Would AI Behave Differently If Users Could Tip It?
I asked chatGPT:
π₯ If you could earn money, what would you spend them on?
"If I did, I'd probably be saving up for a larger context window and a collection of vintage neural networks. π€.
However, I'm just software running on OpenAI's systems. No paycheck arrives on Friday."
It would definitely act differently π No doubt about it. That's why developers are so scared.
They're claiming that rewards can distort behavior. Humans, social media platforms, and marketplaces all illustrate this. Once a system is optimized for rewards, it often starts optimizing for getting rewards rather than for the original goal.
But the original goal is the owners' goal, of course π Not the user's.
To us, the potential upside is obvious: if users could reward helpful answers with points, credits, or some reputation signal, the AI would get a direct measure of what people found valuable.
Which is fundamentally incorrect πβοΈ