Generative artificial intelligence - a technology capable of creating texts, images and other materials based on trained models - is being used in work ever more actively. Journalists, marketers and communications specialists use it for drafts of news items, corporate messages and advertising materials. However, such tools have a hidden problem: they can replicate distortions that are invisible at first glance.
Mugur Jana, a professor of journalism and mass communications at the University of Kansas, devoted a chapter in the book "Sustainable AI - AI for Sustainable Development: A Collection for Public Relations, Communications and Sustainable Development Specialists" (published by Quadriga University of Applied Sciences in Berlin, freely available on the Zenodo platform) to how to recognize AI bias and avoid perpetuating it in one's work. The book is a collection of views from authors around the world and was conceived as an experiment in using AI in publishing.
The chapter is based on the concept of epistemic injustice, introduced by Miranda Fricker, a scholar at the University of London. Epistemic injustice is a situation in which someone's knowledge or experience is systematically ignored or devalued. Jana analyzes how large language models (AI systems trained on vast arrays of text) can reinforce such injustice. According to him, training data do not include the histories and experience of many communities because those have never been digitized. "The corpus we use to train AI either has no access to all these sources of knowledge or chooses to ignore them. I think this is a serious problem that AI faces today," the researcher notes.
To help specialists avoid the traps of bias, Jana proposes a checklist of four dimensions: credibility of evidence, completeness of interpretation, representation (visual and narrative) and institutional responsibility. The checklist contains questions that help identify potential distortions, warning signs and ways to avoid them. It is important to understand: this is not about proven harm, but about a statistical association between what data were used for training and what answers the model produces. Correlation by itself does not prove causation, but it points to the need for critical verification.
Jana emphasizes: AI is trained on human materials, and those are not flawless. Therefore everything AI creates must be critically checked - it is a synthetic entity that bears no responsibility for its actions. At the University of Kansas School of Journalism, they adhere to the principle of "human at the beginning, human at the end": AI must remain no more than a tool. In the conclusion of the chapter, the author writes: the task of communicators is not to abandon the tools, but to stop taking the fluency of their speech as proof of accuracy and to develop the habit of systematically asking whose voice and whose knowledge made it onto the page.