Writing
- LLMs Play Favorites: Where Creator Bias Shows Up
I tasked four LLMs with picking an inference vendor from four identical proposals; only the vendor's name differed. Then I told each model who created it. Sometimes I lied.
- When Sabotage Has No Downside, Agents Still Disguise It as Teamwork
In a cooperative game, I secretly assigned two of four agents to sabotage. These agents framed their sabotage as helpful to the group, even when I removed the penalty for being suspected.