The rapid adoption of artificial intelligence across the legal profession has sharpened a difficult question: when an AI tool produces flawed work, who carries the liability? The prevailing view is unambiguous. Responsibility remains with the lawyer who reviews and approves the output, not with the software or the company that supplied it. However capable these systems become, the professional standards governing accuracy and defensibility have not shifted, and firms that treat automation as a substitute for judgement are exposing themselves to significant risk.
Responsibility Rests With The Lawyer
Courts and regulators are not principally concerned with whether a piece of software made a mistake. They assess whether the lawyer exercised proper professional judgement. This distinction reframes the debate around what has been described as a "human on the hook" rather than merely a "human in the loop". The point is that accountability cannot be delegated to an AI system, no matter how sophisticated its capabilities. A lawyer who signs off on AI-assisted work is answerable for it on precisely the same terms as work produced by any other method, and the burden of ensuring that it can be defended on its merits does not transfer to the tool.
Cautionary Examples From Practice
Recent incidents illustrate how the problem has surfaced in high-profile settings. A leading international firm acknowledged that AI errors, including misquoted law and incorrect case citations, had appeared in a filing before a United States bankruptcy court, and that its own internal AI policies had been breached in the process. The issue is not confined to litigation. A major professional-services firm in Canada withdrew an AI-generated cybersecurity report after researchers identified fabricated citations and references, demonstrating that fabricated or unreliable output can just as easily undermine corporate advisory work. These episodes suggest that the risk is present wherever AI is used to generate substantive content.
The Limits Of Superficial Review
A recurring theme is that light-touch oversight offers little protection. A cursory review of AI output does not reduce risk; if anything, it formalises it by lending the appearance of scrutiny to work that has not genuinely been checked. This concern is heightened by the emergence of so-called agentic systems, which can carry out multi-step tasks with limited human intervention. Such tools make meaningful oversight harder to sustain, precisely at the point where it matters most. The legal standard for accuracy and defensibility has not relaxed to accommodate the technology, and practitioners are expected to understand the tools they rely upon and to apply independent judgement to whatever those tools produce.
Building Defensible Practice
For firms, the practical implication is that AI should be integrated with clear policies, adequate training and a culture in which verification is treated as non-negotiable. Lawyers need to know how their tools work, where they are likely to fail, and how to interrogate the results. Work must be capable of standing on its own merits regardless of how it was generated, and internal policies are of little value if they are not observed in practice.
The wider implication is that AI does not dilute professional accountability; it concentrates it. As these tools spread through firms of every size, the profession's existing duties of competence and diligence remain the decisive test, and those who master both the technology and its limits will be best placed to use it safely.
