IAgen sem alucinação

segurança no processo judicial

Abstract

A review conducted by Oxford researchers, published in Nature, investigates the phenomenon of hallucinations in machine learning-based artificial intelligence systems, with particular emphasis on IAgen. The researchers identified that approximately 80% of hallucinations stem from a probabilistic entropic framework preceding their emergence. This subset was designated as “confabulations.” They proposed a heuristic and procedural optimization to be implemented within the algorithm—namely, the incorporation of a clustering function based on bidirectional implication—to assess whether the entropy is semantic in nature. This approach enables both the prevention and mitigation of the issue.

Furthermore, from a methodological standpoint, the authors outlined additional sources of hallucinations that were not addressed within the scope of the research,  but are nonetheless relevant for legal and regulatory considerations. These factors are extrinsic to the algorithm itself and pertain to the management of training datasets and the oversight of training supervision protocols.