Nobel says AI collects data that is ‘garbage’ and threatens all the world’s information
AI won’t just reshape work and markets; Joseph Stiglitz says it can also silently erode the information these systems depend on. As great language models (LLMs) sift through our snarky comments on Reddit and the loudest fringe voices on extremist forums, the Nobel laureate warns of a world in which everything seems more data-driven, but the underlying data is increasingly “junk.”
“In the case of AI, I think there are some other deeper problems,” the economist told Fortune. “Not only do we have a problem in the job market… but there is another side to what I would call information externalities,” which Stiglitz describes simply as garbage in, garbage out.
Also read: YouTube releases deepfake detection for Brazilian politicians. Understand how it works
The risk is not just job losses; it is a broken feedback loop between truth and the systems we use to interpret reality — from prediction markets to financial models to political debate. In essence, AI is only as smart as the data it receives, and when it continues to collect unreliable information, the result becomes as distorted as the information absorbed.
In his view, current models are built on a flawed compromise: They voraciously collect journalism, research and conversations online while undermining the very institutions that produce high-quality knowledge.
The result, he fears, is a world in which people are driven by the online rhetoric they see being replicated by AI — think of the market crash triggered by a Citrini Research report touting a “phantom GDP” or Matt Shumer’s apocalyptic viral AI essay — rather than by anything anchored in reality.
AI is “stealing information” from the sources it needs
Stiglitz would like to thank you for reading this article, even though his starting point is direct and blunt. “AI is basically stealing information from traditional media,” he said, “and that means traditional media doesn’t have the resources or incentives to produce information.”
It’s true that some AI companies pay for certain journalistic content. OpenAI, for example, has a content deal with News Corp, owner of the Wall Street Journal.
Still, Stiglitz said, AI has neither the interest nor the capacity to produce new quality information. “And the result of all this is that there is a real risk of deterioration of the general information ecosystem.”
If the best sources of information are slowly “defunded” while the cheapest forms—like comment sections, partisan memes, and low-effort content—proliferate, training data will skew toward what is most abundant and cheapest, meaning chatbots will tend to repeat what they absorb from online forums.
This is the first way that AI’s hunger for what’s online could backfire: by cannibalizing the business models that support serious work and changing the makeup of what’s there to be collected in the first place.
Garbage in, garbage out on an industrial scale
Stiglitz, who mentions the information ecosystem in his 2024 book, “The Road to Freedom: Economics and the Good Society”, again resorted to the cliché “garbage in, garbage out”. “If you’re processing and disseminating waste, all you get in the end is garbage — garbage in, garbage out.”
The expression may be old, but Stiglitz says it remains very relevant. AI systems are excellent at processing what we give them, but they are nowhere near as good at distinguishing knowledge from noise.
“There is a real risk that, despite the potential of new technologies to improve the information ecosystem in critical areas, we could end up in a worse situation,” he said. The more garbage that comes in — unverified claims, conspiracy theories, artificial opinion campaigns, low-quality comments — the more refined garbage comes out.
He worries that users will mistake this veneer for truth. “They will think they received highly processed information without fully realizing the extent to which all they did was reprocess waste,” he said. “Processing junk with AI is no substitute for a single good academic paper.”
When anti-vaccine activists outnumber scientists
Nowhere is this risk clearer than in the far corners of the internet, where radical views are often the loudest.
Think of a typical community forum on a given topic. Thanks to the anonymity of the internet, users feel free to express their opinions on the latest political decision or cultural event.
As a result, these spaces are environments where misinformation is most abundant — and the science that debunks it receives little attention, if at all. Vaccines are a perfect case study, says Stiglitz.
“Anti-vaccine activists are much more active on the internet than people who say vaccines work,” he said. Scientists run tests, publish some dense papers, and move on. Conspiracy theorists flood forums and social media every day.
“So there may be a lot more writing on the contra side than that one crucial article that says, ‘Here’s the vaccine trial, and it works. Here’s the efficacy,'” Stiglitz explained. “Do today’s AIs have the ability to say that a single article is all we need? They don’t.”
For models trained based on raw frequency and engagement, the loudest voices win. AI’s hunger for more information can distort reality by boosting the passionate minority above the cautious majority, especially in areas where the public interest depends on trust in slow, methodical science.
Prediction markets based on lack of information
In a 1980 paper with Sanford Grossman, Stiglitz argued that there is a paradox at the heart of efficient markets: if prices fully reflect all available information, then no one has an incentive to pay to collect that information, so the very information that makes markets “efficient” disappears.
He says AI and modern prediction markets are repeating this story on a larger scale. “It’s interesting that you mention Grossman-Stiglitz,” he told Fortune, “because I wrote a paper with one of my graduate students, Max Ventura, extending the Grossman-Stiglitz model to AI, and the result I described earlier about how we can make the information ecosystem worse was actually a reference to that extension.”
When “you don’t force the AI companies that collect data from Fortune and all the other media producers” to pay for what they take, “they don’t get a return, and so the incentives to do the quality basic research that leads to a good information ecosystem are weakened.” Prediction markets and trading algorithms then start to rely on the outputs of these models, further moving their bets away from any real investment in reality.
“This has weakened the incentives to produce high-quality information, increased the capacity to produce low-quality information, and therefore there is more garbage coming in and more garbage going out,” he said. A system designed to aggregate knowledge ends up amplifying what is cheapest and most abundant.
AI as support, not as an oracle
Despite all this, Stiglitz doesn’t think the answer is to ban or ignore AI. He uses it himself and tries to teach his students to do the same — without confusing an elegant response with a solid argument.
“We try to teach them how to use AI as a research tool,” he said. “So, you know, we’re not abandoning AI. I use AI as part of my research. It’s an incredible research tool, but it’s not a replacement for thinking or analysis.
“It can help find sources, develop ideas,” he added. “But in the end, you have to do the hard work.” For him, the results of a model are “actually supports for me to start thinking about things perhaps a little differently”, not conclusions to be accepted without question.
Still, he believes that there must be some intervention at the government level to prevent the deterioration of information from worsening. “In the absence of government regulation,” he warned, “there is at least a significant risk that we will end up with a worse information ecosystem in several areas of concern.”
