ChatGPT does not appear to offer one stable news diet across all ways of using it. A study from the University of Hamburg and the Leibniz Institute for Media Research found that the web interface and the API can point users toward very different sources, even when the subject is news.
The research looked at more than 24,000 AI answers over five weeks, focused on news-related questions in German-speaking countries. The result is a useful warning for readers, developers, publishers, and anyone treating AI answers as a neutral route into current information.
What the study compared
The researchers examined ChatGPT responses to two kinds of prompts. One set used standard news-related questions. Another set specifically asked for a broad mix of sources.
That distinction matters because many people assume that asking an AI system for more variety will improve the quality of the answer. The study suggests the outcome is more complicated. Broader sourcing can increase the number of cited sites, but it can also pull in sources that are less reliable or harder for users to judge.
The central finding is that access method changes the source mix. The ChatGPT web interface and the API did not simply return the same information through different technical channels. They showed clear differences in which outlets appeared, how often they appeared, and what kinds of sources were emphasized.
The web interface gave more weight to familiar media
In the web interface, outlets connected to OpenAI’s licensing partner Axel Springer had a prominent role. The study found that welt.de and bild.de, described in the source as conservative tabloid-style brands, made up about 13 percent of all references on the web interface.
Those same sites were much less visible in API results, where they accounted for just 2 percent of citations. The ranking gap was also large: on the web, welt.de was the top source and bild.de ranked fifth. In the API, welt.de dropped to 61st and bild.de to 158th.
The researchers said these differences were statistically significant. That is important because the pattern was not presented as a small random fluctuation. It points to a meaningful split between two ways of accessing the same AI product.
The web interface also looked more aligned with widely used German media lists. Compared to the Reuters Digital News Report 2025, the web interface’s source list overlapped 45.5 percent with top German media. The API’s overlap was 27.3 percent.
Public broadcasters received more exposure in the web interface as well. They accounted for 34.6 percent of sources on the web interface, compared with 12.2 percent through the API.
The API leaned more toward encyclopedias and niche outlets
The API showed a different pattern. Unless the prompt explicitly asked for more diversity, the API leaned more toward non-journalistic sources, including encyclopedias and niche sites.
Wikipedia was especially visible in API citations, accounting for almost 15 percent. The API also pointed more often to little-known local outlets with limited reach in Germany. Deutsche-handwerks-zeitung.de was one example that appeared far more often in API results.
This does not automatically mean every API result was poor or every web result was better. The study’s point is narrower and more concrete: the same tool can create different news environments depending on the access route. For users, that makes source awareness part of the task.
For developers, the finding has another implication. An application built on the API may not inherit the same source profile that users see in the ChatGPT web interface. If the product is meant to summarize or recommend news, that difference can shape what users end up reading.
Asking for diversity can widen the risk
When users asked ChatGPT for a wider range of sources, the system listed more sites. Through the API, it listed 1.4 times as many sites compared to standard queries. Through the web interface, it listed 1.9 times as many.
More sources, however, did not necessarily mean stronger information. According to the study, these broader requests led to more citations of politically biased or propaganda outlets. One example was news-pravda.com, which the source says has reported ties to the Russian government.
The system also sometimes linked to fake or nonexistent domains, including news-site1.com. In other cases, it pointed to lookalike sites that generate AI-written "news."
The study did not find that ChatGPT’s average cited source leaned strongly away from the national average. Most sources scored between 3.89 and 3.98 on a seven-point scale, with no meaningful difference from the broader population. But averages can hide practical problems: a generally balanced source mix can still include individual links that users should treat carefully.
Why this matters for AI news search
The researchers caution that ChatGPT’s idea of "diversity" may not match what readers mean by useful informational variety. The AI may surface sources that differ linguistically from mainstream outlets, rather than sources that add meaningful context or trustworthy reporting.
The study also underlines a transparency problem. OpenAI does not explain what causes the differences between the web interface and the API. Users therefore have to make their own judgments about why one route produces one source mix and another route produces something else.
There is also a timing issue. The researchers describe the results as a snapshot because OpenAI regularly changes the system without notice. That means the exact rankings and shares may change, while the broader lesson remains: AI-generated news recommendations are shaped by systems users cannot fully see.
This fits a larger pattern noted in the source article. Generative AI search tools often rely on very different sources than traditional search engines like Google. At the same time, major AI systems are now twice as likely to spread false information as they were a year ago, because they now answer nearly every query instead of sometimes declining when facts are unclear.
For readers, the practical takeaway is simple: do not treat a ChatGPT news answer as a fixed map of the media landscape. Check where the links come from, notice whether you are using the web interface or an API-based product, and be especially careful when asking for broad source diversity.