Marta Serra-Garcia: "The use of ChatGPT is not correlated with trust in ChatGPT"

BarcelonaThe conversation about artificial intelligence has focused mostly on the product. ChatGPT (OpenAI) or Claude (Anthropic) have become ubiquitous on users' devices; and, with them, in the social conversation. Like all technological disruptions of the scale that is assumed for AI, however, widespread adoption is the key to any significant advance. This adoption is among the fields of study of Marta Serra-Garcia, a Catalan economist and professor of economics and strategy at the Rady School of Management at the University of California - San Diego specialized in behavioral economics, who has studied public trust in AI. Serra-Garcia receives ARA at the Ciutadella campus of the Pompeu Fabra University, where she studied and where she has returned to participate in the European Meeting of the Economic Science Association, organized by the center's department of economics.

She has been studying the adoption of AI for years, but the technology has made Copernican changes year after year. How do you analyze a field that changes so rapidly?

— We are still learning about it. It is changing very rapidly, and it is a fundamental change in the economy. We need scientific evidence of what is happening right now, but we know it will keep changing. The question is: which things are fundamental, generalizable? And which others will change as models advance? There are certain patterns of technology adoption that have been studied since the 80s, with the first computers, and that we also see reflected in the interaction with AI.

Are the same adoption trends being found as in other technological movements, such as the internet or computers?

— A faster adoption is being seen. Regarding the economic effects, there is no consensus. AI is very accessible to everyone, but using it at scale is expensive. If you want to automate projects, you have to think about access to tokens. From what is known about companies like OpenAI or Anthropic, which are estimates, the use of tokens is somewhat subsidized. When the cost changes, what will happen? It is already beginning to be seen that in certain tasks, for which very high productivity gains were expected, these gains are not as large when compared to the costs. Human labor has a marginal cost and a marginal productivity; and both compensate for each other. AI is so volatile that, in the business world, we will see how it is applied.

And at a user scale?

— There are applications that are very accessible. Everyone can open ChatGPT and ask it things. It provides highly personalized information and can be very useful as a consumer, because it aggregates and summarizes in a way that Google, until recently, did not do. There will also be growth when they start integrating automated tools: that they search for our flights, make our reservations, our purchases... But all of this has a cost. As the AI market changes and the cost of tokens rises, we will see whether it is adopted more or less.

Often the evolution of AI is compared to the dotcom bubbleof the early 2000s.

— One thing is past experience, and how informative it is regarding the future, and another thing is adoption. The dot-com bubble is an experience to keep in mind; and we know that financial markets often generate bubbles. But it is very difficult to say that before it happens. There are experts who point out that the productivity gains of AI are good, but that they do not justify the valuations of Anthropic or OpenAI. But there are also those who say that the promise is a paradigm shift. Now there are a few large companies with a lot of power; but it is possible that all this will change with the entry into the market of other models. There are those who argue that there will be a differentiation: that basic models will arrive, accessible to everyone; and others more complex for those who want to do more advanced things. But this is not a certainty.

The dot-com bubblealso showed that the uses initially foreseen for technology are not the final ones. Are we in a similar phase with AI?

— It is hard to say. It makes sense to think that we are still in the early days. Right now, one of the most requested functions for AI is information searching: there is less Google and more personalized search. Will this be adopted widely? It will depend on how information markets change. There is a clear correlation between trust in the media and trust in AI. But it is just one of the uses that AI will have, and what is done with it will depend a lot on each market, each culture, each society. We still need data to know if its effects are the same on the US population as they are on the European one. There are different regulations, for example.

How do the regulatory environments of the USA and the EU differ?

— The USA is one of the countries that appear most skeptical regarding AI. And there is skepticism regarding what artificial intelligence says; but there is also skepticism regarding regulation. Those responsible for regulating—congresspeople—are mostly older, are not up to date with technology, and it is very difficult to attract the young talent that is at the technological frontier to identify the risks and place limits on them.

Less trust in institutions, in line with the conclusions of your study.

— Yes, that too. Everything is related. In Europe, in fact, there is a bit more trust in AI than in the US, although the levels are similar. There is a greater tradition of thinking about the protection of citizens. Society must choose and design regulation to strike a balance between protecting more and giving space for innovation. In Europe, the balance still favors protection; in the United States, a little less so.

Part of the European tech world considers that AI regulation is economically positive, because it generates safer products for potential customers. Does this idea exist in the USA?

— If the user is sophisticated, and understands that AI has problems—regarding privacy, regulation, protection—then companies will have intrinsic incentives to generate trust. But we are not all perfectly sophisticated. From this perspective, do companies have these incentives to make AI safe if the user does not ask for it? That is not there yet. As skills with technology advance, either regulators will come to decide which things cannot be done, or users will ask platforms to offer more guarantees.

Have AI solutions reached the market too soon, considering their progress in security?

— A company may consider that if it has a model that works, it can make it public, because users also provide data so that the model can improve. A year and a half ago there were promising capabilities, but it was not known how far they would go. It is the chicken and the egg: without the critical mass of users, the model cannot be trained; without training the model, it does not improve; and, if it is not perfected, it does not reach a critical mass. The technology itself needs usage to learn.

As users have become accustomed to using technologies, has it become easier to trust them?

— I haven't seen data on this, but it is plausible. However, often the use of technology and trust in technology are not the same thing. There are those who use it, but do not trust it; and vice versa. We have seen that the use of ChatGPT is not correlated with trust in ChatGPT. For example, if a user uses AI a lot, they may also know its limits better. If they don't use it as much, perhaps they are guided by what their environment says.

Since Trump's return, there was much talk about the confrontations with academia, with cuts, for example, in certain departments. How has this conflict evolved?

— It is a fact that the National Science Foundation (NSF) practically no longer provides funding for internal research projects in social sciences. And they say they will cut more things. California, for example, has increased its funds to compensate for it. There are very good people leaving for Canada, Switzerland...

Do you consider it dangerous that the academy is losing resources in such a large technological expansion as that of AI?

— It could be dangerous, but I believe we still have tools. I want to be optimistic: in Europe there are very good researchers; and in the USA, despite the cuts, there are too. The fact that the most cutting-edge researchers are leaving for the corporate world could be a risk. But there are still capable people left who can collaborate with both companies and regulators. In fact, the companies themselves are the ones that want independent studies to ensure public trust. A company researcher will only study that company's model; but the academy is not tied to any model. There will be changes, but there is still plenty of talent to move forward.