When the production of science speeds up, means for ensuring trust must catch up

Imagine a research paper that seems to appear almost instantly. The language is polished, the argument looks plausible, the references are neatly arranged. But then one of the cited sources does not exist. Another one says something different from what the paper claims. Somewhere in the process, speed has started to take priority over quality.

This was one of the problems Prof. Dr. Johannes Breuer brought into focus during his visit to the Graduate School of the Research Center Trustworthy Data Science and Security (RC Trust) in Duisburg. In his talk, “AI and trust in research in the social and behavioral sciences,” Breuer examined a question that is becoming increasingly difficult to avoid: What happens to trust in science when artificial intelligence becomes an integral part of the way science is produced?

Breuer did not approach AI only as a technology to be studied from the outside. He described three relationships between AI and science: AI is a product of scientific research,as a research subject,  and as a tool for research. This third role was also the focus of his talk.

The role of AI as a tool for research is expanding quickly. AI systems can support all steps of the research process: idea generation, literature search, study design, data collection, data processing, analysis, writing, publishing and quality control. They can edit language, generate code, help structure data, summarize literature or assist with reproducibility checks. For Breuer, this development raises a central issue for trustworthy research. AI can make scientific work faster, more efficient and, in some cases, better documented. But it can also introduce new uncertainties. If researchers rely on AI systems whose training data, model behavior or guardrails are not fully transparent, it becomes harder to understand how outputs came into being. If AI-generated text contains hallucinated citations, the problem is not simply cosmetic. It affects the credibility of scientific claims. And if AI tools are used in ways that are not disclosed, reviewed or verified, they may weaken the very standards they are supposed to support.

Breuer structured this tension around three dimensions: transparency, research quality, and replicability and reproducibility.

Transparency concerns the question of what can be understood and disclosed. Which model was used? For what purpose? With which prompts or settings? Was the system open, or was it provided by a commercial company whose internal workings remain inaccessible? In traditional statistical procedures, researchers can usually describe how a result was produced. With large language models, that chain of explanation is often much harder to reconstruct.

Research quality is more ambivalent. AI can help improve texts, detect coding errors or support data documentation. It may also open up new possibilities for working with large amounts of text or complex research materials. At the same time, it can produce fluent but unreliable outputs. A paper may read well while containing false references, weak arguments or unverified claims. In that sense, AI challenges one of the familiar signals of quality: polished writing no longer necessarily means careful work.

The third dimension, replicability and reproducibility, is especially important for the social and behavioral sciences. Breuer pointed to long-standing concerns in these fields: data and code are often not shared, not sufficiently documented or difficult to reuse. AI can intensify this problem when research depends on opaque, changing or non-deterministic systems that may also not be accessible for everybody. But it may also become part of the solution. AI tools can help check code, structure documentation, compare reported results with data and code and support automated reproducibility assessments.

This double role made the talk particularly relevant for the Graduate School. Doctoral researchers are not only learning what to study; they are also developing the habits, standards and practices that will shape their work as researchers. Breuer’s talk invited them to look at AI not as a distant disruption, but as something that already affects everyday academic practice: how to search, write, check, , and review research.

The practical message was not to reject AI altogether. Breuer argued for a careful and transparent approach: researchers should be mindful about when they use AI, explain what they used it for, verify its outputs and prefer open models where possible. Proprietary systems may be useful, but their use in academic research should not be treated as self-evident. It requires justification.

Prof. Dr. Johannes Breuer is Professor of Digital Social Science at the University of Duisburg-Essen, Institute of Political Science, and heads the team Research Data & Methods at the Center for Advanced Internet Studies (CAIS) in Bochum. His research focuses on the uses and effects of digital media, digital trace data, computational methods open science, and meta-science.

His visit reflected a core aim of the RC Trust Graduate School: to involve doctoral researchers in current academic debates, expose them to interdisciplinary perspectives, and connect them with experienced scholars from different fields. In this case, the discussion went to the heart of trustworthy research itself. AI may accelerate science, but trust still depends on something slower: transparency, documentation, verification, and responsible scientific practice.