Generative artificial intelligence is increasingly being used in the preparation of scientific publications, including in the selection of a topic, the development of a structure, the drafting of an abstract, language editing, the formatting of bibliographic references and the preliminary assessment of a manuscript. However, alongside the practical advantages, the question arises as to the reliability of the generated outputs. These may be well-structured and convincingly formulated, but may contain factual inaccuracies, omissions, incorrect sources or changes to the meaning of the scientific text. This paper proposes a methodology for assessing the reliability of outputs generated by artificial intelligence systems in the preparation of scientific publications. The methodology is based on a multi-criteria assessment and includes criteria such as factual accuracy, relevance to the task, verifiability of claims and sources, preservation of meaning and terminology, completeness of the output, and the need for human correction. An overall reliability index is proposed, which assists the author in deciding whether a given output should be accepted, corrected or rejected.

generative artificial intelligence,AI-generated outputs,reliability assessment,scientific publications,multi-criteria assessment,academic writing,human oversight