Disclosure: DROZlegal publishes this explainer and builds a practice-automation product for Canadian law firms. This guide is written to be useful whether or not you ever evaluate DROZlegal — it earns one contextual mention below, in a sentence that would hold up if a competitor published it.
What generative AI actually is, without the hype
Generative AI is software trained to predict the most statistically plausible next word, given everything that came before it. Feed it a prompt and it generates new text, one likely token at a time, based on patterns it learned from a vast body of writing during training. It doesn't "know" your file, your jurisdiction, or your client — it composes language that sounds like a competent answer, whether or not the underlying facts are correct.
Think of it less like a database you query and more like an extremely well-read colleague working entirely from memory, with no access to your matter, who will still answer confidently even when they're guessing. That's the mechanism worth holding onto through everything below: fluent output is not the same thing as verified output.
Generative AI for lawyers vs. traditional legal-tech automation
Most software your firm already runs is deterministic: a document-assembly template takes the data you enter and produces the same output every time, because it's following a fixed rule, not composing anything new. Generative AI is probabilistic: the same prompt can produce different wording — occasionally a different substantive answer — on separate runs, because it's generating plausible language rather than executing a rule.
| Traditional legal-tech automation | Generative AI | |
|---|---|---|
| How it works | Fills a fixed template or rule with the data you provide | Composes new text from a learned pattern of language |
| Predictability | Same input, same output, every time | Same prompt can vary in wording or substance |
| Where errors come from | Bad data you entered | Fluent-sounding text that's still factually wrong |
| Best use | Repetitive, rule-bound documents (NDAs, standard notices) | First drafts, summaries, and research starting points a lawyer verifies |
Where generative AI genuinely helps today
Stripped of the hype, three use cases hold up in daily practice:
- Drafting assistance. A first pass at routine correspondence, a memo skeleton, or a plain-language summary of a document — something you'd otherwise start from a blank page, now started from a rough draft you edit.
- Summarization. Long discovery productions, deposition transcripts, or medical records condensed into key points, so a lawyer's first read is targeted rather than starting cold.
- Research triage. Surfacing candidate issues, arguments, or authorities worth investigating — a starting map of where to look, not a citation list you file without checking.
Every one of those three shares the same shape: generative AI produces a starting point, and a lawyer's judgment turns it into something reliable. That pattern is the whole story of where this technology fits — and where it stops.
Where it breaks: hallucinated citations and no independent judgment
Hallucination is not a rare glitch — it's a structural property of how these models generate text. Because the model is optimized to produce plausible-sounding language rather than verified fact, it can generate a case name, a quotation, or a statutory citation that looks exactly like a real one and simply doesn't exist.
This isn't theoretical. In Whiting v. City of Athens, Tennessee, a U.S. Sixth Circuit panel sanctioned attorneys $15,000 each in punitive damages — plus fees, double costs, and a disciplinary referral — in a March 2026 ruling, after finding that a brief contained more than two dozen fabricated or misrepresented citations.
A public database maintained by researcher Damien Charlotin had tracked more than 1,400 court cases where courts addressed AI-generated errors in a filing as of May 2026 — a count that had grown to nearly 2,000 by August 2026. Damien Charlotin, AI Hallucination Cases Database, 2026
The second limitation is quieter but just as important: generative AI has no independent judgment. It doesn't know your client's risk tolerance, doesn't owe anyone a duty of loyalty, and doesn't weigh the strategic cost of one argument over another. Legal judgment stays with the lawyer no matter how fluent the draft reads — the tool can accelerate the first ninety percent of a task and contribute nothing to the decision that matters most.
Your professional-conduct obligations don't change
For Ontario lawyers, using generative AI doesn't create a new rulebook to learn. The Law Society of Ontario's guidance confirmed that three existing duties — competence, confidentiality, and supervision — already cover it: you have to understand what a tool can and can't do before relying on it, know how it handles client information before that information goes in, and supervise any AI-assisted work the way you'd supervise a person's. Our full explainer on the LSO's generative AI guidance walks through what those three duties require in day-to-day practice, plus the 2026 Legal Aid Ontario attestation now tied to it.
If you're comparing how firms across the country are actually adopting these tools rather than just what they are, our look at AI adoption in Canadian firms covers the 2026 numbers and the courts' own emerging AI rules. Both posts sit inside the Lawyer AI Academy, DROZlegal's free curriculum for lawyers learning this material from the ground up.
Starting without the risk
The safest way to use generative AI in practice is to keep the drafting fast and keep the judgment human — treat every output as a first draft from a very well-read colleague who has never seen your file, not as a finished answer. Tools that route generative output through a named, audit-tracked workflow rather than a bare chat window make that easier to enforce: DROZlegal's agents, for instance, prepare drafts on a job queue with a visible lifecycle and route anything client-facing to a lawyer's approval gate, rather than sending or filing anything unsupervised.
Frequently asked questions
What is generative AI for lawyers? Generative AI for lawyers is software that produces new text — drafts, summaries, or first-pass research — from a plain-language prompt, rather than executing a fixed template or rule. It's built on large language models trained to predict plausible-sounding language, which is why its output reads fluently even when it's factually wrong.
How is generative AI different from the automation already in my practice-management software? Traditional legal-tech automation is deterministic: the same template and the same data produce the same document every time. Generative AI is probabilistic: the same prompt can produce different wording — or a different substantive answer — on separate runs, which is why its output needs the same independent review you'd give a junior associate's first draft.
Can I trust generative AI's case citations? Not without checking them yourself — generative AI has produced fabricated case citations that read exactly like real ones, and courts have sanctioned lawyers for filing them. A U.S. Sixth Circuit panel fined two attorneys $15,000 each in punitive damages in a March 2026 ruling over a brief with more than two dozen fake or misrepresented citations, and a public database tracked over 1,400 similar court cases as of May 2026 — a count that had grown to nearly 2,000 by August 2026 (Damien Charlotin, AI Hallucination Cases Database).
Do Ontario lawyers need to follow new rules to use generative AI? No new rules — the Law Society of Ontario's guidance confirmed that existing competence, confidentiality, and supervision duties already cover generative AI use, and Legal Aid Ontario roster lawyers must now attest annually that they comply with it. See the full explainer for what those duties require day to day.
Not ready to subscribe? Join the DROZlegal waitlist instead.