AI for Social Sciences: Generative AI Modelling beyond LLMs

Artificial intelligence is no longer confined to chatbots and text generation; it is fast becoming a powerful tool for solving complex, real-world problems. From managing investment portfolios to simulating dynamic corporate environments, today’s AI leverages goal-driven algorithms and generative models to support high-stakes decision-making across industries.

This event featured Prof Will Cong, Rudd Family Professor of Management and Professor of Finance from Cornell University, widely recognised for his leadership in financial economics. The session was moderated by Prof Byoung-Hyoun Hwang, Head of Division of Banking and Finance and Provost’s Chair in Finance at Nanyang Business School.

The following is an edited transcript of the Q&A segment:

Prof Hwang: So, you talked quite a bit about the possible benefits of AI, but some people are wondering about the possible risks or downsides. I think some of them are worried about their jobs, but I think they’re asking more broadly, whether this is going to destabilise things.

Prof Cong: That’s actually a question I’ve been thinking about for a long time. And, you know, having worked on information economics, I think the immediate, simple answer would be that with more generative AI, at least information-wise, we might face a lot of misinformation.

What if an influencer uses some AI tool? They could basically modify the content of another authentic information source and generate something that’s very entertaining and reaches a broad audience. Some of the audience members might have seen the original source. Do they treat this as an independent news source? How do they deal with the information?

That becomes quite complicated, and we’re already seeing that in the influencer or creator economy.

More importantly, and I think this is something computer scientists have thought about, in finance we’ve seen a few studies looking at something related to the idea of emergence.

We have individual AI agents or models we use. Looking at them in isolation, maybe we don’t detect any issues, but when they act together, do they lead to a phenomenon or macro-outcome that’s undesirable in financial markets?

There are already studies showing that if you use some simple reinforcement learning agents for trading, at some point they start to collude, and that’s not something we want to see in real markets.

Those types of phenomena are very hard to predict. So, I think a lot more effort is required to understand them. And pushing this even further, there’s the concern of alignment, AI alignment. How do we make sure they serve the purpose and goals of humans?

I don’t think there’s a solution. Researchers from every field are trying to provide some answer to this question. I do think that’s actually a very scary outcome we might face.

If AI agents are not aligned with our objectives, we may quickly lose control of them. One example I typically use is how we are, in some way, agents of our genes or DNA. They have certain goals in mind, reproduction, propagation of the genes, but once human agents become intelligent enough, we may not always serve that purpose.

So similarly, if AI agents reach that level, it could be potentially disastrous for humanity and society. I know it sounds a bit like science fiction, but it’s a possibility.

Prof Hwang: I mean, in medical research, when we think there’s a possible treatment or drug, we have to go through all kinds of trials and examine side effects. Only if we’re really sure that it’s safe, we go for it, but with AI, it seems like we just close our eyes and go for it, right?

Prof Cong: There’s that camp pushing for progress, the next version after the next, which could be beneficial, but compared to the existential threat, I think it makes sense to think about AI safety. In particular, I think Singapore is a great place to gather different camps, thoughts, and groups to discuss these issues.

Prof Hwang: I see -- that’s interesting. We have a bit of a naughty question. One person is asking: they felt your presentation was a bit more on the technical side, and they're asking whether they really need to know all that technical stuff in order to take advantage of AI. They said they don’t know how to use ChatGPT properly, but it just seems to work fine. What are your thoughts on that?

Prof Cong: Yeah, that’s a great question. If you’re not under any pressure to publish research and so on, then probably you don’t have to get into the nitty gritty of these details. If anything, I think the need has been reduced, right? Even for coding, we can use ChatGPT to code rather than doing it all ourselves.

I do think we don’t need to know the technical details, but it’s still helpful to have a high-level picture of what these algorithms are doing; what their potential functions are, what the issues might be. And in a way, that relates to a question I get a lot. Right, as students what should we learn in the modern digital AI age? Of course, understanding some AI tools is helpful. Maybe trying ChatGPT is fine. It’s not too difficult to utilise. More generally, I think the ability to manage AI agents, to manage a set of AI tools, is actually quite crucial.

And that’s very much tied to business schools, right? We teach people how to manage organisations, how to manage large teams. Now, maybe it’s useful to teach people how to manage a mix of human agents and AI agents, or even just AI agents. And that ties back to my earlier concern. If we’re worried about AI taking over, maybe we’re not going to yield full control to them. So, we’ll still need people to manage a set of AI agents working together to complete certain tasks. In that sense, learning managerial skills, surprisingly, becomes important again, perhaps even more so than technical skills.

Prof Hwang: That’s actually a good transition, because I think there are a lot of students online, which is great. There are also quite a few students here, and many of them are asking: If I’m very new to this topic and I want to learn more, what are some good starting points? Like, what do you tell your undergrads at Cornell?

Prof Cong: That’s a great question. I think various universities have been revising their curriculum to provide the best education for our undergrads and students. I actually don’t have a perfect or complete answer to this.

A lot of the learning materials are available online. If you want to learn machine learning, you can go to Coursera, there are basic courses. If you want to hear about capacities, there are YouTube videos talking about AI. That’s a good source. There are also blogs by OpenAI and a newsletter called, I think it’s called The Rundown AI, which gives you periodic updates.

That type of exposure and learning through osmosis is helpful. Of course, at some point, it would be better if schools developed more structured curriculum to help students.

I was actually looking into the courses at NTU, there’s one called Bachelor of Computing (Hons) in Artificial Intelligence and Society (AISC) for undergrads – a structured approach to AI education. It’s a great start.

Prof Hwang: AI systems are often seen as rational optimisers, but in the real world, people can behave irrationally. So it seems hard to use AI to predict human behaviour. Like, how do you predict how a drunk person will walk next? That’s very difficult. Is AI actually able to do that? You alluded to something on your slide 4, how does it work?

Prof Cong: In the example I was showing, behavioural bias analysis of AI agents, we did see that AIs don’t always behave like humans. There are biases, but some are machine-like rather than human-like. I think the tools are still useful. For one, you can use AI tools to study certain biases and to better identify human biases. That’s helpful. If you want AI agents to produce human-like behaviour, the training data becomes crucial. If we have structural disparities in the way we grant loans or patents, that training data may bias the AI model against certain gender or risk groups.

At the same time, we need to be careful.

AI agents themselves generate new types of biases. That’s not something that’s talked about a lot yet. So, I think these are useful exercises to do. It’s not directly answering the question, but I think there is hope, we can gain more understanding of both human behavioural patterns and AI agents’ biases.

And actually if I may add to that, it’s actually very hard to simulate the interactions of human beings or the real economy. That’s where some discussion about the combination of AI and the on-chain economy, or the Web3 space, is potentially interesting. Because when you’re in a digital environment where most of the interactions are rule-based, driven by automated smart contracts, it’s a little bit easier to simulate the interactions and the environment.

Suppose I want to try a trading strategy in the traditional financial market. I have to worry about: are the traders rational? Are they doing the right thing? If I trade digital assets or crypto assets on-chain, a lot of the transaction costs and price impacts are well specified at decentralised exchanges. So, in essence, it’s interesting to apply AI in that environment. Of course, whether we move most financial assets on-chain is a separate question, maybe that never happens. But just in terms of a clean environment to think about using AIs to simulate real interactions, that’s actually a pretty good setting.

Prof Hwang: So let’s say I don’t want to create a model from scratch. I want to use an off-the-shelf product like GPT or Gemini. And then some people are asking, they’re mostly trained on data from the West. Let’s say I want to use that to simulate, and as a result find the best decision for here in Asia, let’s say. How do I do that, do I just prompt it with “assume you’re in Asia”? Does that actually work, or do I have to start everything from scratch and feed it mostly data from Asia, from the local context?

Prof Cong: So, I think our computer science colleagues have probably done a lot more work along that dimension. Just speaking from an economic perspective: we do know in finance or economic research, there’s the issue of “garbage in, garbage out”. If the data quality is not informative, we can’t do too much. I’m not saying the Western data is of lower quality, but if you want to incorporate some local culture or specific responses, my view is that you have to use local data.

I’ve seen some apps trying that. I can’t recall the name, there’s one called Ada Health, and also MazaCAM, I think, in Africa. What’s interesting about language models is, you have to use local language to train the model. Which makes sense if you study classics, you have to use Asian language to train the model.

The Ada Health one is interesting because certain illnesses or symptoms are very localised. Seeing the same symptom within a particular group in one location might imply something different compared to seeing it in another location. So I think all of this points to the importance of utilising local data to train the model. Of course, there’s a challenge, what if we don’t have enough local data? And that’s where generative models are very interesting.

Can you simulate data that’s more congruent with local culture or practices? Can you simulate the environment, the interactions, and use that data in turn to train the model?

That seems a reasonable way to go if data is limited.

Audience Member 1: That was a very engaging talk, thank you so much. So, my question is more in the finance domain. I’m a finance guy. I’m wondering if we can back out or reverse engineer the implied objective function of a central banker or a portfolio manager by observing their actions and then compare that with their actual stated mandates. Like, for a portfolio manager who has to invest a certain amount of the portfolio in ESG stocks, or a central banker who has a mandate to keep inflation in check.

Prof Cong: Are they abiding by the mandate, or are they trying ...? (Yes)

Prof Cong: Yeah, that’s a great question. It’s related to the corporate finance setting, right? We’re trying to learn about the objectives. So, the argument, at least what we tried, is a revealed preference argument. If these are the potential candidates for the economic goal or objective of the decision-makers, what combination of these objectives must they have in order to make their historical actions optimal, right?

Supposedly, they are rational and optimising their decisions. In that case, we would need a few candidate objectives: maybe the mandate is a default objective. There are some other potential objectives, maybe empire-building type objectives, or some behavioural ones. At the same time, we also need to know the optimal action corresponding to each objective in order to do the decomposition.

Suppose I buy 10 shares of Apple and sell 2 shares of NVIDIA.

Maybe one objective would say “buy 10 shares of Apple”, and another might say “sell one share of NVIDIA”. With that, I can do a decomposition, what weights do I have on these objectives? That’s provided I know the optimal action.

In the corporate finance or firm environment, we typically can’t say “this is the optimal action of the CEO”, because in the analysis we usually control for other actions and isolate just one. In reality, they are making high-dimensional decisions. So that becomes quite challenging, and that’s where this particular framework becomes more useful.

Now, in the central bank setting, I think we would need more data. There isn’t a large cross-section.

In the corporate setting, we have many firms in each market, so that allows us to train some of the models, but with macro policy or central bank decisions, do we have enough cross-sectional or long time-series data to fully train such a model? That’s unclear.

I’m not saying it’s not feasible, the models themselves are not super large, so you don’t need a huge dataset to train them, but compared to corporate finance or asset pricing settings, the amount of data becomes more of an issue. Framework-wise, I think it’s essentially a revealed preference argument and that could be a good project to work on. I wish I had more macro background, but I wasn’t really trained in that area.

Audience Member 2: What’s your opinion, from a policy and governance standpoint, given how fast AI is developing, do you think the only way that policy and governance can catch up is by using AI to govern AI? If not, what do you think is a possible solution, given how quickly things are developing now?

Prof Cong: Right, that’s a great question, and I think many governments are thinking about or dealing with this.

In the US, which at least for now is relatively open to innovation, and relatively unconstrained innovation in the tech sector, there are still big AI initiatives being formed. They’re doing red-teaming, stress-testing the system, and so on. That definitely requires coordination. You have to get many players: researchers, practitioners, policy makers, to come together.

I think government has to lead that initiative. ChatGPT is not going to initiate that coordination, because of different incentives, right? As a private firm, it won’t fully resolve the problem. Do we have to try a gazillion different scenarios to figure out the risks and so on?

That’s still a very open question.

This afternoon, with some of our computer science colleagues, we were chatting about how to improve alignment. How do we make sure it’s safe, because these agents are very creative, we allow them to be creative. So, it’s hard to write out all the possible scenarios.

In economic jargon, we talk about incomplete contracts and we can’t spell out all the contingencies. So, maybe we can borrow ideas from how human societies are governed.

When I take a Grab or Uber ride, we see a list of expected behaviours from the driver, but that’s not a full contract. There are actions not mentioned, but we don’t worry too much, because if the driver takes a very dangerous action, the background law in society will penalise bodily harm and so on.

So accumulating cases over time is helpful, but AI is moving too fast. We took centuries to accumulate legal cases. Another thing that’s helpful is deterrence. The death penalty, for example, is a powerful deterrent in some setting, but do AIs feel the same?

If you tell an algorithm, “If you do something that is unruly to humans, I will shut you down”, do they care? Are they fearful? Are they concerned? That’s unclear. I haven’t died yet, but I fear death. Why?

There’s a long history of evolution. The species that didn’t fear death have already died. The more careful ones survived.

So in a way, do we need AI algorithms to go through a similar process? Does that mean we need to quickly iterate AI generations, and if something goes wrong, we shut it down? How do we make sure that knowledge or memory is passed down to the next generation of models?

These are all very open questions – but they require research and expertise from all fields; legal, social science, humanities, computer science, data science. This is a pressing issue we need to think about before it’s too late.