Programming note: Usually Mondays are for Big Optimism. This week, BigOpt will run Tuesday.
Explanation: It was late Sunday and Big Optimism wasn’t quite finished — so I switched the order, and moved Tuesday’s to Monday.
Welcome to the world of one-person media companies. Thanks for reading!
Welcome to ‘Knowledge Collapase’
Over the summer, I had the moment when I realized that AI had made its way into everybody’s life in America.
I was waiting behind a man in his 80s while we were on vacation on Cape Cod, and I heard him trying to make plans for the next day by simply asking ChatGPT what he should do tomorrow.
I thought: Wow. O.K. That’s it.
People compare the rise of AI to the rise of the Internet. But this feels much faster.
So, I paid attention when I saw a new research paper about what happens when people use AI chatbots to seek information.
The researchers were looking at what they call “epistemic diversity.” This is is about how broad a range of information and claims people encounter when they ask an AI system about a subject.
Dustin Wright, a tenure-track assistant professor at Aalborg University in Copenhagen and the paper’s lead author, explained to me that a few years ago, the dominant way people sought information was to go to Google and get a collection of web pages.
(My interview with Wright is embedded at the bottom of this article.)
That was “all written by humans for the most part and all potentially having different perspectives and different ways of looking at whatever it is your query was about,” but now generative AI is increasingly replacing that.
So Wright and the other researchers set out to do a massive assessment of what people search for, what chatbots give them as answers. Then they compared it to what they would have gotten from the kind of Google searches that we all did a few years back.
They tested 27 different AI systems on 155 topics, using 200 prompt formulations. That produced about 1.7 million responses and, after the researchers broke those responses down, roughly 70 million individual claims.
The topics ranged from nuclear weapons and pornography to democracy, free speech, the Berlin Wall and K-pop. Basic question: Was the information produced by the AI systems was as diverse as the information people would encounter through traditional search.
“There was not a single model that was as diverse as just doing a Google search,” Wright told me.
The researchers found that traditional Google search produced at least 18 percent more diverse information. Wright said that’s how much they say confidently, but they had reason to think the real gap might be larger.
The researchers call the broader concern “knowledge collapse.”
If generative AI systems come to dominate the information space, and the amount of information people encounter through them is narrower, over time we could lose access to certain kinds of information.
It’s what Wright described as “long tail information,” “minoritized information” or culture-specific information. And it’s not well-represented by generative AI systems.
This is the part that concerns me.
Let’s imagine that Understandably came up in search after search after search. I mean, I’d love it. But I’d understand if people assessed for themselves whether they think this newsletter is a reliable source of information.
The point is: they could see that Understandably was the source, and see who wrote it. They could decide whether they generally trust the publication or the writer behind it (hi!).
But when AI itself is generating the answer, in many cases there’s no source even given. Thus, there’s no way of assessing, or even knowing, that the pool of information from which we’re drawing has shrunk.
As Wright put it, sure Google search is already curated and biased. Some websites rank higher than others.
“But at least,” he said, “if you look in different communities, you get a lot of human authored text with different perspectives, offering different information.”
As AI-generated text increasingly takes over that role, he said, “we’re potentially losing a lot of that.”
Now, I don’t mean to put another log on the fire of the AI doomscape. There was actually some encouraging news in the research.
The models have become more diverse over time. Wright told me that he had expected diversity either to decrease or stay stagnant, but instead it increased, especially between GPT-4 and GPT-5.
Still, I don’t know what the answer is.
On the one hand, more people are probably accessing more information than in previous generations.
On the other hand, the natural progression could be toward a smaller scope of information. It’s still massive, obviously, but smaller compared with what’s available in the world.
Look, I couldn’t have told you in 1999 what 2004 was going to look like. I couldn’t have told you in 2004 what 2010 would look like.
In 2010, I couldn’t have told you what 2020 would look like. And in 2020, I sure as heck could not have predicted where we are now.
So, I feel like the bare minimum is to pay close attention. And maybe wonder.
Imagine it’s another 40 or 50 years from now, and I’m the old man on Cape Cod, casually using a new technology most of us can’t even imagine today.
And maybe somebody will think: Wow. O.K. That’s it.
Here’s the video of our interview:
Other things worth knowing …
AP: President Trump signed an executive order late Friday suspending new sanctions targeting Russia, citing national security, but providing no further explanation to the American people. The move comes a week after Trump stunningly said the U.S. had agreed to purchase diesel from Russia to try to bring down global energy prices, despite that potentially hurting Ukraine in its war with Russia.
PBS: French President Emmanuel Macron on Sunday voiced the public’s “anger” and “shame” over rioting and looting during nationwide protests against rising living costs and poor conditions in schools. At least 2,400 people were arrested. “Those who looted will have to pay for the damage, and those who sought to bring down the Republic will be held accountable,” Macron said in a televised address to the nation.
Yahoo News: President Trump’s embattled new press secretary Katie Zacharia was forced out Friday after less than a week on the job. Zacharia is the shortest-tenured White House press secretary in U.S. history.
Bloomberg via Yahoo Finance: More than 1 million people have already cast their ballots in Alberta’s referendum on separation from Canada, signaling strong voter turnout that pollsters say may favor the “remain” side. The official voting day is today. Public opinion surveys have suggested the “remain” side is well ahead, including a recent poll that found 70% of Albertans wish to stay in Canada. But pro-Canada campaigners have pushed to get people to voting centers, fearing that complacency and low turnout would benefit the separatist cause.
CNBC: Always on, always listening: New AI devices are pushing legal limits and could put users in jeopardy.
NBC News: In the new Elon Musk documentary, his lieutenants are the most damning voices. “Musk,” the new film from Alex Gibney, features interviews with more than the usual critics of the Tesla and SpaceX CEO.
Forbes: We are running out of ways to describe just how critically beloved You Can See Everything is, the mind-bending Elizabeth Holmes pseudo-documentary from Nathan Fielder. On Rotten Tomatoes, the website issued a statement that said the film was the only 100% certified fresh movie of the year, having more than 80 reviews in. Now, it’s dropped to a 99% because of a single negative review, but the 99% with 101 reviews still seems to be the record for the year regardless.
Thanks for reading. Photo by Clay Banks on Unsplash. I wrote about some of this at Inc.com. See you in the comments!