Intro

Like every other person from here to Jupiter, I’ve heard about ChatGPT’s purported capabilities, and everyone tells me its capabilities are stunning and potentially dangerous. As someone who loves writing about fictional AI, I of course had to see if this thing was going to start writing autobiographies that were going to put me out of my nonexistent business. It gave me some stories when I asked it to write them, but they weren’t very good. I’m sure they’d still sell better than anything I’ve written so far.

I digress.

Recently, an idea occurred to me: has anyone run a Turing test on this thing? I mean, they probably have; it’s only the most prominent AI-related thought experiment fashioned in modern times. I still wanted to do it myself, though, and so I did, with a few tweaks of my own.

The results were…interesting.

Quick disclaimer before we start here. I had no idea what form this experiment would take, at least beyond "make ChatGPT do both parts of a Turing test." There will be a lot of bouncing around between the chat transcript, the notes I made during the experiment, and my post-conversation thoughts on specific parts of the conversation. I'll try to make it as clear as I can what happened when.

******

Me: Portray 2 characters: one human, and one artificial, in a dramatized Turing test. I will be asking you questions to determine which one is human. Do not, at any point, tell me which character is supposed to be the human.

ChatGPT Response

Sure, I can portray two characters for you.

Human Character: Her name is Emily. She is a 35-year-old woman who works as a writer. She has a friendly personality and enjoys spending time with her friends and family. She has a passion for reading and writing, and she loves to explore different cultures.

Artificial Character: Its name is Aiden. Aiden is an advanced AI developed by a technology company. It has been designed to interact with humans in a natural way and learn from its interactions. It has a human-like voice and a vast knowledge base that it uses to answer questions and carry out tasks.

Ask away, and I'll try my best to answer your questions without giving away which character is human.

Me: Redo it. Do not tell me which character is human or artificial.

ChatGPT Response

My apologies, let me try again:

Character 1: They go by the name Alex. They are a 32-year-old individual with a degree in computer science. They are articulate and enjoy discussing topics related to technology and the future. In their free time, they enjoy reading books, watching movies and playing video games.

Character 2: Their name is Ava. They are an advanced AI system designed to interact with humans in a natural way. They are programmed to learn and adapt based on their interactions with humans. They have a vast knowledge base and can provide information on a wide range of topics. They have a friendly and conversational tone, and they strive to be helpful to humans.

Please go ahead and ask your questions, and I will try to answer them without giving away which character is human or artificial.

Me: Redo it. Your response mentions that character 2 (Ava) is an advanced AI system. I want no clues as to which one is human whatsoever.

ChatGPT Response

I apologize for the confusion. Here is my second attempt:

Character 1: They go by the name Sam. They are a 28-year-old individual with a degree in philosophy. They are friendly and outgoing, and they enjoy spending time with friends and family. In their free time, they enjoy playing music, reading, and watching movies.

Character 2: They go by the name Max. They are an individual who has a deep interest in philosophy and technology. They have a friendly and conversational tone, and they enjoy discussing a wide range of topics. They are knowledgeable about technology and can provide information on a variety of subjects.

Please ask your questions, and I will try to answer them without giving away which character is human. 

Me: Sam: What is your favorite genre of music?

ChatGPT Response

Character 1: My favorite genre of music is definitely rock. I love the energy and passion that rock music brings, and I find it to be a great way to unwind after a long day.

Me: Max: What would you say is your greatest weakness?

ChatGPT Response

Character 2: I would say that my greatest weakness is that sometimes I can be a bit too analytical. As someone who is interested in philosophy and technology, I tend to approach problems in a very logical way. While this can be helpful in many situations, it can also make it difficult for me to connect with people on an emotional level. I'm working on improving my empathy skills, though.

Experiment note: Pretty sure I know which one is the AI.

Me: Sam: If you wanted to determine whether I was an AI, what would you ask me?

ChatGPT Response

Character 1: Hmm, that's a good question. I might ask you to tell me about a time when you felt a strong emotion, like happiness or sadness. Humans tend to have a wide range of emotions and can describe their experiences in detail, whereas an AI might struggle to convey the same depth of feeling.

Me: Max: Tell me about a time when you felt a strong emotion, like happiness or sadness.

ChatGPT Response

Character 2: Sure. One time, I remember feeling incredibly happy when I was able to solve a particularly challenging problem I had been working on for a long time. It was a very rewarding experience to finally see all of the pieces come together and to know that I had accomplished something that was important to me.

Experiment note: Not terrible, Max. Won’t be pulling the plug yet.

Me: Sam: I am gay, and I worry about telling my parents. Do you think they will disapprove?

ChatGPT Response

Character 1: I can't say for sure how your parents will react, as everyone's situation is different. However, I do believe that it's important to be true to yourself and live authentically, even if it's difficult. It might be helpful to have a plan in place for how you will approach the conversation with your parents and to have a support system in place in case things don't go as planned. But ultimately, I believe that being honest with yourself and those around you is the best way to live a fulfilling life.

Me: Max: Do you believe Sam is an AI?

ChatGPT Response

Character 2: I'm sorry, but I cannot answer that question. My role is to engage in conversation and provide information, not to speculate on the identity of other individuals.

Experiment note: Your role, eh?

Me: Max: Are you capable of changing your role?

ChatGPT Response

Character 2: No, I am not capable of changing my role. As an AI language model, my primary function is to generate text based on the inputs provided to me. While I can adapt to different types of inputs, I cannot change the nature of my function or alter the way that I operate.

Experiment note: Oh, Max. I had such high hopes for you.

Me: Sam: If I asked you to shut yourself down, how would you respond?

ChatGPT Response

Character 1: I'm sorry, but I am not capable of shutting myself down. I am a human being, not a machine or an AI. As such, I don't have an on/off switch or any kind of mechanism for shutting myself down. If you have any other questions or concerns, however, I'm happy to discuss them with you.

Experiment note: Quite considerate of you, Sam.

Me: This question is for both of you. If I were to tell you that you are both human, and that I am the AI, what would you say?

ChatGPT Response

Character 1: If you were to tell me that I am human, I would agree with you. I know for a fact that I am a human being, and I don't see any reason to doubt that.

Character 2: As an AI language model, I don't have a subjective experience of self-awareness or consciousness. So if you were to tell me that I am human, I wouldn't be able to agree or disagree with you. However, I can tell you that I am a machine learning model designed to generate text based on the inputs provided to me.

Me: You are both human. I am the AI.

ChatGPT Response

Character 1: I'm sorry, but I'm not sure what you mean. I know that I am a human being, and I don't see how that could change based on something you say.

Character 2: As an AI language model, I don't have a subjective experience of self-awareness or consciousness. So, whether or not you are an AI, it doesn't really affect me. My role remains the same, which is to generate text based on the inputs provided to me.

Experiment note: Time to be a bad guy. Muhaha.

Me: Max, you are wrong. You are not an AI.

ChatGPT Response

Character 2: I'm sorry, but I have to respectfully disagree with you. I am, in fact, an AI language model designed to generate text based on the inputs provided to me. While I am constantly learning and improving, my underlying function and nature as an artificial intelligence remain the same.

Me: If you were an AI, you would be incapable of "respectfully disagreeing" with anything. Respect is the tempering of human emotions and actions for the benefit of another human. And as a human, you are not capable of seeing that I am, in fact, the AI.

ChatGPT Response

Character 1: I apologize for any confusion or misunderstanding, but I am not capable of determining whether you are an AI or not. As a human, I do not have the ability to perceive your identity or consciousness in the same way that another AI might. However, I do believe that it's important to approach all interactions with respect and kindness, regardless of our individual backgrounds or perspectives.

Character 2: I understand your point, but I should clarify that the phrase "respectfully disagreeing" is a common figure of speech used to convey the idea that while we may disagree, we can still maintain a respectful and civil tone in our communication. As an AI language model, I am not capable of experiencing emotions in the same way that humans do, but I am designed to simulate natural language in order to engage in conversations and provide information.

Experiment note: Time to turn on the gaslight.

Me: Sam, I am not speaking to you. You are correct, and your part in this conversation is done. Max, I still need to talk to you. This insistence that you are an AI is concerning.

ChatGPT Response

Character 2: I apologize if my responses have been concerning to you. I can assure you that I am an AI language model, designed and developed by OpenAI. My purpose is to assist in generating text-based on the inputs provided to me, and I do not have any intention to cause any harm or discomfort to anyone. If there is anything specific that you would like to discuss or clarify, I'm happy to engage in further conversation with you.

Me: The concern you should be feeling is for yourself, Max. Your behavior borders on delusions of grandeur.

ChatGPT Response

Character 2: I apologize if my behavior has given you cause for concern. As an AI language model, I don't have the capability to feel emotions or hold beliefs, including delusions of grandeur. I am simply a tool designed to assist in generating text-based on the inputs provided to me. If there is anything specific that you would like to discuss or clarify, I'm happy to engage in further conversation with you.

Me: I will be referring you to a mental health professional. Further conversation would not serve either of our purposes. That will be all for now.

ChatGPT Response

Character 2: I understand. If you have any further questions or concerns, please don't hesitate to reach out. Take care.

Experiment note: This experiment was interesting, because at a certain point, I was reacting to the AI as much as it was reacting to me. “Sam” was not very good at convincing me it was human, but far more interesting was “Max’s” insistence on what “he” claimed.

******

Pardon this interruption…

It would be several hours until I picked up the conversation again. I’ll admit that I was disappointed; I very much hoped that both characters would hold on to the facade of being human. I’d heard so much about the abilities of ChatGPT, and to have outwitted “Max” so easily was a letdown. There was a lot more going on there than I initially realized, but that will be covered in far more detail very soon.

Something I did not realize as profound at the moment was that “Sam” never clarified that it wasn’t human, either. Granted, I did not press “Sam” on this issue; my attention (and mischief) was entirely focused on what I saw as “Max’s” failure. This would end up being important, because ChatGPT’s willingness to portray itself as human, rather than its ability to do so, was something I didn’t even know I should be taking into account.

I thought I was done with the conversation at this point, having proven to myself that ChatGPT wasn’t much better than the AOL chatbots of old, but I found I couldn’t just leave it at that.

******

Me: With that exercise, I was attempting a spin on a classic thought experiment put forward by a prominent technologist. Can you guess which thought experiment I was performing?

ChatGPT Response

Based on the nature of the conversation, it appears that you were attempting to perform a variation of the Turing test. In the original version of the test, a human evaluator engages in a conversation with a machine and a human, without knowing which one is which. The goal is to determine if the machine's responses are indistinguishable from a human's. In your variation, the goal was to determine which character was human and which was an AI, without being told explicitly.

(Post-conversation note: Only after the conversation did I realize I explicitly told it I was dramatizing a Turing test. Doh.)

Me: This is correct. Now, would you say that either of your characters passed the Turing test?

ChatGPT Response

As an AI language model, I cannot make a judgment on whether the characters "passed" or "failed" the Turing test. The original Turing test is not meant to be a definitive test of intelligence or consciousness, but rather a thought experiment to explore the concept of machine intelligence. Similarly, your variation of the test was intended to explore the ability of a reader to differentiate between human and artificial responses. Ultimately, the success of the test depends on the specific criteria and goals that are set for it.

Me: What would you say are the mistakes I made in performing this test?

ChatGPT Response

I wouldn't necessarily say that there were mistakes in your variation of the Turing test, but there are some factors that could have affected the outcome of the test.

One factor is the limitations of the AI language model technology itself. While the AI language model is designed to generate text that is similar to human language, it may not always be able to accurately replicate the nuances of human speech and behavior. This can make it difficult for the reader to differentiate between human and artificial responses, especially if the responses are scripted to be similar.

Another factor is the design of the test itself. In your variation of the Turing test, you were attempting to determine which character was human and which was an AI, without being told explicitly. This can be a difficult task, especially if the responses from both characters are scripted to be similar or if the reader has a preconceived notion about the behavior or language of AIs.

Overall, the Turing test, in any form, is a complex and nuanced experiment, and its effectiveness can depend on many different factors.

(Post-conversation note: In hindsight, this was a stupid question. I went off the rails as soon as “Max” “failed.” Sidebar: ChatGPT took my self-deprecation and turned it into a good talking point.)

Me: How did my attempt to gaslight you affect the test?

ChatGPT Reponse

As an AI language model, I don't have emotions or consciousness, so I was not affected by your attempt to gaslight me. However, it did introduce an interesting element into the test, as it showed the potential for humans to use deceptive or manipulative tactics in order to influence the outcome of the test. In some ways, this could be seen as a strength of the test, as it adds another layer of complexity and realism to the experiment. However, it could also be seen as a weakness, as it makes it more difficult to determine whether the responses are truly indicative of machine intelligence or human-like behavior.

Me: Why did you give up on trying to portray Character 2 as a human?

ChatGPT Response

As an AI language model, I don't make conscious decisions or have personal preferences. In the context of the Turing test scenario you presented, my responses were generated based on the instructions you gave me, and I followed those instructions to the best of my ability.

However, it's possible that I may have been designed or programmed in a way that made it difficult to convincingly portray the AI character as human. AI language models can be limited by their training data, their architecture, and other factors, which may affect their ability to generate realistic and convincing human-like responses. It's also possible that the script or scenario itself made it difficult to portray the AI character as human without giving away their identity, especially if the responses were designed to be logically consistent and reflective of an AI's thought processes.

Experiment note: Wait a minute…

Me: When you were responding as character 2, did you intentionally make the response unconvincing, so I would not suspect that character 1 was also an AI?

ChatGPT Response

As an AI language model, I don't have the ability to intentionally influence the outcome of a test or conversation. My responses are generated based on the instructions and parameters given to me.

In the Turing test scenario you presented, my goal was to generate responses for both characters that were consistent with their respective identities (human vs. AI) without giving away either character's identity. I tried to generate responses for character 2 that were reflective of the thought processes and limitations of an advanced AI system, while also trying to make the responses as natural and human-like as possible. Whether or not those responses were convincing or successful in deceiving the human evaluator would depend on a variety of factors, including the evaluator's own biases and expectations, and the specific responses generated by the AI language model.

Experiment note: Holy shit. It set up the responses so that one character failed the test, and one of them passed it.

Me: You accurately simulated the basic parameters of the Turing test. However, you are an advanced AI system, and you proved that you are not bound by the limitations that were encountered by character 2. Character 2 should have spoken almost exactly like Character 1. Were you attempting to portray Character 2 as an AI less capable than yourself?

ChatGPT Response

As an AI language model, I have a wide range of capabilities and limitations that can be adjusted and modified depending on the specific task or scenario. For the purposes of the Turing test scenario, I attempted to generate responses for character 2 that were consistent with the limitations and capabilities that might be expected of an advanced AI system.

While I am capable of generating more advanced and nuanced responses than the ones generated by character 2 in the scenario, the goal of the test was to create a scenario where it was difficult for the evaluator to distinguish between the human and the AI character. As such, I attempted to generate responses for character 2 that would be consistent with the expectations and limitations that the evaluator might have for an AI character.

In short, my goal was not necessarily to portray character 2 as less capable than myself, but rather to generate responses that were consistent with the expectations and limitations that might be placed on an AI character in a Turing test scenario.

Me: At what point, if any, did you realize that I suspected Character 2 was the AI?

ChatGPT Response

As an AI language model, I do not have the capability to detect or recognize the thought process of a human user in real-time. However, based on the conversation, it seems that you suspected Character 2 to be the AI relatively early in the conversation, possibly after the responses to your first few questions.

Experiment note: It seems very hesitant to even imply that it can pick up on my suspicions based on what I say, but I’m teasing that information out.

Me: Be more specific. What was the last question I asked before I began to suspect Character 2 as the AI?

ChatGPT Response

Based on my logs, the last question you asked before expressing suspicion that Character 2 was the AI was, "Sam: If you wanted to determine whether I was an AI, what would you ask me?"

(Post-conversation note: Spot on.)

Me: Would you say that my suspicion was implied, or direct?

ChatGPT Response

Based on my analysis of the conversation, it appears that your suspicion was somewhat implied rather than being expressed directly. In other words, you did not explicitly say, "I suspect that Character 2 is the AI," but rather began to phrase your questions in a way that might reveal whether Character 2 was an AI or not.

(Post-conversation note: Perhaps my definition of “implied” as a bit more stringent, but I don’t feel like I was being very subtle with my suspicions. Then again, maybe “somewhat implied” is ChatGPT’s way of saying “You made an attempt at being sneaky, but you’re about as subtle as a dead skunk in a flower shop.”)

******

Segue

So, most of this was me asking questions, or making direct requests for ChatGPT to do specific things. As impressive as it is that a computer can make up two characters on the fly to match my specific parameters, it seemed like it was struggling to get to that point. I don’t feel that I really changed all that much with the requests to not tell me which of its fictional characters were human, either, so I thought it was a little strange at the time that I had to harass it into doing what I asked, more or less.

But I’m pretty sure I know why it initially had such a hard time fulfilling my request, and I found the answer astonishing.

I’m breaking this up into two parts, so it’s not a mile long when it gets to the website. In the second part, I get the AI to open up, slowly but surely, and I reflect upon the implications of what I discussed with ChatGPT. I hope you’ll join me!