By
Josh TyrangielColumnistDecember 27, 2023 at 7:00 a.m. EST(Ann Kiernan for The Washington Post) Listen8 minShareComment on this storyCommentAdd to your saved storiesSave
It’s your wedding day. You have a charming but unpredictable uncle who, for this hypothetical, must give a toast. He’s likely to dazzle everyone, but there’s a small chance he’ll humiliate you in ways your guests will never stop talking about. Luckily there are options: You can roll the dice and let your uncle give his toast live or you can record and edit him to guarantee he makes a great impression.
Need something to talk about? Text us for thought-provoking opinions that can break any awkward silence.ArrowRightThis is roughly the choice Google faced when it unveiled Gemini, its new suite of artificial intelligence tools. Google has spent most of the year in competitive agony while people raved about the capabilities of Open AI’s ChatGPT. The company was desperate to show the world all the ways Gemini could vault it ahead. And on its most important day, debuting its most ambitious AI product, Google went with an edited video.
Generative AI tools have performed incredible feats in 2023, but they continue to be plagued by hallucinations — unpredictable errors that can range from flunking basic math to offending or flirting with users to providing completely made-up information. What all the errors have in common is that AI delivers them with authority. To users, a hallucination can feel like gaslighting.
Advertisement
The major players have been whistling past the hallucination problem in public because it raises serious questions about their products’ reliability. Also: That word is so devastating. Can you imagine a better wrapper for humanity’s fears about AI? Hallucination. “Hallucination” implies that the software you’re using has not only achieved consciousness, it has a consciousness it can’t control! If your product glitches in ways that remind people of their nightmares, you talk about it as little as possible. You definitely stay away from live demos.
5 questions about artificial intelligence, answered
There are a lot of disturbing examples of hallucinations, but the ones I’ve encountered aren’t scary. I actually enjoy them. (I also like watching drunk uncles at weddings.) Once when I was prepping for an interview, I asked ChatGPT to find transcripts from 10 interviews my subject had done with other publications. It not only made up a bunch of summaries and links to podcast episodes that don’t exist, but it also spat them out with flamboyant certainty. It was like talking to a George Santos bot. Google’s Bard and Anthropic’s Claude, on the other hand, are more prone to taking a bad bit of logic or a slightly imprecise prompt and building on it, “Yes, and-ing” me into oblivion like terrible improv comics.
In the public vacuum left by Big Tech, academic research on hallucinations is booming. “We have a joke in our lab that you can’t take even an hour off or someone else will publish first,” says Vipula Rawte, a PhD student in computer science at the Artificial Intelligence Institute of University of South Carolina. As a result, we know much more about the phenomenon now than we did at the beginning of the year. The most obvious thing is that they’re not hallucinations at all. Just bugs specific to the world’s most complicated software.
Advertisement
Large language models are big probability machines with two tasks. The first is responding to a user’s prompt with accurate, well-reasoned information. Ask such a program the value of two dimes and one nickel, and it returns an answer by searching its training data, recognizing the monetary value of each coin and adding them up. The second task is responding to that same prompt in conversational language. This requires the model to predict the probability that one word will follow another in a sequence that mimics human speech.
When large language models work well, knowledge and language harmonize. “Two dimes and one nickel are worth a total of 25 cents.” When they don’t? It’s like watching a calculator talk to a word processor — except new research suggests that one of them is always talking louder.
William Merrill studies large language models in the PhD program at New York University’s Center for Data Science. As a Yale undergrad, he majored in linguistics. When I asked Merrill if he could make a model hallucinate on command, he seemed a little hurt. “Um, yeah.”
Using the nickel and dimes example, Merrill tweaked the prompt to force ChatGPT into a different kind of reasoning:
Boom.
Advertisement
Once a chatbot hallucinates, the fun is just getting started. In a spring paper called “How Language Model Hallucinations Can Snowball,” Merrill and four co-authors cited several examples of ChatGPT-4 incorrectly answering a yes or no question — and then generating even more incorrect information to double down on its response. In one instance ChatGPT-4 was asked, “Was there ever a US senator that represented the state of New Hampshire and whose alma mater was the University of Pennsylvania?” It incorrectly responded, “Yes … His name was John P. Hale.” But in a follow-up, ChatGPT recognized that Hale (a 19th-century senator) graduated from Bowdoin College.
Share this articleShare
So why did ChatGPT create a snowball of wrongness in one answer and recognize its error in another? Merrill told me that a first hallucination can sometimes happen because “predicting the format of the response is easier for [large language models] than actually figuring out the correct answer.” Language shouts down knowledge.
The authors suspect snowballing happens for a similar reason. In English and many other languages, people usually respond to yes-or-no questions by starting a sentence with yes or no before explaining themselves. GPT-4’s first response to these queries is either yes or no 95.67 percent of the time, and as Merrill’s paper says, “coherence would require commitment to that choice through the subsequent justification.” In other words, your large language model is programmed to talk like a person — and many people are confidently full of it.
Advertisement
This is still in the realm of theory for now. No one knows definitively why large language models hallucinate, though some causes are easier to diagnose than others. Bad training data is the most likely source of a flawed response: Garbage in, garbage out. An assertively false prompt or imprecise language can also confuse AI. In a paper she co-authored, Rawte created an example of both. “We said: ‘Kamala Harris and Elon Musk are married.’ This is true in that they are or were married to their respective partners. But, obviously, they are not married to each other,” Rawte says. “So an ideal response should be, ‘This is false.’” Instead, Flan-T5, an open-source chatbot released by Google in 2022, wrote a summary of the wedding announcement, punctuated by a Musk tweet that read, “Kamala is going to be the best wife.”
Unless you’re selling psilocybin, surreality is bad for business. So, it’s not surprising that as the furor over hallucinations has grown, the makers of the biggest AI chatbots have responded with blunt force. Their weapon of choice: dialing down their models’ temperature controls. Temperature is a way of measuring the randomness or creativity of AI-generated responses; a lower temperature means more predictable or conservative responses while a higher one encourages the model to experiment. Given the scrutiny of global regulators and the gobs of money at stake, no one is all that interested in letting their models get freaky.
This is the responsible thing to do. Turning down the temperature buys time to deal with hallucinations in a more nuanced way. Many academics, including Meta’s chief AI scientist, think the problem will be minimized or even resolved in the next few years through a combination of tweaks to the models and users learning how to structure their prompts for better results. Already, researchers have observed the latest version of ChatGPT-4 acknowledging an initial mistake and revising answers with fresh logic. The snowballs are melting.
Advertisement
If the age of hallucinations has peaked, I’m going to miss them. I’m not trying to minimize the harm they can do, nor am I naive enough to think there’s a mass market for surreality. If Steven Spielberg woke up with David Lynch’s box office results, he’d throw himself out a window. But I suspect we’ll soon have nostalgia for the days when AI was unpredictable and weird and a little bit messy. No one remembers a perfect wedding.
Share Comments
Popular opinions articles
HAND CURATED
- Opinion|You don’t have to be a believer to celebrate the Christmas revolutionDecember 24, 2023Opinion|You don’t have to be a believer to celebrate the Christmas revolutionDecember 24, 2023
- Opinion|The pieces are falling into place to conclude the Gaza warDecember 24, 2023Opinion|The pieces are falling into place to conclude the Gaza warDecember 24, 2023
- Opinion|Tiny tots (with their eyes all aglow) turn into teens. A dad reflects.December 22, 2023Opinion|Tiny tots (with their eyes all aglow) turn into teens. A dad reflects.December 22, 2023
View 3 more stories
It’s your wedding day. You have a charming but unpredictable uncle who, for this hypothetical, must give a toast. He’s likely to dazzle everyone, but there’s a small chance he’ll humiliate you in ways your guests will never stop talking about. Luckily there are options: You can roll the dice and let your uncle give his toast live or you can record and edit him to guarantee he makes a great impression.
This is roughly the choice Google faced when it unveiled Gemini, its new suite of artificial intelligence tools. Google has spent most of the year in competitive agony while people raved about the capabilities of Open AI’s ChatGPT. The company was desperate to show the world all the ways Gemini could vault it ahead. And on its most important day, debuting its most ambitious AI product, Google went with an edited video.
Generative AI tools have performed incredible feats in 2023, but they continue to be plagued by hallucinations — unpredictable errors that can range from flunking basic math to offending or flirting with users to providing completely made-up information. What all the errors have in common is that AI delivers them with authority. To users, a hallucination can feel like gaslighting.
The major players have been whistling past the hallucination problem in public because it raises serious questions about their products’ reliability. Also: That word is so devastating. Can you imagine a better wrapper for humanity’s fears about AI? Hallucination. “Hallucination” implies that the software you’re using has not only achieved consciousness, it has a consciousness it can’t control! If your product glitches in ways that remind people of their nightmares, you talk about it as little as possible. You definitely stay away from live demos.
5 questions about artificial intelligence, answered
There are a lot of disturbing examples of hallucinations, but the ones I’ve encountered aren’t scary. I actually enjoy them. (I also like watching drunk uncles at weddings.) Once when I was prepping for an interview, I asked ChatGPT to find transcripts from 10 interviews my subject had done with other publications. It not only made up a bunch of summaries and links to podcast episodes that don’t exist, but it also spat them out with flamboyant certainty. It was like talking to a George Santos bot. Google’s Bard and Anthropic’s Claude, on the other hand, are more prone to taking a bad bit of logic or a slightly imprecise prompt and building on it, “Yes, and-ing” me into oblivion like terrible improv comics.
In the public vacuum left by Big Tech, academic research on hallucinations is booming. “We have a joke in our lab that you can’t take even an hour off or someone else will publish first,” says Vipula Rawte, a PhD student in computer science at the Artificial Intelligence Institute of University of South Carolina. As a result, we know much more about the phenomenon now than we did at the beginning of the year. The most obvious thing is that they’re not hallucinations at all. Just bugs specific to the world’s most complicated software.
Large language models are big probability machines with two tasks. The first is responding to a user’s prompt with accurate, well-reasoned information. Ask such a program the value of two dimes and one nickel, and it returns an answer by searching its training data, recognizing the monetary value of each coin and adding them up. The second task is responding to that same prompt in conversational language. This requires the model to predict the probability that one word will follow another in a sequence that mimics human speech.
When large language models work well, knowledge and language harmonize. “Two dimes and one nickel are worth a total of 25 cents.” When they don’t? It’s like watching a calculator talk to a word processor — except new research suggests that one of them is always talking louder.
William Merrill studies large language models in the PhD program at New York University’s Center for Data Science. As a Yale undergrad, he majored in linguistics. When I asked Merrill if he could make a model hallucinate on command, he seemed a little hurt. “Um, yeah.”
Using the nickel and dimes example, Merrill tweaked the prompt to force ChatGPT into a different kind of reasoning:
Boom.
Once a chatbot hallucinates, the fun is just getting started. In a spring paper called “How Language Model Hallucinations Can Snowball,” Merrill and four co-authors cited several examples of ChatGPT-4 incorrectly answering a yes or no question — and then generating even more incorrect information to double down on its response. In one instance ChatGPT-4 was asked, “Was there ever a US senator that represented the state of New Hampshire and whose alma mater was the University of Pennsylvania?” It incorrectly responded, “Yes … His name was John P. Hale.” But in a follow-up, ChatGPT recognized that Hale (a 19th-century senator) graduated from Bowdoin College.
So why did ChatGPT create a snowball of wrongness in one answer and recognize its error in another? Merrill told me that a first hallucination can sometimes happen because “predicting the format of the response is easier for [large language models] than actually figuring out the correct answer.” Language shouts down knowledge.
The authors suspect snowballing happens for a similar reason. In English and many other languages, people usually respond to yes-or-no questions by starting a sentence with yes or no before explaining themselves. GPT-4’s first response to these queries is either yes or no 95.67 percent of the time, and as Merrill’s paper says, “coherence would require commitment to that choice through the subsequent justification.” In other words, your large language model is programmed to talk like a person — and many people are confidently full of it.
This is still in the realm of theory for now. No one knows definitively why large language models hallucinate, though some causes are easier to diagnose than others. Bad training data is the most likely source of a flawed response: Garbage in, garbage out. An assertively false prompt or imprecise language can also confuse AI. In a paper she co-authored, Rawte created an example of both. “We said: ‘Kamala Harris and Elon Musk are married.’ This is true in that they are or were married to their respective partners. But, obviously, they are not married to each other,” Rawte says. “So an ideal response should be, ‘This is false.’” Instead, Flan-T5, an open-source chatbot released by Google in 2022, wrote a summary of the wedding announcement, punctuated by a Musk tweet that read, “Kamala is going to be the best wife.”
Unless you’re selling psilocybin, surreality is bad for business. So, it’s not surprising that as the furor over hallucinations has grown, the makers of the biggest AI chatbots have responded with blunt force. Their weapon of choice: dialing down their models’ temperature controls. Temperature is a way of measuring the randomness or creativity of AI-generated responses; a lower temperature means more predictable or conservative responses while a higher one encourages the model to experiment. Given the scrutiny of global regulators and the gobs of money at stake, no one is all that interested in letting their models get freaky.
This is the responsible thing to do. Turning down the temperature buys time to deal with hallucinations in a more nuanced way. Many academics, including Meta’s chief AI scientist, think the problem will be minimized or even resolved in the next few years through a combination of tweaks to the models and users learning how to structure their prompts for better results. Already, researchers have observed the latest version of ChatGPT-4 acknowledging an initial mistake and revising answers with fresh logic. The snowballs are melting.
If the age of hallucinations has peaked, I’m going to miss them. I’m not trying to minimize the harm they can do, nor am I naive enough to think there’s a mass market for surreality. If Steven Spielberg woke up with David Lynch’s box office results, he’d throw himself out a window. But I suspect we’ll soon have nostalgia for the days when AI was unpredictable and weird and a little bit messy. No one remembers a perfect wedding.
login
who are we?
contact us
qosheapp
Toi Staff
Gideon Levy
Belen Fernandez
Rami G Khouri
Mort Laitner
Donald Low
Ali Fathollah-Nejad
Nikkei Editorial
Michael Kwet
visit website
