ChatGPT, Gemini and Co. answer message questions directly. But it is completely unclear to users which criteria AI models use to select their sources. A new study reveals a frightening picture. Because language models not only spit out a few established media as sources, but also place disinformation sites next to reputable sources without comment. This is a great danger for the formation of public opinion. A commentary analysis.
Source selection of AI chatbots is difficult to understand
- More and more people are having messages summarized by ChatGPT and Co. instead of reading original articles. Numerous media outlets are increasingly feeling this resulting economic consequences. According to estimates by the analysis platform Sistrix, German websites are now losing 265 million clicks per month due to such so-called zero-click searches – and thus indirectly advertising revenue.
- According to an analysis by think tank Agora Digital, the real problem is not just the errors that AI models produce in their answers, but their role as new gatekeepers. This means: Chatbots make a pre-selection of sources – according to them Criteria that are usually not understandable for users. The study analyzed 675 message queries to ChatGPT, Gemini, Claude, Perplexity and Google AI Overview and evaluated almost 5,000 source references.
- The result: The researchers came across 554 different websites that appeared as sources. Most of the answers from the AI models examined but concentrated on a few established media outlets. According to the analysis, domains that spread disinformation were sometimes listed next to reputable sources without users being informed of this. Which sources appear would depend on the particular AI model, the topic and even the wording of the query.
AI as the new gatekeeper of the news world?
The results from Agora Digital reveal that AI has long since become more than just a tool for finding information. It is becoming a gatekeeper to public debates. Because if you get answers straight away, clicks on original sources less often.
The problem: Many users don’t just rely on AI answers already prone to errors can be, but also to a limited circle of sources and therefore also information. At the same time, the power of attention is increasingly moving from editorial teams and search engines to the operators of large language models.
Far from the fact that one in ten AI answers is wrong or contains errors, this creates one further danger to the formation of public opinion. Because users can hardly understand how AI models cover the digital news table – nor which ingredients are even taken into consideration.
They are also particularly problematic Cooperations between AI providers and publisherslike between Axel Springer Verlag and OpenAI. Not necessarily because ChatGPT per se shouldn’t use Bild newspaper as a source, but because it does so in an inflationary manner.
This not only makes AI responses relatively one-sided, but also undermines them due to a lack of transparency the maturity of many users – not to mention Russian propaganda networks cited as sources.
Voices
- Vivien Benert, lead author of the Agora Digital studydemands to Spiegel: “Transparency and disclosure of the criteria for selecting sources during ongoing operations, a verifiable allocation of statements and sources as well as disclosure of existing collaborations between AI providers and media companies for users. Without insight into the selection of sources, neither liability, diversity nor fair remuneration can be enforced. (…) No provider transparently documents how the selection of sources to answer prompts works.”
- Mika Beuster, Federal Chairman of the German Journalists Association (DJV)in a statement: “Quality journalism thrives on human research. Generative AI, which only regurgitates what has already been thought and said, cannot create new perspectives. I see the danger that readers and publishers’ advertising customers will not be willing to spend money on robot journalism. (…) This raises the suspicion that AI is not being used to support but rather to replace journalistic work. This must not become a model for other media companies. Where it involves, for example, the analysis of huge amounts of data AI is a welcome help for editorial teams.”
- Liz Reid, Head of Google Searchdefends the platform’s own AI capabilities: “We continue to drive billions of clicks to websites every day and believe that the value exchange between search and the web remains strong. (…) We have trained our models to deeply understand the web and know when and how to point to the most relevant pages. And our AI responses contain highly visible links, clear attribution and inline citations.”
What needs to change now
The study not only makes it clear how different the source profiles of the individual AI models now are, but also how few users can understand why certain sources appear and others do not. Demands for more transparency are therefore rightly becoming louder and louder.
Because: The selection of sources must be verifiable so that content can be attributed to its origins and the handling of journalistic content and disinformation is regulated in a binding manner. At the user level, this shows once again that AI answers may provide an introduction to a topic, but not a replacement for the original sources are.
Or: Anyone who relies exclusively on AI summaries of political or social issues will not end up being served the entire debate, but only a section that an algorithm considers relevant. It’s sad enough that many people no longer compare different sources or viewpoints with one another, but rather compare them Confirmation of your own opinion seek.
And as if that wasn’t damaging enough to our culture of debate, public opinion and education, AI exacerbates this simplicity also. In other words: Information diversity is increasingly becoming information convenience – with serious disadvantages such as a loss of empathy, maturity and the ability to discuss.
Also interesting:
