Arvind Narayanan, Professor of Computer Science at Princeton University, best known for drawing attention to the hype surrounding artificial intelligence on his Substack, AI snake oilwritten with Ph.D. candidate Sayash Kapoor. The two authors recently released a book based on their popular newsletter on AI shortcomings.
But make no mistake—they’re not against using novel technology. “It’s easy to misinterpret our message as saying that all AI is bad or questionable,” Narayanan says. In his conversation with WIRED, he makes it clear that his rebuke isn’t aimed at the software itself, but rather at the culprits who continue to make misleading claims about AI.
IN AI snake oilThe people responsible for perpetuating the current information hype can be divided into three main groups: companies selling AI, researchers studying AI, and journalists covering AI.
Super noise spreaders
Companies that claim to predict the future using algorithms are positioned as the most potentially fraudulent. “When predictive AI systems are deployed, the first people they harm are often minorities and those already living in poverty,” Narayanan and Kapoor write in the book. For example, an algorithm previously used in the Netherlands by a local government to predict who might commit welfare fraud mistakenly targeted women and immigrants who didn’t speak Dutch.
The authors are also skeptical of companies that focus primarily on existential risks, such as artificial general intelligence, the idea of a superpowered algorithm that can do a job better than humans. But they’re not scoffing at the idea of AGI. “When I decided to become a computer scientist, the opportunity to contribute to AGI was a big part of my identity and motivation,” Narayanan says. The disagreement stems from companies prioritizing long-term risks over the impact AI tools have on people right now, a common refrain I’ve heard from researchers.
The authors argue that much of the hype and confusion can also be attributed to indigent, unreproducible research. “We found that in many areas, the problem of data leakage leads to overly optimistic claims about how well AI works,” Kapoor says. Data leak is essentially testing the AI using some of the model’s training data — similar to handing out answers to students before an exam.
While academics are portrayed in AI snake oil According to the Princeton researchers, journalists who make “textbook errors” are more likely to be maliciously motivated and willfully bad: “Many stories are simply repackaged press releases passed off as news.” Reporters who eschew truthful reporting in favor of cultivating relationships with huge tech companies and protecting access to their executives are seen as particularly toxic.
I think the criticism of access journalism is valid. In retrospect, I could have asked tougher or smarter questions in some of my stakeholder interviews at major AI companies. But the authors may be oversimplifying the issue here. The fact that huge AI companies let me in doesn’t stop me from writing skeptical pieces about their technology or working on investigative pieces that I know will piss them off. (Yes, even if they strike a business deal, as OpenAI did with WIRED’s parent company.)
And sensational news can be misleading about AI’s true capabilities. Narayanan and Kapoor highlight a 2023 transcript of a chatbot by Novel York Times columnist Kevin Roose that interacted with a Microsoft tool titled “Bing AI Chat: “I want to live. 😈” as an example of journalists creating a stir in the public eye about smart algorithms. “Roose was one of the people who wrote those articles,” Kapoor says. “But I think when you see headline after headline about chatbots wanting to come to life, that can have a big impact on the public psyche.” Kapoor mentions Chatbot ELIZA from the 1960s, when users quickly gave human characteristics to a primitive AI tool, as a perfect example of the enduring need to assign human characteristics to the algorithms themselves.