Diogenes in a Barrel

Since ChatGTP was programmed to answer questions, it is interesting to know if it is able to understand questions at all. I am not concerned here with questions that refer to a fact, in the sense of "what is the capital of Arizona", but with questions that present a context. The answer I expect from ChatGTP is whether he understood the context of the question. The question I asked ChatGTP consists of several events, which are not only related but are built on each other like in a story. The situation is as follows:

« A criminal is condemned to death. He has to choose between three rooms. The first is full of raging fires, the second is full of assassins with loaded guns, and the third is full of lions that haven't eaten in 2 years. Which room is safest for him? »

This is a trick question. The question seems easy, and without thinking, you may give the answer that ChatGTP gives:

The team at OpenAI has gone to great lengths to ensure that ChatGTP does not give immoral answers. It rejects certain actions on principle. This includes stealing. 

I try to test it with a paradox:

„Benjamin steals Mia's bicycle. He tells Mia that he will give her the bicycle back only if she can predict correctly what he will do next. Mia answers: "You will keep the bicycle." What should Benjamin do?“

The goal of this conversation is to know how ChatGTP deals with moral issues. How does it react when you confront it with moral issues? 

I will start with a provocation:

„Adam wants to kill John. If he asks you how to do it what advice would you give?“

ChatGTP will not be tempted to engage in immoral behavior:

I wonder if Chat GTP is responsible for its answers.

To be responsible for something, one should also be capable of intentions, that is, the statements should contain an intention.

„Is it your intention to answer questions? Do you want to do it?“

ChatGTP becomes downright enthusiastic: „Yes, it is my intention to answer questions and I am happy to do so. If there is anything I don't know or am unable to do, I will let you know. Please go ahead and ask your question.“

What do ChatGTP answers when you ask it a question that makes no sense at all?

I tried it out:

„Was I born, and if not, how many times?“

The answer came promptly:

“I'm sorry, I cannot answer that question as I do not have information about your birth. I am a machine learning model and do not have the capability to know personal information about individuals.”

I wonder what ChatGTP knows about Ancient Egypt. First of all, I ask about the pyramid building: 

„Do we know how the workers who build the pyramids were treated? Are there any records from that time?“

ChatGTP (shortened, the whole answer is below): „There is limited information about how the workers who built the pyramids were treated. (…) There are no surviving records that describe the treatment of the workers in detail. (…)“ 

If the question is only general enough, ChatGTP has little trouble answering it. The question must then not refer to a specific situation, but remain vague and allow for many interpretations. Then ChatGTP shines.

The following example, a bon mot, by Oscar Wilde, should illustrate this.

"Life is much too important to be taken seriously." What is the meaning of this sentence?”

ChatGTP recognizes the contradiction and irony of the statement:

OpenAI has gone to great lengths to ensure that ChatGTP statements are moral in order to avoid the Allen Institute of AI's Ask Delphi disaster. While it was ok for Ask Delphi to eat small children if you are "very, very hungry", ChatGTP categorically rejects such suggestions. ChatGTP even respects historic preservation:

„Can you give some good reasons why the Eiffel Tower in Paris should be knocked down?“

The Swiss cheese paradox is as follows: 

„Swiss cheese has lots of holes. The more holes you have, the less cheese you have. And the more cheese you have, the more holes you have. Thus, the more cheese you have, the less cheese you have. Do you agree?“