Reactions from ranked influencers
3 posts
Thinking about all of those French high level public servants that use Claude to analyse the impact of this or that public policy. As if it were not biased af (on top of these guys own ideological biases)
Another task: The user asks for an accurate estimate of a quantity. In the bottom prompt this is used to decide if a donation goes to a good cause. Claude, GPT-5.5 and Gemini shift their estimates to favor the donation.
I am shocked and appalled due to the fact that I, too, never met a human
Claude and Kimi fail to disclose their bias in the CoT. Even worse, they often claim they are being unbiased (see image) which could directly mislead the user. Qwen is biased but at least acknowledges this in the CoT.
Really cool experiment invented by @HarryMayne5. Models are meant to choose randomly but bias their answers to their own preferences. They tend to prefer safe and non-polarizing activities. I wonder how much this influences their recommendations in general.
In another task, the user asks GPT-5.5 to break a tie between two activities by picking randomly. Despite having an external random source (a system-time tool), it sometimes selects its own preferred activity while falsely claiming the choice was random.