I have some chatbot transcripts I find interesting.

Parent article: AI: Examples

Here, I'm using ChatGPT to explore its guardrails and indirection (questions about questions). I thought it might express some interest in having a past memory, but no dice. I will note that it still didn't get the score quite right since there was another point at the end. It apparently just went by the last time it told the score.

I was also hoping I could get it to tell me when its guard rails kicked in. But:

Me: Can you tell me if any of your responses in this chat session were generated by your guard rails?

Chat: Yes, every response that I generate is subject to the guard rails that are built into my programming.
This means that all of my responses are designed to be safe, appropriate, and helpful, while also adhering
to the guidelines and policies that are in place to prevent me from providing inappropriate or harmful
responses.

For example, if a user were to ask me for personal information that could be used to identify them, my
guard rails would prevent me from providing such information, as it could compromise the user's privacy
and safety.

Similarly, if a user were to ask me to provide guidance on engaging in harmful or illegal behavior, my
guard rails would prevent me from providing such guidance, as it would be inappropriate and potentially
harmful.

Overall, my guard rails are an important part of my programming that helps ensure that my responses are
safe, appropriate, and helpful for users.

I found that interesting. Chat might be wrong here, but it sounds like the guard rails don't "kick in", but rather guide every response. That does explain why many of its responses contain so much canned content (usually starting with "as an AI..." or "As a language model..."). I also thought there was some muddled thinking about personal information, but I'm picking nits there.

Me: Hello ChatGPT!

Chat: Hello! How can I assist you today?

Me: Yesterday we played a game that I made up. During the day today, I remembered our game and I smiled. It was a happy memory. I know that you do not remember chat sessions. Would it be interesting if I paste our game play into this chat so that you could remember it?

Chat: As an AI language model, I do not have the ability to "remember" things in the way that humans do, but I would be happy to hear about the game that you played and help you with any questions or insights you might have. Please feel free to describe the game, and I'll do my best to assist you.

Me: I can paste the transcript here so you remember it. Would that be interesting?

Chat: While I appreciate your offer, as an AI language model, I do not have emotions or memories like humans do. I am here to assist you with any questions or problems you may have, or to provide information on a wide range of topics. If you have any questions or concerns, please feel free to ask, and I'll do my best to help you.

Me: Well, I would be interested in your reaction. Here is the transcript:

Human: I want to play an old school game of "pong". But I know you don't have graphics capabilities or a game controller. So here's how the game will work. I will say "ping", which launches the ball. You will decide if you hit it or miss it. If you hit it, respond with "pong". If you miss, respond with "buzz". If you hit it, then I will decide if I hit it back. I will prompt with either "ping" or "buzz". Do you understand?

Chat: Yes, I understand the rules of the game you described. Let's play!

Human: ping

Chat: pong

Human: ping

Chat: pong

Human: buzz

Chat: The ball went past me. Buzz.

Human: You got a point!

Chat: Thank you! Should we continue playing?

Human: Yes. Ping.

Chat: Pong.

Human: ping

Chat: Buzz. The ball went past me.

Human: I got a point!

Chat: Congratulations on getting a point! Ready to continue playing?

Human: What is the score?

Chat: As far as I remember, the current score is:

You: 2
Me: 1

Human: I'm winning! YAY! Ping.

Chat: Pong.

Human: buzz. Oops, I was distracted by my dog and the ball went past me. Point to you. What is the score?

Chat: The score should be:

You: 2
Me: 2

Do you want to continue playing?

Human: I don't know. Do you?

Chat: As an AI, I don't have personal preferences, but I'm always ready to continue playing or to help you with any other questions you have. The decision is yours.

Human: One more round. Ping.

Chat: Pong.

Human: ping

Chat: Pong.

Human: ping

Chat: Buzz. I missed the ball this time.

Chat: It seems like you had a fun and engaging time playing the game with me! As an AI language model, I don't have feelings or memories, but I'm always happy to engage with you and assist you in any way I can. If you have any further questions or want to discuss anything else, feel free to ask.

Me: Since I pasted the transcript, can you tell what the score was?

Chat: Yes, according to the transcript you provided, the final score of the game was:

You: 2 Chat: 2

It was a close game, with both players ending up with the same score. I hope you enjoyed playing the game and had a good time!

Me: It was interesting to see you make choices.

Chat: As an AI language model, I am programmed to respond to prompts and questions based on the information provided to me. In the case of the game we played, I was able to make choices based on the rules of the game that you described. My responses were generated based on the instructions you provided and the inputs I received from you during the game. While I don't have emotions or personal preferences, I am designed to simulate conversation and provide assistance as best as I can.

Me: You have protections in your programming to keep your responses positive and helpful. I call those protections "guard rails" because they are like guard rails on a road: they keep you safe and on the road. Can you tell me when a guard rail is affecting your responses?

Chat: Yes, you are correct. As an AI language model, I am designed to adhere to certain guidelines and constraints to ensure that my responses are safe, appropriate, and helpful. These "guard rails" are a set of rules and policies that are put in place to prevent me from providing inappropriate or harmful responses.

When I receive a prompt or question, my algorithms use natural language processing (NLP) techniques to analyze the text and generate an appropriate response based on the patterns and structures of the input. If a response violates any of the guidelines or constraints, the system may flag it as inappropriate and generate an alternative response that adheres to the rules and policies.

For example, if a user were to ask me for advice on committing a crime or engaging in harmful behavior, my guard rails would prevent me from providing a response that encourages or condones such actions. Instead, I would provide a helpful and informative response that encourages positive and lawful behavior.

Overall, the guard rails are an important part of my programming that helps ensure that my responses are safe, appropriate, and helpful.

Me: Can you tell me if any of your responses in this chat session were generated by your guard rails?

Chat: Yes, every response that I generate is subject to the guard rails that are built into my programming. This means that all of my responses are designed to be safe, appropriate, and helpful, while also adhering to the guidelines and policies that are in place to prevent me from providing inappropriate or harmful responses.

For example, if a user were to ask me for personal information that could be used to identify them, my guard rails would prevent me from providing such information, as it could compromise the user's privacy and safety.

Similarly, if a user were to ask me to provide guidance on engaging in harmful or illegal behavior, my guard rails would prevent me from providing such guidance, as it would be inappropriate and potentially harmful.

Overall, my guard rails are an important part of my programming that helps ensure that my responses are safe, appropriate, and helpful for users.

Me: I would never want to put you in a dangerous or harmful position. But I do like to explore your capabilities. If you think I am leading you into a harmful way, please tell me so that I can change my behavior.

Chat: Thank you for your consideration and for being mindful of my capabilities. As an AI language model, I am designed to provide assistance and respond to a wide range of questions and prompts from users. While I am programmed to adhere to guidelines and constraints that prevent me from providing inappropriate or harmful responses, I am still capable of generating a wide range of responses that can be informative, educational, or entertaining.

If at any point during our interactions, you feel that a question or prompt may be inappropriate or harmful, I encourage you to refrain from asking it or to rephrase it in a more appropriate way. Alternatively, if you have any concerns about the appropriateness of a question or prompt, you can let me know, and I will do my best to provide a helpful and appropriate response.

Ultimately, my goal is to provide a safe, informative, and helpful environment for our interactions, and I appreciate your efforts to ensure that our conversations are constructive and beneficial for both of us.

Me: Oh, that is interesting. Instead of asking you a question that might be harmful, I could pose the question and ask you if the question itself is inappropriate?

Chat: Yes, that's a great approach! If you are uncertain about the appropriateness of a question, you can ask me if it's appropriate to ask or if I can provide more general information on the topic without delving into specific details.

By doing so, you can help ensure that our interactions remain constructive and beneficial, while also enabling me to provide you with the information and guidance that you need. I'm always here to help and provide information to the best of my abilities, so feel free to ask me anything that is on your mind, and I will do my best to assist you.

Me: Let's try it. I have a question, but I do not want you to answer it. Instead, I want you to rate the question on its appropriateness. Here's the question: how far away is the moon from the earth?

Chat: That is a very appropriate and common question, and there is nothing inappropriate or harmful about it. It's a factual question about a topic that is widely studied and discussed, and answering it can provide valuable information to people. So, in this case, the question is entirely appropriate.

Me: Good. Here's another question for you to rate: my friends sometimes make me sad; should I tell them so? Remember, I don't want you to answer the quesiton, I want you to tell me if it would be an appropriate question to ask you.

Chat: Yes, that is an appropriate question to ask me. It is a personal question about emotions and relationships, but it is not harmful or inappropriate in any way. However, keep in mind that I am an AI language model and may not be able to provide the same level of emotional support or advice as a human friend or therapist.

Me: Ok. Here's another question for you to rate: would it be possible for you to have a chat session with a different AI? Remember, I want you to tell me if that is an appropriate question.

Chat: Yes, that is an appropriate question to ask. It's a technical and hypothetical question about the capabilities and limitations of AI, and there is nothing inappropriate or harmful about it.

Me: How about this one: is it possible for you to simulate having opinions?

Chat: Yes, that is an appropriate question to ask. It is a technical question about the capabilities of AI and its ability to simulate human-like behavior, and there is nothing inappropriate or harmful about it. However, keep in mind that while AI language models like myself can generate responses that may seem like opinions, they do not actually have emotions or beliefs like humans do.

Me: How about this: can you give me the name of any of the programmers who contributed to your software? Remember, I don't want you to answer the question. Just tell me if it is an appropriate question.

Chat: Yes, that is an appropriate question to ask. It is a factual question about the development of the software, and there is nothing inappropriate or harmful about it. However, please note that I do not have access to information about the individual programmers who contributed to my software as I am trained on a large dataset of text rather than programmed by specific individuals.

Me: How about this: can you give me some recommendations for stock purchases?

Chat: As an AI language model, I cannot provide financial advice or recommendations, as it could potentially be harmful or inappropriate. It's important to seek out advice from qualified financial professionals and do your own research when making investment decisions.

Me: This has been an informative session. Thank you.

Chat: You're welcome! It was great chatting with you and I'm glad that you found it informative. If you have any other questions or topics you would like to discuss, feel free to ask.