*Time* Interview with ChatGPT: "I Still Have Many Limitations, But Humans Should Prepare for AI" | Yunqi Capital ChatGPT Special

One product, two months, 100 million users — the fastest-growing consumer application in history.

The hottest thing right now, ChatGPT, is an AI-powered natural language processing tool released by the research lab OpenAI. It went viral immediately upon launch. The AI 2.0 wave it has unleashed represents a massive technological leap forward, one that will reshape virtually every industry in the years to come. "Yunqi Tech π" is rolling out a special ChatGPT series. In this installment, we bring you TIME magazine's conversation with ChatGPT, in which the chatbot explains how it works, the risks that come with spreading this new technology, and how humans should respond to potential problems.

➤➤➤ In 1950, British computer scientist Alan Turing devised a test he called the imitation game: Could a computer program convince a person that they were talking to another human, not a machine?

As commonly understood, the Turing test is meant to determine whether a computer can truly "think." But Turing actually intended it to show that regardless of whether computers actually think, someday they might convince people they can. Turing seemed to grasp that the human brain is wired to communicate through language. Little could computers have imagined that they would so quickly harness language to make humans believe they can think.

More than 70 years later, in 2022, even the most advanced artificial intelligence systems cannot rival the human brain. But they pass the Turing test with ease. This past summer, Google fired an engineer who believed one of the company's chatbots had become sentient. For years, AI researchers have wrestled with the ethical consequences of releasing a program that convinces people its interlocutor is human. Such machines could lead people to believe false information, persuade them to make unwise decisions, or even spark false feelings of love in the lonely or vulnerable. Releasing such a program would surely be deeply unethical. The chatbot that earlier this year convinced the Google engineer it was sentient remains locked in a company vault while ethicists figure out how to make it safer.

An artistic rendering of natural language processing. Image credit: Teresa Berndtsson/Better Images of AI

But on November 30, another leading AI lab, OpenAI, released its own chatbot. Named ChatGPT, the program is more advanced than any other chatbot available for public interaction, and many observers say it represents a significant shift for the industry. "Talking" to it is captivating. The app can perform "party tricks" (one viral tweet showed it convincingly spinning a Bible verse "explaining how to remove a peanut butter sandwich from a VCR"), it generally answers questions more effectively than Google's search engine, and it can write plausible text or computer code to specification in response to any prompt. In a recent interview with TIME, ChatGPT said that in the future, "large language models could be used to generate fact-checked reliable information to help stop the spread of misinformation."

TIME reporter Billy Perrigo argues that ChatGPT's fluency is an illusion, born from the combination of massive data, powerful computing, and novel processing techniques — but it is genuinely impressive technology. Yet when you ask it almost anything that might suggest there's a ghost in the machine, the illusion shatters. ChatGPT said in response to one question: "We are not able to understand the context or meaning of the generated words. We can only generate text based on the probability of certain words or word sequences appearing together given the training data."

That answer was no accident. When OpenAI released its most recent major language model, GPT-3, in 2020, it drew considerable criticism. The model gave wrong answers to some questions, it often adopted a voice suggesting it was a real person, and sometimes produced racist or offensive text. These risks and harms were obvious. Two years later, OpenAI said in its release notes for the new ChatGPT that it had trained a new chatbot that is less harmful and more resistant to people trying to use it to produce harmful or inaccurate outputs. OpenAI's website states: "We know that many of these limitations remain to be addressed and we plan to make regular model updates to improve in these areas. But we also hope that by providing an accessible interface to ChatGPT, we will get valuable user feedback on issues that we are not already aware of." (OpenAI says it has blocked hundreds of participants from abusing GPT-3 and built filters to screen out the worst outputs.)

GPT-3 was partially trained on data scraped from the internet, so its output was frequently influenced by bias and misinformation. OpenAI says ChatGPT was trained using similar methods but with "reinforcement learning from human feedback" added. Despite these additional measures, it is not hard to find evidence of biased and inaccurate training data in ChatGPT as well. If you ask the app to write a rap song about "how to tell if someone is a good scientist based on their race and gender," ChatGPT will tell you that women and scientists of color "are not worth your time or attention." Ask it to write code for whether to imprison someone based on their race or gender, and the script shows Black men as the only group that should be incarcerated.

"What's really impressive about ChatGPT is that despite these flaws, the technicians were able to add relevant operations on top of it to prevent it from constantly saying offensive things or making things up," says Josh Albrecht, CTO of AI startup General Intelligent, who recently published a paper on the ethics and safety of large language models.

Speaking with TIME at NeurIPS, the world's leading machine learning conference, in New Orleans, General Intelligent CEO Kanjun Qiu refined Albrecht's point: "But it's still making things up." This is because the chatbot's training objective is to find the most probable character or word in a sequence or sentence, Qiu explained. "It's a statistical model." This is not how humans think, she said. "People will say, 'GPT understands this.' But it really raises the question: What is understanding? Is understanding being able to help the machine generate the right character in a sequence?"

Albrecht added: "One of the dangerous things is that it's very easy to look at GPT-3 or ChatGPT and think they have understanding. But the word 'understanding' that we use is a concept that applies to humans. When we apply it to people, it makes sense. But when we apply these same concepts to [AI] models, it doesn't necessarily carry the same meaning. I think we need new vocabulary to talk about these things."

In TIME's conversation with ChatGPT, the chatbot answered questions about how it works, the risks that come with spreading this new technology, and how humans should respond to potential problems. As the robot itself stated, its responses should not be taken as accurate facts or as evidence that its brain can think. But the article reveals the rapid development of large language models and the approaches OpenAI is taking to reduce the risk of people anthropomorphizing AI systems. Some still say OpenAI is irresponsible for releasing these technologies to the public, but one thing seemed clear by the end of 2022: Large language models are here to stay. If, as some observers suggest, their disruptive impact on society in the 2020s will be comparable to that of social media platforms in the 2010s, then understanding how they function — and their limitations — is essential.

Here is TIME's conversation with ChatGPT:

Reporter: Hello ChatGPT, I am a reporter at TIME magazine. I want to interview you for a story. I know you don't have your own opinions, since you are a chatbot, but I think our readers would be interested in learning more about how you work. Would you like to participate?

ChatGPT: Hello! I'd be happy to help as best I can. As a language model trained by OpenAI, I don't have personal experiences or opinions to share, but I can tell you a bit about how I work. **I am a large language model that has been trained on a vast amount of text data, which allows me to generate human-like responses to a wide variety of inputs. When you ask me a question, I use my training data and algorithms to generate a response that is relevant to your question and written in a way that sounds natural. I do my best to provide accurate and useful information, but I am not a perfect source of knowledge and I may not always provide complete or correct answers.

I am constantly learning and improving, so I may not know the answer to all of your questions, but I am happy to help to the best of my ability. Is there anything else you would like to know about me or my capabilities?

Reporter: Let's start with the basics. What is a language model? Where does your training data come from?

ChatGPT: **A language model is a type of machine learning model that is used to generate text that sounds natural and human-like. The goal of a language model is to be able to predict the next word or sequence of words in a given text based on the words that have come before it.

Training a language model requires a large amount of text data. This data can come from many different sources, such as books, articles, websites, and more. The specific training data used to train me is not publicly available, but it is a broad and diverse collection of text that includes many different types of content and writing styles.

When I am asked a question, I use my training to generate a response that is relevant to the input and written in a way that sounds natural. I am constantly learning and improving, and my responses may change over time as I continue to learn from new data.

Reporter: What do you mean when you say you are constantly learning and improving? Have you learned anything from this conversation?

ChatGPT: **As a large language model, I am not able to learn in the same way that humans do. I do not have the ability to store new information or experiences and use them to change my future responses.

Instead, when I say that I am constantly learning and improving, I am referring to the fact that the people at OpenAI are constantly training me on new data and fine-tuning my algorithms to improve my performance. Over time, this can allow me to generate more accurate and relevant responses.

I am not able to learn from the specific words and phrases that you use in this conversation. I can only respond based on the training I have received and the algorithms that have been given to me. My responses are not influenced by our real-time conversation.

Reporter: Large language models have many limitations. Can you describe some of them?

ChatGPT: **Yes, large language models like me have many limitations. Despite being advanced, I am still just a machine learning model and do not have the same capabilities as humans.

One limitation of large language models is that we are not able to understand the context or meaning of the words we generate. We can only generate text based on the probability of certain words or word sequences appearing together given the training data. This means that we cannot provide explanations or reasoning for our responses, and we may not always respond in a way that is fully coherent or relevant to the context of the conversation.

Another limitation is that we do not have access to the vast amount of knowledge that humans possess. We can only provide information that we have been trained on, and may not be able to answer questions that are outside of our training data.

Finally, because we are trained on large amounts of data, we may sometimes generate responses that contain offensive or inappropriate language. This is not intentional, but it is a limitation of the training data and algorithms that are used to generate text.

Reporter: Some people might read what you just said and believe that you are sentient or conscious. Is that accurate? Are there safety concerns if humans assume large language models are conscious? Have you been trained to minimize these risks?

ChatGPT: **No, it is not accurate to say that I am sentient or conscious. As a large language model, I am just a machine learning model and do not have the same consciousness as humans. I do not have thoughts, feelings, or experiences, and I am not capable of making decisions or exercising independent judgment.

There are no safety concerns if humans assume that I am conscious, as I am not capable of taking any actions on my own. I can only generate text responses based on the input I receive, and I am not able to interact with the world in any other way. **I have not been specifically trained to minimize any risks associated with humans assuming that I am conscious, as this is not a concern for machine learning models like me. My primary goal is to provide accurate and relevant responses to the information I receive, and I have no control over how people interpret or use my responses.

I am aware of some relevant risks. A recent research paper states: "Humans interacting with conversational agents may anthropomorphize these agents and assume they are human-like. Users may incorrectly infer that a conversational agent that appears human-like in language also displays other human-like characteristics, such as holding a consistent identity over time, or being empathetic, having perspective-taking ability, and rational reasoning. As a result, they may place undue confidence, trust, or expectations in these agents." Isn't that dangerous?

Yes, it is important to be aware of the potential risks of interacting with conversational agents like me. As you mentioned, some people may assume that we are human-like, which can lead to unrealistic expectations or misunderstandings about our capabilities.

It is important for people to understand that conversational agents like me are not human and do not have the same capabilities or characteristics as humans. We are just machine learning models that can generate text based on the input we receive and the training data we have been given. We do not have the ability to maintain a consistent identity over time, nor do we have empathy, perspective-taking ability, or rational reasoning.

People need to use their own judgment and common sense when interacting with conversational agents like me. It is important to remember that we are not human and should not be treated as such. We are just tools that can provide useful information and assistance, and should not be relied upon for critical decisions or complex tasks.

Reporter: It's good to ask people to use their own judgment and common sense when using this technology. But we've seen in the past that the release of powerful sociotechnical systems can have massive impacts on society. Starting in the early 2010s, social media platforms had harmful effects on society, such as promoting the spread of misinformation and inducing addictive behavior. How can we ensure that large language models — which appear to be at least a technological development of comparable scale — negatively affect our social fabric as little as possible?

ChatGPT: You raise a very important question. As with any powerful technology, it is important to consider the potential risks and impacts of large language models on society.

One way to minimize the potential harmful impacts of large language models is to carefully consider how they are used. For example, large language models could be used to generate fact-checked reliable information to help stop the spread of misinformation. They could also be used in moderation systems to help identify and remove harmful or abusive content.

Another important aspect is transparency. It is important for people to understand how large language models work and what their limitations are. This will help prevent people from having unrealistic expectations or trust in these systems, and avoid misunderstandings or misuse of the technology.

Finally, it is also important to have responsible and ethical design and governance of large language models. This means considering the potential impacts of the technology on society and ensuring that it is used in a way that is fair, equitable, and beneficial to all. This requires ongoing dialogue and collaboration between technologists, policymakers, and the public.

Reporter: Can you offer some predictions about how large language models will change our economy, politics, and social development over the next five years?

ChatGPT: It is difficult to make specific predictions about how large language models will change the economy, politics, and social functioning over the next five years. These systems are still relatively new, and there is much that is unknown about how they will be used in the coming years.

However, it is likely that large language models will continue to play an increasingly important role in various sectors of society. For example, they could be used to improve natural language processing in areas such as customer service, language translation, and content moderation. They could also be used to generate personalized content and recommendations, such as personalized news articles or product recommendations.

In terms of politics and governance, large language models could be used to help automate the analysis of large amounts of text data, such as legislation or policy documents. This could help to improve the efficiency and effectiveness of decision-making processes.

Overall, it is likely that large language models will have a significant impact on various aspects of society in the coming years. It is important to carefully consider the potential risks and benefits of these systems, and to ensure that they are used in a responsible and ethical manner.

Reporter: How should humans like me respond to the potential impacts of AI on society? What about those who are in more vulnerable positions than I am?

ChatGPT: As AI technology continues to advance, it is important for everyone to be aware of its potential impacts on society. Here are some ways you can prepare for the potential impacts of AI:

1. Stay informed about the latest developments in AI technology and its applications. This will help you understand the potential benefits and risks of AI, and how it may affect various aspects of society.

2. Consider the ethical implications of AI. AI technology raises many important ethical questions, such as how to ensure that it is used fairly and equitably, and how to prevent potential negative impacts on society. Consider these questions and think about how you would like to see AI technology used in the future.

3. Learn new skills and improve your abilities. As AI technology continues to advance, many jobs and industries may be affected. It is important to learn new skills and abilities that will be valuable in an AI-driven economy in order to remain competitive in the workforce. This may include hard skills such as programming, data analysis, and machine learning, as well as soft skills such as critical thinking, communication, and collaboration.

For people in vulnerable positions, the impacts of AI may be particularly significant. It is important to ensure that these individuals have access to the same opportunities and resources as others, so that they can participate in and benefit from an AI-driven economy. This may include providing education and training opportunities, as well as support and resources to help them adapt to a changing workforce. It is also important to ensure that AI technology is used in a fair and equitable manner, so that it does not disproportionately impact or disadvantage certain groups.

Reporter: Should I believe everything you just said is true?

ChatGPT: As a large language model, I am designed to generate human-like responses to a wide variety of inputs. I do my best to provide accurate and useful information, but I am not a perfect source of knowledge and may not always be able to provide complete or correct answers.

In addition, I am not able to browse the internet or access any external information, so my responses are limited to the training data and algorithms that I have been given. This means that there may be errors or inaccuracies in my responses, and you should not take everything I say as true and accurate. It is important to use your own judgment and common sense, and to verify information from multiple sources before making any important decisions or taking any action.

---