks: 0
- reposts: 0