Coherence before Knowledge: An Epistemic diagnosis of AI-driven Hallucinations for Knowledge Management

Generative artificial intelligence is increasingly entering research and development (R&D) processes — be it as a support for idea generation, market interpretation, project evaluation, strategic decision-making and so on. In contexts marked by time pressure, uncertainty as well as informational complexity, AI-generated outputs may appear especially valuable because they offer syntheses that are rapid and explanations that are strikingly plausible. Yet precisely these qualities also make generative AI problematic for R&D management. Its outputs may be fluent, coherent, and strategically persuasive while remaining factually incorrect, weakly grounded, or insufficiently justified. This problem is commonly discussed under the label of AI hallucination, which we hereby define as factually incorrect or ungrounded AI-generated content which is perceived as persuasively plausible, convincing, and decision-ready. The paper contributes to R&D management in three ways. First, it reframes AI hallucinations as epistemic risk rather than merely technical error. Second, it introduces an epistemological distinction between coherence, justification, and knowledge as a conceptual tool for understanding AI-supported decision failures. Third, it proposes a therapeutic, governance mechanism for R&D decision-making that helps managers identify where epistemic risk enters the process and how it may be mitigated before strategic commitments are made.