Large Language Models - The Limitations & Risks Explored

Introduction

A marketing executive whose AI-generated post contains errors faces embarrassment. A solicitor who files AI-generated grounds of appeal citing five non-existent authorities faces something altogether more serious: potential contempt of court proceedings. As the High Court made clear in Ayinde v Haringey, the consequences for lawyers are fundamentally different and so is the education they require.

This short webinar examines, in clear and practical terms, the ways large language models (‘LLMs’) fail in legal practice and why those failures occur. You will discover the mechanics behind the technology, the ‘clever parrot’ that predicts plausible language rather than reasons and why hallucination is not a defect to be patched away but an inherent characteristic that demands careful supervision.

Drawing on the rapidly developing case law, the webinar explores the lessons from Mata v Avianca, Harber v HMRC, Ayinde, Bandla v SRA (twenty-five fictitious authorities and struck-out grounds), Olsen v Finansiel Stabilitet (where a contempt summons was declined only ‘narrowly and somewhat reluctantly’) and Zzaman v HMRC, in which the tribunal took the unusual step of setting out practical measures to reduce the risk.

But fabricated authorities are only part of the story. The webinar also addresses knowledge cut-offs and why integrated web search is no cure-all; the limits imposed by context windows when analysing lengthy documents; confidentiality, privilege and data protection risks, including the ICO’s position, SRA warnings and recent developments concerning privilege; bias and Equality Act 2010 exposure; and three realistic scenarios showing how AI mistakes can evolve into negligence claims.

By the end of the webinar, you will understand not only the risks, but exactly what effective professional supervision and verification require. You will leave with practical safeguards, a clear understanding of your professional obligations and the confidence to use AI responsibly without exposing yourselves, your clients or your firms to unnecessary risk.

What You Will Learn

The webinar will cover the following:

• How large language models work (next-token prediction rather than reasoning) and why this core mechanism underpins every downstream failure mode
• Hallucination - the underlying mechanics, why models cannot reliably detect their own uncertainty and why legal research represents a particularly high-risk use case
• Case law in depth - Mata, Harber, Ayinde, Bandla, Olsen and Zzaman - and the escalating consequences now in play, including contempt exposure, strike-out, wasted costs and regulatory referral
• Knowledge cut-off limits and how integrated web search mitigates but does not remove the risk, including the new categories of error introduced by search tools themselves
• Context windows - what they do and do not retain and why long-form documents, bundles and data rooms can behave unpredictably when processed by AI systems
• Confidentiality, privilege and data protection - differences between consumer and enterprise tools, training data risks, ICO and SRA expectations and recent developments affecting legal privilege
• Bias in AI outputs and how this can engage Equality Act 2010 duties in advice, drafting and decision-support contexts
• Three concrete loss scenarios - the broken break clause, the liability cap misdraft and the fabricated authority as well as the practical verification discipline that prevents all three

This pre-recorded webinar will be available to view from Tuesday 16th February 2027

Alternatively, you can gain access to this webinar and 2,400+ others via the MBL Webinar Subscription. Please email webinarsubscription@mblseminars.com for more details.