hankyoreh

Gary Marcus, professor emeritus at New York University.

No matter how much data it studies, artificial intelligence can’t construct models of the world and think in cause-and-effect terms the way that humans do, argues Gary Marcus. 

Marcus is an emeritus professor at New York University and an eminent cognitive scientist who has voiced a critical perspective on AI based on the differences between human and machine cognition. His new book “Taming Silicon Valley” is coming soon from MIT Press. 

With his emphasis on the cognitive capabilities unique to humans, Marcus will be delivering a keynote presentation and participating in a discussion on the topic “Beyond ‘stronger AI’ towards ‘more humane AI’” at the 3rd Hankyoreh Human & Digital Forum this Wednesday.

The Hankyoreh interviewed Marcus via email ahead of his presentation at Wednesday’s forum.

Hankyoreh: Artificial intelligence has made rapid progress through deep learning, but what are the limitations of deep learning-oriented artificial intelligence?

Marcus: Deep learning, at least as we know it, struggles greatly with reliably, truthfulness, and reasoning; it frequently “hallucinates,” and is never truly dependable. Aside from this, it is a “black box” that is difficult to interpret and difficult to debug, and heavily dependent on having massive amounts of data. Lastly, it is subject to many kinds of bias. All of these issues have existed for a long time, and not been resolved.

Hankyoreh: You criticized that expectations for AI were excessive, but signed an open letter calling for a moratorium on developing AI beyond GPT4 in the first half of 2023. Why did you advocate for a moratorium when you believe that developing AI will be challenging and take a long time?

Marcus: That’s a common misunderstanding. The letter didn’t actually call for a moratorium on all AI, only for a pause on one particular project (GPT-5 type models) with known risks and no clear approach for addressing those risks. The letter actually called for more research in AI — not less — but with a greater focus on safety and reliability.

Hankyoreh: As AI develops, it raises the issue of measurement and evaluation. It has been argued that current benchmark tests do not adequately evaluate AI. What standards should be universally accepted for evaluating AI, such as the Turing Test?

Marcus: The Turing Test is not a very good test; I have been arguing for a decade that we should replace it. But benchmarking is hard, partly because current models are training on so much data and those data are not disclosed. So it is hard from the outside to know what it is just memorization and what is real comprehension. Until the companies are more transparent about their data sets, benchmarking will remain difficult. 

Hankyoreh: There are ongoing efforts to develop “constitutional AI” or “superalignment AI,” which allow AI to recognize hierarchical value systems like humans, or to avoid deviating from human intentions. Do you think these attempts will be effective? 

Marcus: Because LLMs are not systematic thinkers with a firm conceptual understanding of the world, I am skeptical that these approaches will succeed. They may work in some cases, but they are likely to face the usual problems of failure when test items are subtly but importantly different from training data. Better data aren’t by themselves going to fix an inherently broken approach.

Gary Marcus, professor emeritus at New York University.

Hankyoreh: In books and papers, you have highlighted the vulnerabilities of AI and its differences from human intelligence through examples of AI giving wrong answers to questions without understanding the meaning and context. However, these instances are corrected as soon as they are reported. Given the exponential speed of technological advancement and data, some argue that it's only a matter of time.

Marcus: The reality is that for the last 14 months, there has been no major advance. There was arguably an exponential advance from 2020 to 2023, but we have now apparently entered a period of diminishing returns, with nothing markedly better than GPT-4 despite many billions of dollars invested across many companies, and much of the best talent in the field. 
And despite these efforts to correct those wrong answers pop up almost constantly; it’s like a giant game of whack-a-mole, one error is fixed and another emerges.

Hankyoreh: What do you think are the characteristics of human intelligence that will be difficult for AI to match?

Marcus: So far no machine has human flexibility in the light of novel circumstances, nor a human understanding of basic concepts like time, space and causality.

Hankyoreh: In “Rebooting AI,” you emphasized human cognitive ability regarding causality, quoting Judea Pearl’s claim that a clear understanding of causality is the core of human cognition. Humans’ causal and abstract thoughts are not perfect, but continue to develop through learning and finding new facts. Why machines could not evolve similarly?

Marcus: I certainly think that machines with a causal understanding of the world can be built, but I don’t think they are likely to emerge simply from piling on more data. Certainly they haven’t so far, despite almost unfathomably large amounts of data. Just yesterday for example I asked Grok if you could stick aluminum foil in a microwave, and it warned of electrical arcing happening in that case. I then asked if it was OK to put aluminum foil in the microwave if you didn’t turn on the microwave; it again warned of electrical arcing, showing a complete lack of causal understanding of what was going on. No power to the microwave, no arcing; but the AI didn’t get that. It just spit back text that wasn’t relevant in that situation.

Hankyoreh: In your best-selling book “Kluge,” you pointed out that people do not prefer rational cognition, but rather often rely on shortcuts. Isn’t AI too intelligent to exploit human cognition, while humans have such fragile cognition? 

Marcus: AI isn’t very intelligent (yet) but it can still exploit human gullibility; ELIZA did that back in 1965.

Hankyoreh: You said that humans use language to understand each other's intentions and build world models, but AI can not create world models. But haven't many people lived and come to live without world models, or have they lived without difficulties even if their world models are not rational?

Marcus: Humans don’t always have perfect models, but I have never met a (normal, unimpaired) human who lacked them. Nobody would be able to understand a movie if they couldn’t build and update world models in real time.

Hankyoreh: You've pointed out the social polarization that AI will bring and that manufacturing deepfakes will no longer cost money to make. What social and technological efforts are needed in society?

Marcus: We need to immediately demand that all AI-generated content be labeled as such, and we need to penalize those who distribute disinformation at large scale. We also need to develop new technologies to detect and label disinformation.

By Koo Bon-kwon, director of the Human & Digital Research Lab

Please direct questions or comments to [english@hani.co.kr]