Spectres of the Algorithm -AI Hallucinations between Failure of Meaning and Possibility of Imagination

This article examines so-called AI "hallucinations" and questions the common binary that treats them either as a bug to be eradicated or as a machine imagination to be celebrated. It argues that the single word covers two distinct phenomena: factual fabrication in language models, a failure fuelled by training and evaluation incentives, and latent-space drift in image and sound models, an aesthetic potential that artists exploit. It then examines a borderline case where the two meet, "resonant hallucination": sentences falsely attributed to real authors that nonetheless read as plausible extensions of their thought. Resonance, the article contends, is no criterion of truth, and such texts become legitimate only when disclosed. All of these are distinguished from the "algorithmic unconscious", that is, models' bias in what they show and erase. Hallucination becomes a danger when it touches memory, archives and local identity, for which the article proposes models grounded in local archives and evaluated with criteria that reward calibrated abstention. It becomes a creative resource when framed and disclosed by a "curatorial engineer". The cultural and economic value of "error" lies not in its strangeness but in its disclosure: declared error is an aesthetic; hidden error is deception.