The Ethics of Using AI in Legal Work

A generative model will happily write you a factum. It will do it in seconds, in confident prose, complete with citations. The problem is that some of those citations do not exist, and confidence is not the same thing as accuracy. That gap, between what the tool produces and what you are actually allowed to file, is where the ethics of AI in legal work live.

None of the old rules went away when the new tools arrived. Competence, confidentiality, candour to the tribunal, supervision, and honest billing are still the whole game. AI just gives you faster ways to get each of them wrong. So the useful question is not "may I use AI?" You almost certainly may. The question is what your existing duties require once you do.

Competence did not become optional

Every provincial law society expects you to provide competent representation. Nobody wrote "unless a chatbot did the first draft" into that rule. If you put your name on a document, you own every word, every case, and every proposition of law in it.

That means the model is a starting point, never an ending one. It can surface an argument you had not considered or restructure a clumsy paragraph. It cannot verify that the authority it cited is real, still good law, or from the right jurisdiction. There have already been lawyers, in more than one country, sanctioned for filing submissions built on cases the AI invented. The tool did not get disciplined. The human who signed did.

A workable habit: treat AI output the way you would treat a memo from a bright but unsupervised summer student. Useful, occasionally brilliant, and absolutely not something you would file without reading every line and checking every citation yourself.

Confidentiality is where most people slip

Here is the quiet risk. When you paste a client's facts into a public AI tool, you may be handing that information to a third party whose servers you do not control and whose data practices you have not read. Depending on the product and its settings, your input could be stored, reviewed by humans, or used to train future versions of the model.

Solicitor-client privilege is not something you want to test against a vendor's terms of service.

A partner I know puts it plainly to every new articling student: if you would not email it to a stranger, do not paste it into a tool you have not vetted.

Before you feed anything sensitive into a system, ask a short list of questions:

• Where does the data go, and who can see it? Consumer tools and enterprise tools often have very different answers.
• Is my input used to train the model? Many business tiers let you turn this off. Many free tiers do not.
• Has my firm approved this tool? If there is a policy, follow it. If there is no policy, you may be the person who prompts one.
• Can I strip identifying details first? Sometimes you can get the same drafting help from anonymized facts.

The safe default while you are still learning the ground: keep real client information out of any tool your firm has not formally cleared.

Candour to the tribunal, and to your own client

You owe the court honesty, and you owe your client honest work. AI complicates both in a subtle way, because the output is so fluent that it feels finished. Fluency is a trap. A paragraph can read beautifully and still misstate the law.

If a submission relies on authority, you confirm that authority exists and says what you claim it says. Full stop. The same goes for advice you give a client. Running a question through a model and relaying the answer without your own judgment is not legal advice. It is repetition, and the client is paying for judgment.

There is also a growing conversation about disclosure: whether you should tell a court or a client that AI was involved. Some courts have started issuing practice directions on exactly this. The specifics vary and they are changing quickly, so check the current rules of the court you are appearing before rather than assuming last year's answer still holds.

Supervision cuts both ways

If you are an articling student or a junior, AI can make you look faster and more polished than your experience alone would. That is fine, as long as your principal understands how the work was produced. Hiding your process is a bad instinct in a profession built on trust, and it robs you of the feedback that turns a junior into a lawyer.

If you are the one supervising, the duty is yours too. You cannot delegate judgment to a model any more than you can delegate it to a junior and then not read their work. Someone competent has to own the final product, and that someone is a person.

This is also a career point, not just a compliance one. The articling students who thrive are the ones who use these tools to do more thinking, not less. If a partner senses that AI is doing your reasoning for you, that is a problem. If they see you using it to test arguments and free up time for harder analysis, that is a superpower. If you are still working out how AI fits into day-to-day practice, our overview of AI in legal practice in Canada is a good next read, and it pairs well with our roundup of legal tech tools worth knowing.

Billing honestly when the work takes minutes

The billable hour and the speed of AI are on a collision course, and the ethical answer is the boring one: you bill for value delivered and time actually spent, not for time you would have spent before the tool existed.

If a task that once took three hours now takes twenty minutes plus careful review, you cannot bill three hours because that is what it "used to" cost. That is not a grey area. Some firms are moving toward flat fees or value-based pricing partly because AI makes the hourly model harder to defend with a straight face. Whatever your firm does, the underlying rule against overbilling has not moved an inch. If you want the mechanics, our explainer on the billable hour walks through how the model works and where it strains.

A short, practical checklist

When you reach for an AI tool on a real matter, run through this before you rely on the output:

1. Verify every fact and citation the model produced, from primary sources.
2. Confirm the tool is approved for the sensitivity of the information you are entering.
3. Keep client identifiers out of anything not formally vetted.
4. Apply your own judgment to the substance rather than passing the output along.
5. Bill honestly for the time and value, not the old rate card.
6. Check current court rules on AI use and disclosure for your jurisdiction.

None of this is exotic. It is the same professional responsibility you already signed up for, applied to a faster pencil.

The tools change, the duties don't

It is tempting to treat AI as a special case with its own rulebook. It is not. The Federation of Law Societies and the provincial regulators have been clear that existing obligations apply to new technology, and the Federation of Law Societies of Canada is a reasonable place to watch for guidance as it develops.

The lawyers who come out of this era well will not be the ones who avoided AI, and they will not be the ones who trusted it blindly. They will be the ones who used it hard and checked it harder. That posture, curiosity paired with skepticism, is worth building now, early in your career, because these tools are only going to get more capable and more tempting.

If you are early in the search and thinking about the kind of practice you want to build, browse current articling and early-career roles with this in mind, and keep an eye on the rest of our legal tech coverage as the ground keeps shifting. The firm you join will have opinions about these tools. Having your own, grounded in the duties above, is how you show up as a lawyer rather than a user.