What Lawyers Get Wrong About Artificial Intelligence

Artificial intelligence arrived in the legal profession accompanied by an unusual quantity of noise. Every conference acquired a panel on it, every software supplier discovered that its product contained it, and perfectly sensible lawyers divided themselves into those announcing that the profession was about to be transformed beyond recognition and those pointing out that they had managed without ChatGPT for thirty years and proposed to continue doing so. Both positions are comforting because neither requires much thought.

The first mistake is to treat AI as though it were a person. The language encourages this. We say that a system “thinks”, “knows”, “understands”, “decides” or “hallucinates”. Sometimes that language is useful shorthand, but it can easily become an explanation rather than a metaphor. The practical problem is not whether a large language model possesses some inner state which philosophers will eventually agree to call thought. A barrister does not need to settle the nature of consciousness before asking a machine to summarise a bundle. The useful questions are more mundane: what material has it seen, what has it been asked to do, what sort of output is it capable of producing, where is it likely to go wrong and how will the result be checked?

The second mistake is to confuse fluency with authority. Modern AI is extremely good at producing the outward appearance of competence. It explains, summarises and drafts in confident prose, which makes it particularly seductive to a profession in which confident prose has traditionally been regarded as evidence that somebody knows what he is talking about. Unfortunately, the machine can produce the form of a good answer without possessing the foundations which make the answer reliable. An invented case with a neutral citation still looks very much like a case until somebody takes the elementary precaution of opening it.

That leads to the third mistake, which is to confuse assistance with delegation. There is nothing particularly alarming about using a machine to prepare a provisional chronology, compare two documents, extract dates, suggest lines of enquiry or produce a first structure for a draft. Lawyers have always used other people and other tools to do preparatory work. What matters is whether the person ultimately responsible for the work understands what has been done and is capable of checking it. A pupil’s draft does not cease to be counsel’s responsibility because the pupil typed it. The same proposition does not become more complicated merely because the junior has been replaced by software.

The fourth mistake is to assume that the choice is between adoption and rejection. It is perfectly possible for a lawyer to use AI extensively for one class of task and not at all for another. I am content to let a machine reorganise a table, extract dates or suggest alternative structures for an argument. I am considerably less interested in allowing it to decide whether a witness is truthful, whether an allegation ought to be pleaded or whether a client’s commercial interests are served by pursuing a point simply because it is legally available. The line is not between “AI work” and “human work”. It runs through individual tasks.

The fifth mistake is to imagine that the existence of the technology settles the question whether it ought to be used. Sometimes the answer will plainly be yes. Sometimes the cost of checking the machine’s work removes most of the supposed efficiency. Sometimes confidentiality makes a particular system inappropriate. Sometimes a junior lawyer ought to do the difficult work personally because the purpose of the exercise is not merely production but learning. A machine which can prepare a first draft in thirty seconds is of limited educational value if the result is that nobody in the team ever learns how to prepare one without assistance.

None of this is an argument for technological asceticism. AI is already useful and will become more so. Refusing to use it because it sometimes makes mistakes makes no more sense than refusing to use an internet search because the internet contains rubbish. The sensible response is to understand the tool sufficiently well to know when it is useful and when its apparent convenience is merely moving risk somewhere less visible.

The legal profession does not need to worship artificial intelligence, and it does not need to fear it. It needs to become competent in using it. That is a less dramatic conclusion than either salvation or extinction, which may explain why it receives rather less attention.