It’s Not Magic — It’s Just Math

FORUM: Everyone is talking about artificial intelligence now. But what exactly do we mean by it?

Stuckenschmidt: There are different approaches, but I favor a definition that holds that AI is about developing computer programs that solve problems that would typically only be considered solvable with the help of a certain amount of human intelligence. But that does not necessarily imply that the computer reaches a solution in the same way a human would.

Ponzetto: Exactly! There’s no magic to it. It’s just math. We use mathematical methods to teach computer programs — to teach AI — how to behave like humans. We feed them with vast quantities of data to achieve this. All the successful AIs that members of the public use in their everyday lives are based on data and statistical methods.

FORUM: And why is there such a buzz about AI now, specifically?

Ponzetto: The decisive factor driving the tremendous speed of current advances is the sheer quantity of data that has now become available and that we are now able to use. Those opportunities did not exist in the past in their current form. Right now, everyone is abuzz about AI systems like ChatGPT that are based on a category of models referred to as neural networks. These neural networks are structured in such a way that the number of connections between neurons is extremely high. Giving such a network access to more data makes the connections between individual neurons stronger and the model more powerful.

Stuckenschmidt: We are currently in a phase of AI hype again. By now, there have been a few different phases of hype. Amusingly, we even had a phase already in which neural networks were hailed as the panacea to every problem — and we seem to be back at that point again now. I would say that two factors are driving the current hype. The quantitative leap in data availability that has already been mentioned has been flanked by recent progress in some traditional AI domains that has been quite striking. Strides have been made in the processing of unstructured data such as text, images, and combinations of both. The other factor is simply marketing. What began with IBM’s Deep Blue beating the reigning world champion at chess has peaked, for now, with Microsoft marketing ChatGPT. The technical development of systems like ChatGPT has been ongoing for more than a decade, but the hype has only just begun.

FORUM: What are the main research focus areas of your chair? And why is continuing research in this field especially crucial now?

Stuckenschmidt: We work on a diverse range of topics. In basic research, we are working on the development of symbolic machine learning methods that use and generate symbols — in contrast to neural networks, in which data and results are simply numbers. We hope to make machine learning outcomes more transparent and traceable in this way. This is not only an interesting theoretical problem — it has huge practical relevance in light of current debates about regulating the use of AI. We are looking at the applications of AI methods in a range of domains as we take this forward. At the moment, these are mainly the areas of supply chain management, mobility and smart city applications, and psychology and learning research. We always work together with experts from each specific application area to ensure that our ideas generate genuinely useful results. These experts are often colleagues from other schools, especially the Business School and the School of Social Sciences, but we also work with enterprises across the region, and we work with the City of Mannheim on mobility.

Ponzetto: We seek to analyze and improve existing AI programs together with our research team. One of the issues that we are interested in, for instance, is the question of what ChatGPT has actually read. What books or websites was it trained on? We are interested in discovering more about this because ChatGPT is a proprietary model and no transparent record of the data set used to train it is publicly available. And what influence does the input that was used have on the texts it generates? This is highly significant because input texts can contain bias in many forms — gender stereotypes, racist stereotypes, and so on. Conducting research in this area is immensely important, in my view, because the goal we strive for must always be to produce AI that is accessible to all, compliant with democratically agreed rules, and, above all, not beholden to capitalist interests. Research must shoulder the responsibility of working towards this.

FORUM: Would you say that there are areas in which AI is already highly proficient?

Stuckenschmidt: Deep Blue and ChatGPT are good examples of things that AI can do well: handling vast quantities of data that a human could never keep track of, let alone analyze, and systematically searching for optimal solutions are capabilities we see realized there. Humans tend to perform poorly in both of those areas. And that is where the opportunities to gainfully deploy AI methods arise: they can take on the tasks at which humans do not tend to excel.

Ponzetto: I use AI to revise my own writing — when I need to structure or shorten texts, ChatGPT can be quite useful. Tools like this have also reached the point at which they produce passable results when tasked with small translation jobs.

Stuckenschmidt: I agree. To be frank, I don’t find ChatGPT’s efforts to generate new texts on a given topic convincing. The texts it generates often feature strings of factual errors alongside a few correct statements. They also frequently contain very general statements that are not wrong as such but add no real value. I must say that I am not plagued by worries that students will all gain top grades with artificially generated dissertations in the future.

FORUM: Let’s stick with ChatGPT for now. How is this particular AI system breaking new ground?

Ponzetto: ChatGPT is the first AI system that the general public can access and use with an awareness that they are using AI — that makes ChatGPT different from, say, the AI capabilities built into cellphones. On top of this, ChatGPT appears to be relieving us of tasks that the education system and entire professions are based on.

Stuckenschmidt: Research on language models based on neural networks has been ongoing for years already. ChatGPT is only one of a host of models that have been developed. What distinguishes it from the other models, as I see it, is that a vast amount of money has been plowed into engineering this model to make it robust enough for use by the general public. ChatGPT also seems to have exactly the capability that is needed to “pass” the Turing test, a famous test of artificial intelligence based on the premise that a program can be considered intelligent if a person chatting with it cannot distinguish it from a human. ChatGPT copes very well with that challenge.

FORUM: What do you identify as the biggest opportunities and threats that the use of AI is creating for society?

Stuckenschmidt: I see AI as a tool that can be hugely useful when it is handled properly and hugely damaging when it is misapplied. It can make particularly useful contributions to analyzing large quantities of data, systematically searching for optimal solutions, and making rational decisions. But this is only true, of course, when AI methods are applied correctly and in reasonable ways. This becomes especially clear when we consider the subfield of machine learning, an area that is currently attracting a lot of attention. Learning processes can only assimilate what the data presented to them actually contains. When the data contains incorrect or biased representations of reality, the learning processes will also yield incorrect or biased outcomes. This problem is compounded by the fact that unfair decisions made using AI in turn generate further data that flows into the learning process and amplifies the original bias to create outcomes that are even more unfair.

Ponzetto: One of the threats that certainly arises with systems such as ChatGPT is that the data basis it draws on is not clear. “Bad” data also yield bad results, for instance by reproducing, spreading, and therefore reinforcing bias. We are also already starting to see models with capabilities that were not predictable in advance. ChatGPT can produce translations, for instance. Copyright issues also remain unclear. Achieving greater transparency and tighter regulation is essential. I can identify a major opportunity to foster educational justice. AI can make it possible for anyone with internet access to take advantage of education. It can even provide learning opportunities tailored to individual needs.

Interview: Dr. Maartje Koschorreck, Jule Leger/December 2023