Training juniors when the AI writes the first draft
A first year solicitor hands you a 900 word advice on a commercial lease dispute at 4pm. It is well structured, it cites four cases, and the tone is right. You ask why they preferred one authority over another and the answer is a shrug. They did not choose it. The tool did, and the draft read well enough that nobody had a reason to argue with it.
Every partner who has supervised a graduate through their first two years knows the drill: give them the memo, let them flounder, mark it up in red, repeat. That loop worked because building a document from nothing forces you to read the whole file, hunt for the authority, and commit to a position you then have to defend. The output was worth maybe a third of the exercise. The other two thirds were the training.
The tools are not going back in the box, and there is no serious argument for making juniors type slowly on purpose. The question is narrower and more useful: if the machine now does the part that built judgment, what do you hand the junior instead?
The first draft was the training, not the deliverable
Firms have described first drafting as low value work for a decade, mostly because it was billed at a rate that embarrassed people. That framing was always wrong about what was happening. A junior building an advice from scratch has to open every document in the file, work out which facts matter, find authority that fits those facts rather than the general topic, and then put their name to a conclusion. Reviewing a draft that already exists asks for none of that.
Reviewing is a real skill, but it is a different one. It teaches you to spot what looks off in a document that is already coherent. It does not teach you how a coherent document gets built, which is the skill you need at year five when a client rings with a mess and there is nothing to review.
The four skills that erode first
If you want to know what to replace, be specific about what is disappearing. In practice it is four things, and they do not go at the same speed. Reading the file properly is the first casualty because the draft already contains a plausible summary of the facts. Committing to a position is the last, and it is the one that shows up in front of a client or a judge.
None of these are recovered by telling a junior to be careful. They are recovered by giving them tasks that cannot be completed without doing them.
Give the junior the half the tool cannot do
The fix is a sequence, not a policy document. Put the junior in front of the question before anything is generated, and put them in front of the sources after. The generation sits in the middle where it belongs, and it stops being the thing they are assessed on.
The practical shift is that you stop reviewing their prose and start reviewing their markup. If a junior hands you a clean draft you have learned nothing about them. If they hand you a draft with every proposition annotated to a source, and a short note on what they rejected, you can supervise properly in ten minutes.
Verification is a skill, so teach it like one
Most firms have told their juniors to check the citations. Very few have told them what checking means, and the difference matters. The failure that gets firms into trouble is rarely an invented case name, which is easy to catch. It is a real case, correctly cited, attached to a proposition it does not support, or a real case that has since been overturned or distinguished into irrelevance.
The rule that works is blunt: nothing goes in unless the junior has opened the judgment or the section, and can point to the paragraph. Not the headnote, not a summary of the case, the passage itself. Make them paste the quote into the markup. It takes longer for the first month and then it becomes the way they read.
Rebalance where their time goes and say so out loud, because a junior who is measured on output speed will quietly stop verifying the moment they get busy.
What supervision looks like on paper
Supervision obligations did not change because the drafting tool did. The solicitor who signs the advice or files the document owns every word in it, and a firm that cannot show how a junior's work was checked is exposed twice over, to the client and to the regulator. Several Australian superior courts have now published practice notes or guidelines dealing with generative AI in litigation, and the requirements differ between courts and between jurisdictions. Check the current one for the court you are actually in before you file, rather than relying on what a colleague told you last year.
Internally, the useful move is a written progression with dates on it, so that supervision is a stage a junior moves through rather than a permanent condition. Something like this.
- Junior states which parts of the work were AI assisted, in writing, on every piece
- Every authority verified against the primary source before it leaves their desk
- Supervisor pulls two citations at random on each piece and reads the passage
- Junior submits the markup and the disagreement note with the draft, always
- Junior keeps a running source log per matter that the supervisor can audit
- Supervisor reviews the log rather than re-checking every citation
- Junior explains why they rejected a passage, not just that they did
- Spot checks continue but drop to one piece in three
- Junior sets the research plan unaided and defends the scope of it
- Review focuses on conclusions, risk framing and what was left out
- Junior starts supervising the verification of someone more junior
- Sign off responsibility stays with the supervisor, and everyone knows it
Choose tools that make checking cheap
All of this depends on one thing being true: that verifying a proposition is fast. If a junior has to leave the draft, open three tabs, work out which version of the Act was in force at the relevant time, and hunt through a judgment for the paragraph, verification will get skipped under deadline pressure. Not because they are lazy, because it costs twenty minutes a citation and they have a 5pm.
So the tool choice is a training decision, not just a procurement one. A research tool that ties each proposition to the Australian primary source and shows you the passage it relied on turns verification into a ten second habit instead of a twenty minute detour, which is the difference between a rule juniors follow and a rule they nod at. That is the whole design principle behind Legal Brain, and it is worth applying as a test to whatever you end up using: can a first year check this in under a minute, and can they see what the answer was built from?
The firms that come out of this well will not be the ones that banned the tools or the ones that handed them over without a word. They will be the ones that noticed which part of the old apprenticeship was load bearing, and deliberately built something to carry the weight.
Frequently asked questions
Should junior lawyers be allowed to use AI for first drafts?
Yes, with a defined sequence around it. Have the junior write the issue statement and the research plan before anything is generated, and require them to verify every proposition against the primary source afterwards. Banning the tools mostly produces hidden use, which is far worse than supervised use.
How do you train a graduate lawyer if AI does the research?
Shift what you assess. Stop grading the prose and start grading the research plan, the source markup and a short note on what the draft got wrong. Those three artefacts show you their reasoning, which is the thing you are actually trying to develop.
Do Australian courts require disclosure of AI use in filed documents?
Several Australian superior courts have issued practice notes or guidelines dealing with generative AI in litigation, and some impose disclosure or verification requirements. The obligations are not uniform across courts or jurisdictions, so check the current practice note for the specific court before you file rather than assuming a general rule.
Who is responsible if an AI assisted draft contains a bad citation?
The solicitor who settles and signs the document. The tool is not a party to anything and cannot hold a practising certificate. That is also why the verification step belongs to a named person with a date against it, not to a general firm policy.
Two quick questions
No score is stored. Pick an answer to see why it is right.
-
1A junior hands you an advice citing a real Australian case for a proposition the judgment does not actually support. What is the most useful supervisory response?
The failure is not carelessness with names, it is accepting a proposition without reading the source. Requiring the quoted passage for every citation makes that impossible to skip and builds the habit across the whole document, while fixing it yourself or adding a proofread leaves the underlying reading gap untouched.
-
2In a supervised workflow, which task should the junior complete before any draft is generated?
Framing the question and naming where the answer should live is the part the tool cannot do for them, and it is the step that reveals whether they have read the file. Doing it first also gives you a benchmark to test the generated draft against instead of assessing it on its own terms.
Research Australian law without handing over client data
Legal Brain searches Australian legislation and case law, shows you the source behind every answer, and anonymises client-identifying detail before anything reaches a model.