Paolo Rosso: "It's too easy to blame artificial intelligence for misinformation, but at the end of the day, it's a tool people behind it."

It is no secret that hate speech, misogyny, and misinformation are rampant on digital platforms and social media, sometimes having a significant impact on the general public. Detecting these messages has therefore become essential to prevent the potential harm and misinformation they can cause. Today, this is possible thanks to artificial intelligence that uses natural language processing models to detect this subject discourse. To explore this topic further, we spoke with Paolo Rosso, Full Professor computer science at the Polytechnic University of Valencia and an expert in natural language processing and speech analysis speech social media. Rosso attended the congress “Public Opinion and Democratic Civic Engagement: Expanding Reflection in Public Agendas,”organized by the School of Communication in collaboration with the Institute for Culture and Society at the University of Navarra. As framework this framework, he delivered a lecture Reactions to Information Warfare and Polarization: AI Models for the Detection of Disinformation and Conspiracy Theories.”

To provide some context, could you share a few insights into how programs are trained to detect speech , misogyny, or“fake news”?
Artificial intelligence models are trained using large amounts of data, from which they are able to identify common patterns. This also applies to images, such as memes or videos. Nowadays, these AI models come pre-trained by large companies. The next step is to fine-tune them so they focus on the specific problem at hand. This is called fine-tuning.

The programs are given a few examples, which they use internally to adjust the final layers based on deep learning. You can provide them with just a few examples—a process known as Few-Shot Learning—of sexist and non-sexist memes and videos, along with a description of the problem, to help them distinguish between them more effectively.

How does the use of videos and images affect model training? Sometimes, the text can seem very neutral, but depending on the image you use, it can be completely unrelated, right?
Yes, there are models called contrastive learning models where, on one side, you input the text and, on the other, the image to contrast the patterns across modality. As you say, for example, in sexist messages in memes, people play with quotation marks, humor, irony, sarcasm… The text is generally the most informative part, but in these cases, the image makes all the difference.

Are these tools that are trained to detect this subject discourse used only by you as academics?
Nowadays, it’s difficult to train an tool from scratch. A few years ago, yes. Now, you can’t compete with “Big Tech.” As I mentioned earlier, Large Language Models come pre-trained with a Issue data would be unthinkable for a research center. In other words, having your tool is out of the question in many of these cases.

So, there are tools available that can be used; some are open access. These artificial intelligence tools are available to everyone, and the goal is to fine-tune them so that they focus on and address the problem you want to solve in the best possible way using some of the techniques I mentioned earlier—but they are not tools we developed ourselves.

Why is it important to identify this subject messaging in memes and videos?
We receive a lot of memes through digital platforms, such as WhatsApp. But this is especially important today for younger generations because they spend a lot of time watching videos on platforms like TikTok. On these platforms, users consume short videos where the content is increasingly tailored to what they’ve seen before due to the algorithmic bias behind the system. Therefore, it’s important to address this issue because the algorithm drives an increase in the consumption of these messages.

Is this only a problem on TikTok?
No, no. It does seem to be the most dangerous platform for receiving that subject message because, these days, young people use TikTok. We’re currently investigating this platform. We also previously investigated what used to be called Twitter, now known as X.

I've also seen that they were studying irony. How is it possible to program an AI to detect something as human as topic ?
Yes, irony certainly makes it harder to process figurative language. It's even harder than processing natural language among humans. Sometimes we notice that the person we're talking to doesn't understand the irony, and we play around with that ourselves, don't we?

These models are able to extract, understand, and then generalize these patterns when they need to detect irony in new social media posts based on large volumes of data. Thus, they are able to extract these patterns, although in some cases, a human must always be involved. A person—an expert, a linguist—always has to be there because it’s true that, in many cases, the patterns aren’t so obvious. It’s not that easy for Large Language Models to detect irony, but they’re getting better at it all the time.

Do we know if most of this hate speech, misogyny, or fake news is also generated by artificial intelligence?
It is we who generate hate speech, above all. We live in an increasingly polarized society. You just have to look at what’s happening. The fragmentation of information creates an environment where misinformation or biased and partially false information can spread. So, it would be a bit simplistic to blame artificial intelligence. It’s true that certain messages can also be generated automatically and can, let’s say, spread rapidly.

A few years ago, the Massachusetts Institute of Technology (MIT) conducted a study on what was then Twitter, which empirically demonstrated that messages containing false information spread 6 or 7 times faster than messages containing accurate information. But it is people who spread them, regardless of who may have originally created the message. There are also programs of study have shown that many of these messages may have been automatically generated in troll farms located, for example, in Russia.

So these messages—whether created by humans, trolls, or even generated automatically—can further polarize society and our democracy. But who spreads them? We do. At the end of the day, we are the ones to blame.

Why?
Because we’re driven by confirmation bias. We come across an opinion about a politician we really don’t like, and we don’t check to see if it’s true. It reinforces what we already think. In this way, we’re the ones who “help” spread false information—not to mention hate speech.

For example, this was evident after the DANA disaster in Valencia. Pseudo-journalists and malicious actors have taken advantage of the disaster to further polarize society. It’s all too easy to pass the buck by blaming artificial intelligence models, which certainly don’t make things any easier in this regard. On the other hand, yes, they do help us. It’s a tool there are people behind and in front of it. It can be very useful for detecting these cases, but it can also automatically help spread false messages. 

If these tools can be trained to detect such discourse, can they be trained in such a way that they also “learn” to create it?For several years now, there has been this great fear regarding false news and disinformation—often called “fake news,” although we shouldn’t use that term because it isn’t fake from start to finish. There was a great fear that these artificial intelligence tools would indeed be used to create deepfakes that would be difficult to distinguish from the real thing. At first, it was relatively easy to tell them apart, but lately that’s not so true, and it’s easier to fall for them and believe them.

But there are programs of study such as theIberifier project at the University of Navarra— programs of study show this isn’t such a major problem; in other words, false content is generated primarily by people—by humans—rather than by artificial intelligence. Ultimately, AI is a tool can also be used as a weapon. It can help us—and greatly—solve many problems, but at the same time, it could also be exploited by malicious actors.