The number of cases involving hallucinations is on the rise. In R (Ayinde) v London Borough of Haringey [2025] EWHC 1383, hallucinated cases, or “fake authorities,” were regarded as professional misconduct. The Court lost its patience with the use of AI by technically uninformed counsel and signaled the grave consequences:  “Placing false material before the court with the intention that the court treats it as genuine may, depending on the person’s state of knowledge, amount to a contempt. That is because it deliberately interferes with the administration of justice.”  In 2026, it is no longer possible to feign ignorance of the fact that Large Language Models, or “LLMs,” cannot be trusted and that their outputs require meticulous verification. Although the problem of hallucinations was supposed to be fixed in “the near future,” in late 2025 OpenAI finally admitted hallucinations were an inevitable side-effect of text generation. Hallucinations are here to stay: LLMs will continue to produce plausible fluent and seemingly helpful text that will be ridden with inaccuracies and falsehoods. Some confusion is introduced by the excellent benchmark performance of many LLMs and the incessant use of anthropomorphic terms like “understand” or “reason” by AI vendors. Does benchmark performance, we must ask, translate into real-world tasks? Can AI vendors be trusted?   Additional problems result from the lack of a universally accepted definition of hallucinations. Consequently, there is no reliable way of measuring their occurrence and evaluating the reliability of specific LLMs or LLM-based tools.

The seminar addresses the following questions:

• What is the legal landscape surrounding hallucinations? How are judges in various jurisdictions reacting to this phenomenon?
• Are there ways of counteracting hallucinations? Are some LLMs more reliable than others?
• What are the legal tasks particularly prone to invoke hallucinations?
• To what extent, if any, can lawyers rely on any of the benchmarks measuring the capabilities of LLMs?

The seminar will also revisit some technical concepts, such as “context engineering” and “retrieval augmented generation” that are indispensable to better leverage the capabilities of LLMs and LLM-based tools.