Content
summary Summary

A new study from the University of Hamburg and the Leibniz Institute for Media Research shows that ChatGPT’s news recommendations change significantly depending on how you access the tool.

Ad

Over five weeks, researchers analyzed more than 24,000 AI answers to news-related questions in German-speaking countries and found clear differences between the web interface and the API.

The team compared ChatGPT’s responses to standard questions and to queries that specifically asked for a broad mix of sources. They found that unless you ask for more diversity, the API leans toward non-journalistic outlets, with a bigger share from encyclopedias and niche sites.

Bar chart comparing the proportions of journalistic media, news agencies, encyclopedias, organization/business, and miscellaneous sources cited by ChatGPT's API and web interface under regular and diverse query conditions.
If you don’t request broader coverage, the ChatGPT API pulls more from non-journalistic sources like encyclopedias, while the web interface sticks mostly to journalistic media. | Image: Schatto-Eckrodt et al.

Web interface boosts Springer outlets, API leans on Wikipedia and niche sites

The gap between the web interface and the API is striking. The web version features outlets from OpenAI’s licensing partner Axel Springer, with welt.de and bild.de—both conservative tabloid-style brands—making up about 13 percent of all references.

Ad
Ad

Meanwhile, these sites barely show up in API results, making up just 2 percent of citations. On the web, welt.de is the top source and bild.de ranks fifth; in the API, they drop to 61st and 158th. The researchers confirmed these differences were statistically significant.

Scatterplot mapping citation rates of different news domains in ChatGPT’s API (x-axis) versus web interface (y-axis), using color to indicate which interface favors each source.
The API is more likely to cite niche and tech-focused domains, while the web interface sticks to mainstream outlets like tagesschau.de. | Image: Schatto-Eckrodt et al.

By comparison, the API often points to encyclopedic sources like Wikipedia—almost 15 percent of API citations—and to little-known local outlets that have limited reach in Germany. Deutsche-handwerks-zeitung.de, for example, appears far more often in API results.

Side-by-side scatterplots showing differences in domain citation frequencies for the API (left) and web interface (right) when comparing regular and diverse queries, with points above or below the line indicating shifts in prominence.
The API surfaces more specialized sources, while the web interface favors established brands like deutschlandfunk.de and rnd.de. | Image: Schatto-Eckrodt et al.

Compared to the Reuters Digital News Report 2025, the web interface’s source list overlaps 45.5 percent with top German media, while the API overlaps just 27.3 percent. Public broadcasters also get more exposure on the web interface (34.6 percent of sources) than through the API (12.2 percent).

Broader source requests can surface fringe and unreliable sites

When users asked ChatGPT for a wider range of sources, the system listed 1.4 times as many sites through the API and 1.9 times as many through the web interface, compared to standard queries. But more variety didn’t always mean better information. According to the study, these requests led to more citations of politically biased or propaganda outlets, like news-pravda.com, which has reported ties to the Russian government.

Four-panel line chart showing weekly trends in source diversity and concentration metrics (Herfindahl-Hirschman Index, Shannon Diversity Index, Rao’s entropy, and unique sources) for regular and diverse queries over a five-week period.
Source diversity and concentration changed week to week, highlighting how unpredictable ChatGPT’s news results can be. | Image: Schatto-Eckrodt et al.

The system also sometimes linked to fake or nonexistent domains, like news-site1.com, or to lookalike sites that generate AI-written "news."

Recommendation

Even though some sources were polarizing, the average political leaning of outlets cited by ChatGPT was close to 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.

However, the researchers caution that ChatGPT’s idea of “diversity” doesn’t always mean actual informational variety. The AI may simply highlight sources that stand out linguistically from mainstream outlets.

The study also highlights how little is known about ChatGPT’s internal processes. OpenAI does not explain what causes the differences between the web interface and the API, so users have to make their own judgments. The researchers add that these results are just a snapshot, as OpenAI regularly changes the system without notice.

This fits a broader trend: 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, since they now answer nearly every query instead of sometimes declining when facts are unclear.

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • ChatGPT gives different news source recommendations depending on how it is accessed: while the web interface favors established and licensed media, the API tends to suggest Wikipedia and more niche sites.
  • When asked for greater diversity, ChatGPT includes a wider range of sources, but this can also mean listing politically biased, propagandistic, or even fictitious sites.
  • Although the political orientation of its recommendations is generally balanced, the study points out the lack of transparency in how sources are chosen and urges users to examine the results critically.
Sources
Jonathan writes for THE DECODER about how AI tools can improve both work and creative projects.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.