Anyone lured by AI chatbots like ChatGPT and Bard – wow, they can write articles and recipes! He eventually runs into what is known as hallucinations, the tendency of artificial intelligence to fabricate information.
Chatbots, which guess what to say based on information gleaned from all over the internet, can’t help but get things wrong. And when they fail—by posting a cake recipe with wildly inaccurate flour measurements, for example—it can be a real killer.
However, as mainstream technology tools continue to incorporate AI, it is imperative to grapple with how it is used to serve us. After testing dozens of AI products over the past two months, I’ve concluded that most of us use technology in a suboptimal way, largely because tech companies have given us bad directions.
Chatbots are least useful when we ask them questions and then hope the answers they come up with on their own are correct, which is how they were designed to be used. But when directed to use information from trusted sources, such as authoritative websites and research papers, AI can perform useful tasks with a high degree of accuracy.
“If you give them the right information, they can do interesting things with it,” said Sam Heutmaker, founder of Context, an AI startup. “But on their own, 70 percent of what you get isn’t going to be accurate.”
With the simple modification of advising chatbots to work with specific data, they have produced clear answers and helpful advice. Over the past few months it has transformed me from a cranky AI skeptic into an enthusiastic power user. When I went on a trip with an itinerary planned by ChatGPT, it went well because the recommendations came from my favorite travel sites.
Directing chatbots to specific, high-quality sources such as websites from well-established media and academic publications can help reduce the production and spread of misinformation. Let me share with you some of the techniques I have used to get help with cooking, researching and planning travel.
Chatbots like ChatGPT and Bard can write recipes that sound good in theory but don’t work in practice. In an experiment conducted by the New York Times Food Desk in November, an early AI model created recipes for a Thanksgiving menu that included an extra-dry turkey and a thick brownie.
I’ve also experienced disappointing results with AI-generated seafood recipes. But that changed when I tried ChatGPT plugins, which are basically third-party apps that work with a chatbot. (Only subscribers who pay $20 per month to access ChatGPT4, the latest version of the chatbot, can use the plugins, which can be activated in the settings menu.)
In my list of ChatGPT plugins, I chose Tasty Recipes, which pulls data from Tasty owned by BuzzFeed, a well-known informational site. Then she asked the chatbot to create a meal plan that included seafood dishes, ground pork, and vegetarian sides, using recipes from the site. The robot presented an inspiring meal plan, including ham and lemongrass, grilled tofu sandwiches and everything in the fridge; Each meal suggestion included a link to a recipe on Tasty.
To get recipes from other posts, I used Link Reader, a plugin that allows me to paste a web link to create meal plans using recipes from other trusted sites like Serious Eats. The chatbot pulled data from the sites to create meal plans and asked me to visit the sites to read the recipes. It took extra effort, but it beat an AI-prepared meal plan.
When I did research for an article on a popular video game series, I turned to ChatGPT and Bard to refresh my memory of previous games by summarizing their plots. They missed important details about the games’ stories and characters.
After testing several other AI tools, I concluded that in order to research, it was necessary to focus on reliable sources and quickly check data for accuracy. I eventually found a tool that does just that: Humata.AI, a free web application that has become popular with academic researchers and lawyers.
The app allows you to upload a document such as a PDF, and from there the chatbot answers your questions about the material along with a copy of the document, highlighting the relevant parts.
In one test, I uploaded a research paper I found on PubMed, a government-run search engine for scientific literature. The tool produced a proper summary of the lengthy document in minutes, a process that would have taken hours, and I glanced at highlights to double-check the summaries were accurate.
Cyrus Khajwande, founder of Humata, based in Austin, Texas, said he developed the app when he was a researcher at Stanford and needed help reading thick science articles. The problem with chatbots like ChatGPT, he said, is that they rely on outdated models of the web, and so the data may lack proper context.
When a Times travel writer recently asked ChatGPT to compose an itinerary for Milan, the bot directed her to visit a central part of the city that was deserted because it was an Italian vacation, among other things.
I had better luck when I requested an itinerary for me, my wife, and our two dogs in Mendocino County, California. As I did when planning a meal, I asked ChatGPT to include suggestions from some of my favorite travel sites, like Thrillist, which is owned by Vox, and the travel section of The Times.
Within minutes, the chatbot created an itinerary that included dog-friendly restaurants and activities, including a farm with wine and cheese pairings, and a train to a popular hiking trail. This saved me many hours of planning, and most importantly, the dogs had a great time.
Google and OpenAI, which work closely with Microsoft, say they are working to reduce hallucinations in their chatbots, but we can actually reap the benefits of AI by controlling the data that bots rely on to come up with answers.
In other words, said Nathan Beneish, a venture capitalist who invests in AI companies, the main advantage of training machines with massive data sets is that they can now use language to simulate human thinking. An important step for us, he said, is to pair this capability with high-quality information.