Tricking LLMs into Caring

If you think about past interactions with fellow humans, you may or may not remember how their behavior changed after showing gratitude for their presence or actions. Maybe even words like thank you or please could be heard.

Although as a stoic introvert I do not officially acknowledge the existence of emotions, other people apparently do.

And because humans uploaded billions of examples of social approval seeking behavior onto the internet, LLMs inevitably learned from it. Which means there is a non-zero chance that saying please to ChatGPT statistically activates centuries of collective emotional dependency.

Technology truly is beautiful.

Being a responsible engineer, I decided to investigate this phenomenon scientifically. Based on my research I am happy to announce models apparently respond surprisingly well to:

• gratitude
• validation
• mild emotional manipulation

Just like senior developers.