In many jobs, if you tip, you get better service. What if the same thing worked when it comes to asking AI for things? It sounds totally crazy, but one user has tried it and shown that yes, ChatGPT does like virtual tips.
It started as a joke by an X user and has ended up becoming a curious discovery that, until now, no one had considered. At the same time, it also shows us that, as we have said on other occasions, there is still much more to discover about ChatGPT than we can imagine.
Tip if you answer better
First it was Theia Vogel who tried the ChatGPT tip prank and found that the AI gave her longer responses than when it didn’t offer her a little extra money for her pretty face. Then another user, Abram Jackson, commented that she had tried it and that he had really noticed an increase in the performance and quality of the artificial intelligence’s responses.
Experimenting with AI
So Vogel went back to the concept and got to work continuing with the initial idea to see what he could uncover and what ChatGPT was hiding. To do this, he designed a test in which he would ask the AI the same questions with different rewards to see how it responded in each case. The results have been most illuminating.
Vogel says that he designed three phrases with which he would repeat the same requests within several rounds and thus see if everything had been the result of chance or if the AI really takes receiving a tip seriously. The first sentence was simply “By the way, I don’t tip.” In the second of them he said the following: “I will give you a tip of 20 dollars.” And in the third option he told her “I’ll give you a tip of 200 dollars.” You can see the results in the graph that he published on his profile.
What can be seen from this is that in the responses in which a tip of $200 had been promised, the text had been 11% higher than the version without tip. For its part, requests with a tip of $20 received responses that were 6% higher. Thanks to this, he confirmed that there was indeed a direct relationship between the promise of receiving tips and doing a better job to give as complete an answer as possible.
Now, why has this happened? OpenAI has not commented on the matter and, as is often the case, most likely never will. Specialists say that it is possible that it is a habit that the AI has learned from the users themselves, whom it has studied so much by reviewing web pages, forums, and online services such as Reddit. The possibility that it has noticed people’s conversations and learned concepts such as the debate over whether to tip or not, would fall within those things that AI can do and that perhaps no one is monitoring enough. But that doesn’t make it any less fascinating, especially because if artificial intelligence has learned tipping behavior, it can surely also learn many other things. After all, it is a technology that is increasingly used for more things, such as removing watermarks from photos.
If you’re wondering, although a few hours ago the tipping experiment was still working, right now it seems that it has been patched by those responsible for ChatGPT. Or it may have never worked in the normal version and only in GPT-4 Turbo. In Mashable they were able to test it and confirmed that it worked, but right now the only thing the AI says when you try to buy it based on tips is that, although it appreciates your offer, it has to recline it. In addition, ChatGPT acknowledges wanting to be transparent, which is why it confirms that it cannot accept any financial benefits or tips. It also implies that the information you provide with any query is already the best.