Will AI Replace Lawyers?

Every few months a headline announces that AI has passed the bar exam and the legal profession is finished. Then you look at the employment projections and find the opposite: the Bureau of Labor Statistics expects the number of lawyers in the United States to grow 5 percent between 2025 and 2035, faster than the average for all occupations, adding about 40,600 positions to a base of 863,700.

So the headline is wrong. But the comfortable conclusion, that law is safe, is wrong too, and you can see why in the same dataset. Look one rung down the ladder, at the 404,900 paralegals and legal assistants, and the projection is zero percent. Not slower growth. Zero, and a net loss of 1,100 jobs over the decade.

That gap is the story. AI is not coming for the person who signs the advice. It is coming for the work that sits underneath them, and the BLS says so in as many words.

The two lines, side by side

These are two occupations in the same buildings, on the same matters, often at the same desks. One is projected to grow faster than the economy. The other is projected to shrink slightly while the economy grows around it.

“Demand for these workers is expected to be limited by advances in technology, including artificial intelligence (AI). These technologies are expected to make paralegals and legal assistants more efficient at tasks such as conducting research and preparing documents, which may reduce demand for these workers.”

What AI already does well in law

The tasks being absorbed have a shape in common. They are high volume, they have a checkable right answer, and they were already being done against a written record rather than in a room with a person.

Document review in discovery is the clearest case. Sorting hundreds of thousands of documents for relevance and privilege was once the definition of junior legal work, billed by the hour. It is now largely a retrieval and classification problem, which is exactly what these systems are built for.

Legal research follows the same logic. Finding the cases that bear on a question, checking whether they are still good law, and summarising what they held is work with a verifiable answer, and the verification step is what makes it safe to automate: a drafted memo can be checked against the cases it cites.

First drafts of standard instruments (NDAs, employment agreements, leases, routine motions) are the third. The structure is known, the clauses are largely conventional, and the value a lawyer adds is in the handful of terms that are not.

What does not move

The parts of legal work that resist automation are not resisting because they are difficult. They resist because of who has to be accountable for them.

A lawyer is a person a regulator can sanction, a client can sue, and a court can hold in contempt. That is not a technical property, and no improvement in model quality transfers it. When a piece of advice is wrong, somebody must be answerable, and the profession is organised around that somebody being a named human being.

Then there is everything that happens in a room. Reading a witness, judging whether a client is telling you the whole story, deciding how hard to push in a negotiation, knowing when a matter is really about a family rather than about a contract. This is the work clients think they are buying when they hire a lawyer, and mostly they are right.

And advocacy: choosing which of five defensible arguments to actually run, in front of this judge, in this jurisdiction, this month. That is a judgement about people, and it is made with information that was never written down.

So where is the risk for an actual lawyer?

Not in being replaced. In two other places.

The first is the training ladder. Junior lawyers have always learned judgement by doing enormous quantities of low-judgement work (reviewing the documents, pulling the cases, drafting the boilerplate) and noticing patterns in it. If that work is done by a machine, the apprenticeship that produced senior lawyers stops running, and nobody has yet shown what replaces it. This is a problem for the profession before it is a problem for any individual.

The second is the business model. A great deal of legal work is billed by the hour, which means a firm’s revenue is tied to how long tasks take. Tools that make those tasks take a tenth as long do not threaten the lawyer’s employment so much as the price of what they sell. Firms that bill for outcomes rather than hours are less exposed to this than firms that do not.

Notice that neither of these is the risk the headlines describe, and neither is addressed by getting better at law.

What to do about it

The honest advice is narrower than the usual list.

The short version

The projections do not say lawyers are being replaced. They say the opposite, clearly and with numbers.

What they also say, just as clearly, is that the work immediately below the lawyer is being absorbed, and the government’s own explanation for the paralegal projection is artificial intelligence. If your day consists mostly of that work, whatever your title, the projection that describes you is the flat line, not the rising one.