Brendan Fleig-Goldstein is a philosophy professor and Mellon Postdoctoral Fellow at Brown University’s Cogut Institute for the Humanities. His research explores the philosophical foundations of cognitive science and artificial intelligence. In particular, he studies questions surrounding how well cognitive systems make use of limited resources, how cognitive limitations shape epistemic norms, and how irrationality can serve as evidence for cognitive models. He was previously the Director of Cognitive Science at Columbia University and is a coordinator for ALIUS, an international research group on consciousness. He holds a PhD in History and Philosophy of Science from the University of Pittsburgh, an MS in Symbolic Systems from Stanford University, and a BA from Tufts University with a double major in Philosophy and Cognitive & Brain Science and a minor in Religion.
Samrah Syed: There’s a lot of confusion about what large language models (LLMs) actually are. At a basic level, how do they work?
Brendan Fleig-Goldstein: At a high level, what they’re doing is taking natural language input, cutting it up into a bunch of smaller units called tokens, and then mapping those tokens into vectors, which are just a list of numbers. Then, those vectors get passed through successive layers where they are transformed and the model ultimately produces a probability distribution over possible next tokens. So what we really train LLMs to do is next-token prediction. When you use something like a chatbot, it’s generating text one token at a time—predicting the next word, then the next, and looping that process.
SS: A lot of people think AI is just pulling data from its large training databases and regurgitating it. Is that accurate?
BFG: Not really. That’s probably one of the biggest misconceptions about AI—that these models can’t do anything novel. Of course there’s large amounts of data that these models are trained on, but it’s not just regurgitating the statistics of their input. Current models use what are called ‘attention mechanisms,’ where they track how different words in a sentence influence each other. In a sense, you can think of it as learning which words matter more for understanding the meaning of other words in context; so the attention mechanisms essentially bake the context of words into vectors.
So if you have a phrase like “mean dog,” the model doesn’t just treat “dog” the same way every time—it adjusts its internal representation of “dog” based on the word “mean.” As it processes a sentence, it keeps updating these representations, gradually building in more and more contextual meaning. There are also other layers that capture broader patterns—things like factual associations or general world knowledge—but overall, the process is generative, not just retrieval-based.
SS: Can you give an example of how current models create patterns and come to generate text?
BFG: One interesting thing about these models and the way they “learn” is that certain patterns in meaning get encoded geometrically in that space. A classic example is that if you take the vector for “king,” subtract “man,” and add “woman,” you get something very close to “queen.” The model isn’t explicitly told concepts like gender or royalty—it learns patterns like that from data. So while the math is complex, the core idea is surprisingly simple: It’s learning relationships between words and using those relationships to generate text, one step at a time.
SS: Given how these systems work, what are they actually good at, and where do they fall short?
BFG: I like to think back to Simon and Newell’s analysis that an intelligence process is two-fold: You have to generate and then test, or you have to search and identify solutions and then decide which one’s the right one.
LLMs are pretty good at the first part. They’re great at generating ideas, alternative phrasings, or possible arguments. Where they’re weaker is in the evaluation step—figuring out whether something is actually correct or good. So you can use them to come up with some problems with your view that you can then decide whether to fix. But you shouldn’t treat them as a reliable judge of quality. That’s where people tend to go wrong.
SS: There’s a lot of debate about whether students should be using AI for things like readings or assignments. What do you think are appropriate uses of AI?
BFG: I don’t really see a problem with, say, putting in a reading and asking for a summary or clarification. If you’re confused about something, using a tool to help you understand it better seems totally reasonable. You can also use it to ask questions about a text, get alternative explanations, or even generate objections to an argument you’re working on. That kind of interaction can actually support learning. Where it becomes an issue is if it replaces the process entirely. It’s not a substitute for doing the reading or developing your own understanding.
SS: So is the concern that students are offloading the wrong kinds of tasks?
BFG: Yeah, I think that’s a good way to put it. We’ve always offloaded certain tasks to tools, and that’s not inherently a bad thing. The question is which tasks you’re offloading. If you’re offloading something like basic clarification or getting unstuck, that can be helpful. But if you’re offloading the actual thinking, like forming an argument, working through a problem, or making judgments about what’s good or true, then you’re bypassing the skill that you’re supposed to be developing. So it’s less about whether you’re using AI and more about how you’re using it.
SS: A lot of the public conversation around AI is about it “taking our jobs.” So, I have to ask…is AI stealing our jobs?
BFG: Look, one of two things are going to happen. Either AI is going to be everything that all the hypers say it’s going to be, and we’re going to get crazy super-intelligent AI in a few years, and it’s going to take all our jobs—or it’s not. And then the market, which is tied to all of this investment around AI, is going to then have a market correction. So if it doesn’t live up to the hype, then we will have a big recession, and we’re going to lose all our jobs anyways.
[I think Brendan saw the look of horror on my face and decided to give a more uplifting answer].
BFG: But, on a more optimistic note, I think there’s a lot that can be said on this topic and it’s just really hard to say right now. I think it’s less about outright replacement and more about changing what counts as a valuable skill. In a lot of fields, these tools are already being integrated into everyday work. For example, in programming, it’s increasingly common for people to use AI assistance, and some job interviews even allow it. That’s similar to how we think about calculators—you’re expected to use the tools that are available in real-world contexts. So the question shifts from “Can you do this entirely on your own?” to “Can you use these tools effectively to solve problems?” That doesn’t mean skills disappear, but it does change which skills are emphasized.
SS: Does that mean we should be less worried about AI replacing people entirely?
BFG: Yeah, I think the reality is more nuanced than that. There will probably be some displacement in certain areas, but in many cases it’s more about redistribution of tasks. AI is very good at producing outputs quickly, especially in domains that involve language or pattern recognition. But it still relies on humans to guide, interpret, and evaluate those outputs. So rather than fully replacing people, it often ends up changing the structure of work—what parts are automated and what parts still require human judgment.
At the same time, I do think there are certain kinds of skills that are much harder to displace—especially the kinds of skills emphasized in the humanities. Things like interpreting meaning, constructing arguments, thinking critically, and making nuanced judgments don’t map as cleanly onto what these systems do well. Those are still going to be valuable, and in some ways they might become even more important as these tools become more widespread.
SS: So…is it stealing our jobs? Do I even need to be here?
BFG: Well, it’s also just a very uncertain moment. Current systems might be plateauing a bit in terms of what they can do, but that’s not necessarily the end of the story. If there’s another major breakthrough—whether that’s a new approach to training these models or combining them more effectively with robotics and other systems—then you could see another wave of capabilities. And that could change the landscape again in ways that are hard to predict.
So, I think the honest answer is that we don’t really know yet. But it’s probably better to think in terms of adaptation rather than total replacement and to focus on developing skills that involve judgment, interpretation, and understanding—things that are harder to outsource.
SS: Given all of this, what’s the most important thing people should understand about AI right now?
BFG: Probably that these systems are very good at producing outputs that look convincing, but that doesn’t mean they’re reliably correct or insightful. So the responsibility is still on you as a human to evaluate what’s being produced. If you keep that in mind, LLMs can be useful tools. If you don’t, that’s where people run into trouble.
*This interview has been edited for length and clarity.