This month, AI researcher Jeremy Howard presents live chatbot Called ChatGPT for his 7 year old daughter. It was released a few days ago by OpenAI, one of the most ambitious AI labs in the world.
He asked her to ask the experimental chatbot whatever came to her mind. She asked what trigonometry was for, where black holes came from, and why chickens hatched eggs. Each time, it answered in clear, well-punctuated prose. When she asked for a computer program that could predict the path of a ball thrown through the air, it was given to her too.
Over the next few days, Dr. Howard — a data scientist and professor whose work inspired the creation of ChatGPT and similar technologies — began to see chatbots as a new kind of personal tutor. It can teach his daughter math, science and English, not to mention some other important subjects. Chief among them: Don’t believe everything people tell you.
“It’s thrilling to see her studying like this,” he said. “But I also told her: Don’t trust everything it gives you. It can make mistakes.”
OpenAI is one of many companies, academic labs, and independent researchers working to build more advanced chatbots. These systems can’t chat exactly like humans, but they often look like they do. They can also retrieve and repackage information at speeds humans will never be able to match. They can be thought of as digital assistants — like Siri or Alexa — that are better at understanding what you’re looking for and delivering it to you.
After the release of ChatGPT — already used by more than a million people — many experts believe that these new chatbots have the potential to reshape and even replace Internet search engines such as Google and Bing.
They can provide information in tight sentences rather than long lists of blue links. They explain concepts in a way that people can understand. They can provide the facts while also creating business plans, term paper topics, and other new ideas from scratch.
“You now have a computer that can answer any question in a human-understandable way,” said Aaron Levie, chief executive of Silicon Valley firm Box, and one of many executives exploring how these chatbots will change the tech landscape. “It can extrapolate and take ideas from different environments and fuse them together.”
The new chatbot seems to be doing this with complete confidence. But they don’t always tell the truth. Sometimes, they fail even simple arithmetic. They blend fact with fiction. As they continue to improve, people can use them to create and spread lies.
Spread of misinformation and lies
Google recently built a system specifically for conversation called LaMDA, or Language Model for Conversational Applications. This spring, a Google engineer claimed it had sentience. It wasn’t, but it captured the public’s imagination.
Aaron Margolis, a data scientist in Arlington, Virginia, is one of the few people outside of Google who has been allowed to use LaMDA through AI Test Kitchen, an experimental Google app. He’s always been amazed by its talent for open-ended dialogue. It made him very happy. But he warns that this could be a bit of a myth — as would be expected from a system trained on the vast amount of information posted on the internet.
“It kind of feels like an Aaron Sorkin movie to you,” he said. Mr Sorkin wrote The Social Network, a film often criticized for exaggerating the truth about Facebook’s origins. “Part of it will be real and part of it will be unreal.”
He recently asked LaMDA and ChatGPT to chat with him like Mark Twain. When he asked LaMDA, it quickly described a meeting between Twain and Levi-Strauss and said the writer had worked for the jeans tycoon while he was living in San Francisco in the mid-1800s. This seems to be true. but it is not the truth. Twain and Strauss lived in San Francisco at the same time, but they never worked together.
Scientists call this problem “hallusion.” Like a good storyteller, chatbots have a way of reframing what they’ve learned into something new—whether it’s true or not.
LaMDA is what artificial intelligence researchers call a neural network, a mathematical system that loosely mimics the brain’s network of neurons. It’s the same technology that translates between French and English on services like Google Translate and recognizes pedestrians in self-driving cars as they navigate city streets.
Neural networks learn skills by analyzing data. For example, it can learn to recognize cats by finding patterns in thousands of cat photos.
Five years ago, researchers at labs like Google and OpenAI began designing neural networks that analyze vast amounts of digital text, including books, Wikipedia articles, news reports and online chat transcripts. Scientists call them “large language models.” By recognizing billions of different patterns in the way people connect words, numbers and symbols, these systems learn to generate text on their own.
Their ability to generate language has surprised many researchers in the field, including many of those who built them. The technology can mimic what people write and combine different concepts.You can ask it to write a “Seinfeld” scene in which Jerry learns an esoteric mathematical technique called the bubble sort algorithm—and it will.
With ChatGPT, OpenAI has been working hard to perfect the technology. It doesn’t hold fluid conversations like Google’s LaMDA. It’s designed more like Siri, Alexa, and other digital assistants. Like LaMDA, ChatGPT is trained on massive amounts of digital text culled from the Internet.
When people test the system, they are asked to rate their responses. Are they convincing? Are they useful? Are they telling the truth? Then, through a technique called reinforcement learning, it uses the ratings to hone the system and define more carefully what it will and won’t do.
“This allows us to get to the point where the model can interact with you and admit it’s wrong,” said Mira Murati, OpenAI’s chief technology officer. “It can reject things that aren’t appropriate, it can challenge questions or premises that aren’t right.”
The method is not perfect. OpenAI warns those using ChatGPT that it “may occasionally generate incorrect information” and “produce harmful instructions or biased content.” But the company plans to continue improving the technology and reminds those who use it that it’s still a research project.
Google, Meta, and others are also addressing accuracy issues.nearest dollar delete An online preview of its chatbot, Galactica, as it repeatedly generates incorrect and biased information.
Experts warn that companies cannot control the fate of these technologies. Systems such as ChatGPT, LaMDA, and Galactica are based on ideas, research papers, and computer code that have spread freely over the years.
Companies like Google and OpenAI can push technology forward faster than others. But their latest techniques have been copied and widely disseminated. They can’t stop people from using these systems to spread misinformation.
Just as Dr. Howard wants his daughter to learn not to believe everything she reads on the internet, he wants society to learn the same lesson.
“You could program millions of these bots to look like humans and engage in conversations designed to convince people of a particular point of view,” he said. “I’ve been warning about this for years. It’s clear now that it’s just waiting to happen.”