The boom in artificial intelligence (AI) has made a big impact on the internet since the publication of ChatGPT in 2022. Not only are texts, images and videos generated by AI much faster and cheaper to produce, their quality is also improving - which poses a major challenge, because AI generated content is not always close to the truth. Jaro Krieger-Lamina considers this development to be highly concerning. He conducts research at the Institute for Technology Assessment of the Austrian Academy of Sciences (OeAW), including research on the influences of generative AI on democracy. In this interview, he talks about this problem – and what measures we can take.

Sharp increase in AI content on online platforms

The amount of AI-generated content on the internet has exploded in the past two years. Is there quantitative evidence for this?

Jaro Krieger-Lamina: There are no concrete figures, but we are seeing a sharp increase even on platforms where you might not expect it. If you search for scientific publications using Google Scholar, you will now find a surprisingly high number of AI-generated results. This is not yet the case in high-quality scientific journals because the peer review process seems to be working in this regard. But that could change as well if text generators improve.

Can AI content still be distinguished from human-made content today?

Krieger-Lamina: AIs have become so good that it's usually no longer possible to reliably distinguish between the two. The quality is good enough to fool non-experts. But experts are still able to recognize the errors that are almost always present.

The question is what will happen to the collective knowledge of humanity on the internet.

Why is AI-generated content problematic?

Krieger-Lamina: The big problem is the tendency of AIs to “hallucinate” and invent “facts”. AI systems can generate text that sounds plausible but conveys false information. That is why their field of application is still limited today. The question is what will happen to our collective knowledge on the internet if more and more content is generated by machines and a significant portion of the information is false. If future generations of AI systems are trained with this falsified information, it will lead to absurd results in the long term because the false information multiplies with each generation. Language models can only work if they are trained on the basis of human content.

AI content: cheap to produce

How long do we have before the internet becomes worthless as a source of information?

Krieger-Lamina: That depends on the specific field. There is already a lot of AI-generated content on many news portals because it costs practically nothing. Journalistic scrutiny no longer takes place, and that is certainly problematic. Overall, I believe that the proportion of AI content on the internet is not yet very high. But the economic pressure is high in many areas, and if companies are willing to accept a certain error rate in order to produce more cheaply, they will do so. I can imagine, for example, that picture agencies will disappear in the coming years because AI generators produce results that are good enough for many purposes for less money.

People make mistakes and lie too. What's the difference? 

Krieger-Lamina: In their motivation. Most people don't lie on purpose, but they may make mistakes and commit inaccuracies. We also produce content to consciously gain economic or political influence. But AI has no motivation, it doesn't think. The error rates are also strikingly high. A colleague recently worked with ChatGPT and had it create a biography for a well-known scientist on whom a wealth of reliable public information exists. AI then simply invented a federal cross of merit that the colleague never received. That wouldn't happen to a human.

AIs make it possible to lie to people on an industrial scale.

How great is the danger that AIs will be used to deliberately spread disinformation?

Krieger-Lamina: These systems are perfectly suited to deliberately manipulate people. The tendency towards hallucinations is not a problem for disinformation campaigns and AIs make it possible to lie to people on an industrial scale. When combined with personal information about people that can be gleaned from social media, campaigns can be tailored to individual users. This poses a problem for democracy because public opinion, which is increasingly being formed on social media, is being poisoned. This even allows elections to be influenced, as we have already seen on several occasions, including in Europe. Not only AI-generated disinformation campaigns contribute to this, but also algorithms that favor controversial content.

The risk of disinformation

How can states protect themselves from such attacks?

Krieger-Lamina: In Austria, for example, the Army Intelligence Office, is working to uncover such manipulation campaigns. During the last elections, massive attempts to spread disinformation were uncovered. Following the example of the USA, it is usually not so much about discrediting individual parties or politicians. Instead, it is an attempt to paint a pessimistic picture of the world that is intended to discourage certain groups from voting.

What can we do as a society to prevent the misuse of AI systems?

Krieger-Lamina: One suggestion is to develop “good” AIs to protect people. But I don't believe that such an arms race between AIs makes sense. Instead, it's about deciding where AI should be used at all. It's hard to limit the potential for abuse, and laws are probably not enough here either. Experts therefore suggest putting the focus on people's personal responsibility and improving people's critical thinking through education and raising awareness.

Is quality journalism part of the solution?

This has already been suggested as a solution when social media emerged.

Krieger-Lamina: The challenges will be even greater with AI. The two deciding factors are media literacy and trust. I can't verify every piece of information myself and have to ultimately decide who I can believe. Media literacy must be taught in a more nuanced way and given more importance. At the moment, most people still more or less trust the traditional media, and that provides a foundation. Perhaps we need to add digital watermarks to trustworthy content, with journalists digitally signing their content. However, this requires high technical standards and is expensive because there is a risk of an arms race here too. Furthermore, only individual online sources could be protected, and a broad introduction for the entire network is certainly unrealistic.

In a world full of AI-generated misinformation, quality journalism takes on a higher value.

That all sounds very depressing. Is there any light at the end of the tunnel

Krieger-Lamina: Perhaps new markets for curated, verified information will emerge in the future. In a world full of AI-generated misinformation, quality journalism takes on a higher value, and perhaps people will then be more willing to pay for it in one way or another. Press subsidies could also be linked to this. The differentiation of the information ecosystem will probably accelerate and seals of quality and verified recommendation systems could help users to find their way around. The internet is already an information swamp and now it is a matter of protecting the few flowers.

In which areas should AI not be used?

Krieger-Lamina: I think that using it in administration for creating official notifications – or more generally for decisions that affect people's lives and rights – are insane. These systems work on the basis of probabilities and we cannot guarantee that the results will always be the same. That is why demands to use AI to assess, for example, unemployment benefit claims, asylum applications or creditworthiness are dangerous.