We know remarkably little about how AI language models work

AI language models are not humans, and yet we evaluate them as if they were, using tests like the bar exam or the United States Medical Licensing Examination.

The models tend to do really well in these exams, probably because examples of such exams are abundant in the models’ training data. Yet, as my colleague Will Douglas Heaven writes in his most recent article, “some people are dazzled by what they see as glimmers of human-like intelligence; others aren’t convinced one bit.” 

A growing number of experts have called for these tests to be ditched, saying they boost AI hype and create “the illusion that [AI language models] have greater capabilities than what truly exists.” Read the full story here. 

What stood out to me in Will’s story is that we know remarkably little about how AI language models work and why they generate the things they do. With these tests, we’re trying to measure and glorify their “intelligence” based on their outputs, without fully understanding how they function under the hood. 

Other highlights:

Our tendency to anthropomorphize makes this messy: “People have been giving human intelligence tests—IQ tests and so on—to machines since the very beginning of AI,” says Melanie Mitchell, an artificial-intelligence researcher at the Santa Fe Institute in New Mexico. “The issue throughout has been what it means when you test a machine like this. It doesn’t mean the same thing that it means for a human.”

Kids vs GPT-3: Researchers at the University of California, Los Angeles gave GPT-3 a story about a magical genie transferring jewels between two bottles and then asked it how to transfer gumballs from one bowl to another, using objects such as a posterboard and a cardboard tube. The idea is that the story hints at ways to solve the problem. GPT-3 proposed elaborate but mechanically nonsensical solutions. “This is the sort of thing that children can easily solve,” says Taylor Webb, one of the researchers. 

AI language models are not humans: “With large language models producing text that seems so human-like, it is tempting to assume that human psychology tests will be useful for evaluating them. But that’s not true: human psychology tests rely on many assumptions that may not hold for large language models,” says Laura Weidinger, a senior research scientist at Google DeepMind. 

Lessons from the animal kingdom: Lucy Cheke, a psychologist at the University of Cambridge, UK, suggests AI researchers could adapt techniques used to study animals, which have been developed to avoid jumping to conclusions based on human bias.

Nobody knows how language models work: “I think that the fundamental problem is that we keep focusing on test results rather than how you pass the tests,” says Tomer Ullman, a cognitive scientist at Harvard University. 

Read the full story here. 

Deeper Learning

Google DeepMind has launched a watermarking tool for AI-generated images

Google DeepMind has launched a new watermarking tool that labels whether images have been generated with AI. The tool, called SynthID, will initially be available only to users of Google’s AI image generator Imagen.  Users will be able to generate images and then choose whether to add a watermark or not. The hope is that it could help people tell when AI-generated content is being passed off as real, or protect copyright. 

Baby steps: Google DeepMind is now the first Big Tech company to publicly launch such a tool, following a voluntary pledge with the White House to develop responsible AI. Watermarking—a technique where you hide a signal in a piece of text or an image to identify it as AI-generated—has become one of the most popular ideas proposed to curb such harms. It’s a good start, but watermarks alone won’t create more trust online. Read more from me here.

Bits and Bytes

Chinese ChatGPT alternatives just got approved for the general public
Baidu, one of China’s leading artificial-intelligence companies, has announced it will open up access to its ChatGPT-like large language model, Ernie Bot, to the general public. Our reporter Zeyi Yang looks at what this means for Chinese internet users. (MIT Technology Review)

Brain implants helped create a digital avatar of a stroke survivor’s face
Incredible news. Two papers in Nature show major advancements in the effort to translate brain activity into speech. Researchers managed to help women who had lost their ability to speak communicate again with the help of a brain implant, AI algorithms and digital avatars. (MIT Technology Review)

Inside the AI porn marketplace where everything and everyone is for sale 
This was an excellent investigation looking at how the generative AI boom has created a seedy marketplace for deepfake porn. Completely predictable and frustrating how little we have done to prevent real-life harms like nonconsensual deepfake pornogrpahy. (404 Media) 

An army of overseas workers in “digital sweatshops” power the AI boom
Millions of people working in the Philippines work as data annotators for data company Scale AI. But as this investigation into the questionable labor conditions shows, many workers are earning below the minimum wage and have had payments delayed, reduced or canceled.
(The Washington Post) 

The tropical Island with the hot domain name
Lol. The AI boom has meant Anguilla has hit the jackpot with its .ai domain name. The country is expected to make millions this year from companies wanting the buzzy domain name. (Bloomberg)

P.S. We’re hiring! 
MIT Technology Review is looking for an ambitious AI reporter to join our team with an emphasis on the intersection of hardware and AI. This position is based in Cambridge, Massachusetts. Sounds like you, or someone you know? Read more here.