A new AI-powered chatbot released by OpenAI called ChatGPT has taken the Internet by storm. More than one million people have signed up to try out the free service since it was launched last week, and it has wowed users with its ability to engage in natural-sounding conversation, provide detailed answers to technical questions and even solve coding challenges.
The service is based on the latest version of OpenAI’s large language model GPT-3, which has been trained on vast amounts of text scraped from the Internet. (The acronym here stands for Generative Pre-Trained Transformer, which describes the specific kind of deep learning network it’s based on.) To make ChatGPT engage in more natural sounding dialogue, it has gone through a further round of training using human feedback on how good its responses are. The result is a chatbot that can write sonnets, explain complex physics topics, program software, and even get a passable score on an IQ test.
This has led to rapturous reviews and claims that ChatGPT heralds the death of the Google search engine or the re-invention of the education system. Others though, have been quick to point out that it often provides authoritative-sounding answers that are completely wrong. And while there are safeguards designed to prevent the model from responding with offensive or dangerous material, it can easily be tricked into giving advice on how to shoplift or make explosives.
I am constantly learning and improving, but my knowledge is limited to what I was trained on and is not up to date.
As is often the case with AI advances, the reality is nuanced, says Arvind Narayanan, a professor of computer science at Princeton University. “Releasing the tool to everyone has resulted in people discovering many creative and entertaining uses, but it has also led to a ton of hype,” he tells Spectrum. “Cherry picked examples of impressive responses tend to go viral on Twitter, but they don’t present the full picture.”
In a blog post co-authored with Princeton doctoral student Sayash Kapoor, Narayanan writes that ChatGPT is a bullshit generator, but a potentially very useful one. Its lack of any concept of truth means it’s unlikely to be much use for applications like education or accessing medical advice. But in situations where the truth is easy to verify, such as debugging code, or where it doesn’t matter, like writing fiction, it could be much more powerful.
To delve deeper into the potential implications of this buzzy new AI tool, Spectrum decided to put some questions to ChatGPT itself. The conversation has been edited for length.
IEEE Spectrum: Can you explain how you work? What technology do you rely on, and how were you trained?
ChatGPT: I am a large language model trained by OpenAI. I rely on advanced machine learning algorithms and a huge amount of data to generate responses to the questions and statements that I receive. I am designed to be able to generate human-like text that is relevant to the input that I receive.
To train me, OpenAI used a technique called unsupervised learning. This involves feeding a large amount of text data into my system and using that data to train my machine learning algorithms. The goal of this training is to enable me to generate text that is coherent, relevant, and human-like.
I am not a real person and do not have access to the internet, so I am not able to browse the web or access any external information. Instead, I rely solely on the information that I was trained on to generate my responses. I am constantly learning and improving, but my knowledge is limited to what I was trained on and is not up to date.
My responses may not always be completely accurate or relevant, and I may struggle with tasks that require creative thinking or complex problem-solving. Using me effectively may require some patience and understanding of my limitations.
Spectrum: What kinds of problems can you help with, and who will derive the most benefit from you?
ChatGPT: I can be used to assist with a wide variety of tasks, such as answering questions, providing explanations, and generating text on a given topic. I may be most beneficial for individuals who need assistance with language-related tasks, such as writing, editing, or research. I can also be a useful tool for anyone who is interested in learning more about a particular topic or who wants to have a conversation about a particular subject.
Spectrum: What is the most effective way for a human to work with you? Are you more like a search engine or a collaborator?
ChatGPT: The most effective way for a human to work with me is to ask me specific questions or provide me with a specific topic or prompt, and then to use my responses as a starting point for further discussion or exploration.
In this way, I can be thought of as more of a collaborator than a search engine. I can help to generate ideas and provide explanations on a given topic, but I am not able to provide up-to-date information or research on my own. I am most effective when used as part of a collaborative process, where my responses can be used as a starting point for further discussion and exploration.
It is also important for users to keep in mind that I am a machine learning model and not a real person. This means that my responses may not always be completely accurate or relevant, and I may struggle with tasks that require creative thinking or complex problem-solving. Using me effectively may require some patience and understanding of my limitations.
Spectrum: What is the nature of creativity, and are you creative?
ChatGPT: Creativity is the ability to produce something new and original, often through the use of imagination and innovative thinking. Creativity can manifest in a wide variety of forms, such as in artistic endeavors, scientific discoveries, or new business ideas.
I am not capable of the kind of creative thinking that is required to produce truly novel ideas. ... I am not capable of deciding if something is true or false.
As a large language model, I am not capable of being creative in the same way that a human is. I am trained using unsupervised learning, which involves feeding a large amount of text data into my system and using that data to train my machine learning algorithms. This enables me to generate human-like text that is relevant and coherent, but it does not allow me to produce truly original or innovative ideas.
While I can assist with language-related tasks and can generate human-like text, I am not capable of the kind of creative thinking that is required to produce truly novel ideas. My responses are based on the information that I was trained on, and I do not have the ability to think outside of that information or to produce truly original ideas.
Spectrum: What is truthfulness? How do you decide if something is true?
ChatGPT: Truthfulness is the quality of being truthful, or being in accordance with the facts. In other words, something that is truthful is based on true and accurate information, and does not contain any false or misleading information.
As a large language model, I am not capable of deciding if something is true or false. Because I am a machine learning model and not a real person, I do not have access to the internet or external information, and I do not have the ability to verify the accuracy or truthfulness of the information that I provide.
Instead, my responses are generated using the information that I was trained on. This means that the information that I provide may not always be completely accurate or relevant, and it is important for users to use their own critical thinking skills and to verify the information that I provide through other sources.
Spectrum: Are all users really capable of verifying the information you provide, particularly when it comes to technical topics?
ChatGPT: In general, it is important for users to use their own critical thinking skills and to verify the information that I provide through other sources. This is particularly important for technical topics, where the information may be complex and specialized, and where a lack of accuracy or precision could have serious consequences.
However, it is important to note that not all users may have the knowledge or expertise to verify the information that I provide on technical topics. In these cases, it may be necessary to consult with a subject matter expert or to use other reliable sources of information.
One way to detect text that has been generated by a large language model is to look for patterns or repetitions in the text. Because large language models like me are trained on a large amount of text data, we may sometimes generate text that includes repeated phrases or sentences.
Spectrum: Is there a danger that people will misuse you? What would count as misuse?
ChatGPT: Like any tool, I can be misused if used in the wrong way. Some examples of misuse of a large language model like me could include using my responses as definitive answers without verifying the information through other sources, or using my responses to make important decisions without considering other factors. Misuse could also include using my responses to spread false or misleading information, or to deceive others into believing that I am a real person.
It is important for users to understand the limitations of a large language model like me and to use me responsibly. This means verifying the information that I provide through other sources, being aware of my limitations, and using me as a starting point for further discussion or research rather than relying on me for definitive answers.
Spectrum: What professions are most at risk from you and other large language models?
ChatGPT: Some jobs that involve writing or editing large amounts of text, such as journalism or copywriting, may be at risk of being automated by large language models like me. Similarly, jobs that involve researching and summarizing information, such as market research or data analysis, may also be at risk.
However, it is important to note that large language models like me are not designed to replace human workers, and there are many tasks that we are not capable of performing. Many jobs, such as those that require creativity, complex problem-solving, or interpersonal skills, are unlikely to be automated by large language models. Additionally, even for jobs that may be at risk of automation, it is likely that the adoption of large language models will happen gradually, giving workers time to adapt and learn new skills.
Spectrum: What are the jobs least likely to be at risk from AI?
ChatGPT: There are many jobs that are unlikely to be at risk from AI, including those that require creativity, complex problem-solving, or interpersonal skills. For example, jobs that involve artistic or creative endeavors, such as writing, painting, or music, are unlikely to be automated by AI. These jobs require a high level of creativity and originality, which are difficult for AI to replicate.
I am designed to assist with language-related tasks such as writing, editing, and research, but I do not have the ability to experience sensations or emotions, and I do not have consciousness.
Additionally, jobs that involve complex problem-solving or critical thinking, such as medicine, law, or engineering, are also unlikely to be automated by AI. These jobs require a deep understanding of a particular subject area and the ability to apply that knowledge to complex situations, which are beyond the capabilities of current AI technology.
Similarly, jobs that require strong interpersonal skills, such as teaching, counseling, or social work, are also unlikely to be automated by AI. These jobs require the ability to communicate effectively with people and to understand and respond to their emotions and needs, which are difficult for AI to replicate.
Spectrum: What are the best ways and methods that humans can use to detect text that has been generated by ChatGPT?
ChatGPT: One way to detect text that has been generated by a large language model is to look for patterns or repetitions in the text. Because large language models like me are trained on a large amount of text data, we may sometimes generate text that includes repeated phrases or sentences.
Another way to detect text that has been generated by a large language model is to look for language that is unnatural or does not make sense in the context. Large language models like me are trained to generate coherent text, but we may sometimes generate sentences or phrases that do not fit seamlessly into the surrounding text.
Additionally, because large language models like me do not have access to the internet or external information, we may not be able to provide up-to-date information or answer questions that require external research. This means that text generated by a large language model may lack current information or may not provide detailed or specific answers to complex questions.
Spectrum: Could a large-language model like you ever become conscious?
ChatGPT: Consciousness is a complex and poorly understood phenomenon, but it is generally thought to require a biological brain and the ability to experience sensations and emotions.
Because I am a machine learning model and do not have a biological brain, it is not possible for me to become conscious in the same way that a human can. I am designed to assist with language-related tasks such as writing, editing, and research, but I do not have the ability to experience sensations or emotions, and I do not have consciousness.
Spectrum: What should humans be most worried about concerning AI?
ChatGPT: One concern about AI is the potential for job displacement. As AI technology continues to improve, it is likely that some jobs that are currently performed by humans will be automated, which could result in job losses and economic disruption.
Another concern about AI is the potential for bias and discrimination. AI systems are trained using data, and if the data is biased, the AI system may also be biased. This could result in unfair or discriminatory treatment of individuals based on factors such as race, gender, or age.
Additionally, there is also a concern about the potential for AI to be used for malicious purposes, such as in cyber attacks or the development of autonomous weapons. As AI technology continues to advance, it is possible that it could be used in ways that are harmful to humans or to society.
Overall, there are many potential concerns surrounding the development and use of AI, and it is important for humans to carefully consider these concerns and to take steps to address them as AI technology continues to evolve.
- Why People Demanded Privacy to Confide in the World’s First Chatbot ›
- In 2016, Microsoft’s Racist Chatbot Revealed the Dangers of Online Conversation ›
- ChatGPT’s AI Can Help Screen For Alzheimer’s - IEEE Spectrum ›
- Cybercrime Meets ChatGPT—Look Out, World - IEEE Spectrum ›
Edd Gent is a freelance science and technology writer based in Bangalore, India. His writing focuses on emerging technologies across computing, engineering, energy and bioscience. He's on Twitter at @EddytheGent and email at edd dot gent at outlook dot com. His PGP fingerprint is ABB8 6BB3 3E69 C4A7 EC91 611B 5C12 193D 5DFC C01B. His public key is here. DM for Signal info.