Professor Mik opened by unpacking the word “hallucination” itself. For humans, it evokes seeing or hearing things that do not exist. For AI, it has come to mean something more mundane but no less troubling: the “confident” production of false or fabricated information. Her examples ranged from the obviously absurd, such as adding non-toxic glue to pizza sauce, to the far more insidious, where the model’s response appears perfectly plausible yet is legally unsound — a discrepancy that would only be detectable by a trained legal professional. The core problem, Professor Mik explained, lies in how LLMs work. Outputs are generated by recombining statistically probable word sequences. LLMs are designed to be helpful for users. However, they are not trained to assess the correctness of their own responses, the legal soundness of an argument or to even respond “I do not know”.
The discrepancy between a model’s answers and their factual accuracy is no longer just theoretical. Professor Mik referred to recent incidents in which lawyers had relied on case law generated by AI that turned out to be entirely fictitious, resulting in very real professional consequences. What may sound like a harmless “hallucination” in marketing terms, she argued, in fact masks a structural defect: the outcome is an imitation of legal reasoning, without any justification that would meet legal standards of explanation and accountability.
The industry’s response has been to tweak rather than transform. Feeding models external documents, linking them to databases, or “making them bigger” promises fewer errors but does not change the underlying logic, namely that the model is still based on frequency, not veracity. Against this backdrop, Professor Mik urged that LLMs should be treated as tools requiring close human supervision rather than as ready‑made junior associates. For now, however, the only reliable safeguards are simple, carefully phrased prompts, consistency in the choice of model and a healthy dose of scepticism.
A warm thanks to Professor Mik for her insightful presentation and to all participants for their thoughtful questions and contributions.
The lecture was organised by Professor Iris Eisenberger and Professor Christiane Wendehorst.
