Study: One-third of new websites are AI-generated

Many people do not trust Artificial Intelligence-generated content. However, it is becoming increasingly difficult to escape it online. This is according to a recently published study by researchers from Imperial College London and Stanford University, who collaborated with the Internet Archive to analyze the web on a sample basis. Their finding: Around 35 percent of all newly created websites are AI-generated or at least AI-assisted. The phenomenon of cheaply produced AI content, also known as AI Slop, has long since spread from websites to platforms like YouTube and Instagram. This 35 percent surprised even the researchers, as the proportion was zero before the release of ChatGPT at the end of 2022. However, there are positive surprises regarding concerns about the qualitative impact of this trend.

For their study, the researchers took a representative sample from the Internet Archive (Wayback Machine) and examined it with the AI detection tool Pangram v3. Additionally, human fact-checkers reviewed extracted claims from AI-generated pages to investigate how AI generation affects the content. They are dealing with one of the fastest transformations in digital history, researcher Jonáš Doležal told the online medium 404media.

Is diversity on the web declining?

The content was checked against six common assumptions. These include fears that AI is reducing diversity on the web, that language is being artificially enlivened, that misinformation is increasing, that individual writing styles are disappearing, and that AI texts link to fewer external sources. The concern that texts are becoming longer with less information density was also investigated.

The result: Only two of the hypotheses proved true. The semantic similarity between AI-generated websites was 33 percent higher than among human-written pages. Thus, the diversity of ideas and perspectives on the web is decreasing due to AI. A study from Duke University comes to a similar conclusion: AI models are also less creative as a group than humans and level out linguistic diversity. And AI is also noticeably more cheerful: AI-generated content showed a 107 percent higher proportion of positively rated documents than human texts.

Two changes, four false assumptions

One of the most widespread assumptions, that there is more misinformation online, turned out to be false. No statistically significant correlation was found between the proportion of AI and the rate of misinformation. However, a study by the European Broadcasting Union had shown that AI chatbots incorrectly reproduce news content in 45 percent of cases. According to surveys, 75.1 percent of surveyed US adults believe that AI is leading to more misinformation online. The researchers also could not prove the disappearance of individual writing styles, the lack of links, or a tendency towards longer texts with less information.

However, the researchers see no reason to breathe a sigh of relief: The decrease in diversity on the web could affect society like disinformation. And the fact that AI is on par with human texts regarding misinformation might simply mean that the web fundamentally has a problem with truth. Furthermore, AI could be more adept at spreading misinformation that is simply difficult to verify.

The lack of labeling for AI pages is also a issue because it leads users to question online information wholesale. Another study proves that users often uncritically adopt AI answers: AI makes users lazy thinkers – and more confident, as researchers from the University of Pennsylvania showed. The researchers recommend cryptographic verification of human authorship. Search engines and recommendation algorithms should also be geared towards highlighting human-generated content and contributions that serve diversity. How far AI-assisted search is from traditional search engines is shown by a recent analysis: Researchers found significant differences between AI search and Google, for example, in source diversity and temporal consistency.

(mki)