TL;DR: Machine learning is a branch of AI in which systems learn patterns from data to classify or predict. Generative AI is a type of machine learning that creates new content - text, images, code - rather than only scoring or sorting it. In legal work, classic machine learning powers technology-assisted review and outcome prediction, while generative AI drafts, summarises, and answers questions. Both need human oversight.
The legal-tech market blurs two distinct technologies under one buzzword. A vendor saying its AI "reviews your documents" might mean a classifier trained to sort relevant from irrelevant files, or a large language model writing a summary in fluent prose - very different tools with very different failure modes. Knowing which is which helps you ask better questions, set realistic expectations, and avoid the trap of treating a confident paragraph as if it were a verified result. This article draws the line clearly and shows where each technology earns its place in practice.
What is machine learning?
Machine learning is a branch of artificial intelligence in which systems learn patterns from data to classify or predict, rather than following rules written explicitly by a programmer. You show the system many labelled examples - documents marked relevant or not, contracts tagged by clause type, past cases with their outcomes - and it infers a statistical model that can then label or score new, unseen examples. The defining feature is that the behaviour is learned from data, not hand-coded.
Most machine learning used in law is what specialists call discriminative or supervised: it draws boundaries between categories (relevant or not, high-risk or low-risk) or estimates a number (a probability, a score). It does not write anything new. Ask a classic machine-learning model to review a document set and it returns a sorted, scored pile - this file is 92% likely to be responsive, that one is not - which is exactly what you want for tasks like e-discovery. We cover one major legal application in depth in what is TAR and predictive coding.
What is generative AI?
Generative AI is a type of machine learning that creates new content - text, images, code - rather than only classifying or scoring existing content. The large language models behind today's chat assistants are the best-known example: trained on vast amounts of text, they predict plausible next words and so can draft a clause, summarise a judgment, or answer a question in natural language. The output is generated, not retrieved or merely labelled, which is what makes it feel like writing.
That generative ability is powerful and genuinely new for legal work, but it comes with a characteristic flaw: because the model produces statistically plausible text, it can state things that are fluent and confident yet false - the so-called hallucination, including invented case citations. Generative AI does not know whether its output is true; it knows what looks like a likely continuation. For a deeper look at how these models work and how retrieval can ground them, see legal LLMs explained and RAG for legal AI.
How do machine learning and generative AI relate?
The relationship is one of nesting: generative AI is a subset of machine learning, which is itself a subset of artificial intelligence. Every generative model is a machine-learning model - it learned its behaviour from data - but not every machine-learning model is generative. The classifier that sorts your documents and the language model that drafts your letter are cousins built on the same statistical foundations, differing in what they are trained to produce.
This is why "AI" alone is an unhelpful label. Academic centres and standards bodies both stress precise terminology because the risks differ by type; the Stanford HAI research community has documented how conflating these categories leads to misplaced trust. A classifier's errors are measurable and bounded - it mislabels a document. A generative model's errors can be unbounded and invisible - it can invent an authority that never existed. The two demand different kinds of verification.
What is the practical difference between them?
The practical difference is what each technology produces and therefore how you check it. Discriminative machine learning sorts and scores; generative AI writes. The table maps that distinction onto concrete legal examples.
| Dimension | Machine learning (classify / predict) | Generative AI (create) |
|---|---|---|
| Core task | Sort, score, or predict from examples | Produce new text, images, or code |
| Typical output | A label, a probability, a ranking | A draft, a summary, an answer |
| Legal example | Flag responsive documents; classify clauses; predict outcomes | Draft a clause; summarise a deposition; answer a research question |
| How you verify | Sample and measure recall and precision | Read every output; check each citation against the source |
| Main failure mode | Misclassification you can measure | Fluent, confident, fabricated content |
Read across the rows and a pattern emerges: classic machine learning gives you a result you can audit statistically, while generative AI gives you a draft you must read and verify line by line. Strong legal platforms increasingly combine the two.
How are machine learning and generative AI used in legal work?
Both technologies are already in daily use, but for different jobs. The rule of thumb is that classic machine learning is best where you must sort or predict across a large volume of documents, and generative AI is best where you must produce language. Many real workflows chain them: a classifier narrows a million documents to the few thousand that matter, and a generative model then summarises or drafts from that smaller set.
Classic machine learning: classify and predict
Classic machine learning shines in technology-assisted review and predictive coding, where a model trained on reviewer decisions prioritises the documents most likely to be relevant in e-discovery. The same family of techniques powers clause classification (tagging a contract's provisions by type), document categorisation, and outcome or litigation-risk prediction. In each case the model is sorting or scoring against learned categories, and its accuracy can be measured with recall and precision - a discipline that makes it defensible in court.
Generative AI: draft, summarise, and answer
Generative AI is the engine behind first-draft contracts and pleadings, plain-language summaries of long documents, and conversational legal research that answers a question in prose. Its strength is producing usable language from messy inputs; its weakness is that the language can be wrong in ways that read perfectly. That is why the better legal tools ground generation in retrieved source documents and cite every claim, so the lawyer can check the output against the original rather than trusting the prose.
Why does the distinction matter for lawyers?
The distinction matters because the two technologies fail differently and so must be verified differently. When a classifier is wrong, it mislabels a document, and you can estimate how often that happens by sampling - the error is measurable and bounded. When a generative model is wrong, it can produce a fluent paragraph citing a case that does not exist, and no amount of sampling will catch it unless you read the output and check the source. Treating both as a single "AI" invites you to apply the wrong kind of scrutiny.
Standards bodies have made this concrete. The US National Institute of Standards and Technology publishes an AI Risk Management Framework that distinguishes risks by system type and use; the NIST guidance reinforces that you cannot manage what you have not correctly categorised. For a lawyer, the takeaway is practical: ask vendors which kind of AI a feature uses, how its errors are measured, and what you must verify - then build that verification into your workflow.
Where does Judicio fit?
Judicio uses both kinds of AI deliberately and keeps a human in the loop for each. On the generative side, Legal Research and Drafting produce answers and first drafts - but every finding, answer, and date is cited to the exact page and quoted passage, and web sources are archived as permanent PDFs, so you verify against the original rather than trusting the prose. The labels and section references attached to citations are deterministic, never AI-generated, which keeps the part you rely on for navigation from being part of the generation.
On the structured side, Document Review and the Review Matrix apply AI to sort, extract, and answer across multiple files in a single run, with up to 25 questions in a matrix, each cell cited to its source. The combination mirrors the argument of this article: use the right kind of AI for the job, ground generation in retrieval, and keep verification fast. Judicio does not train on your data, hosts on Google Cloud Platform, and provides role-based access with an audit trail.
How do you choose the right tool?
Choosing the right tool starts with naming the task honestly: are you sorting and scoring a large set of documents, or producing language? If the job is to find the relevant files in a haystack, you want classic machine learning with measurable recall and precision. If the job is to draft, summarise, or answer, you want generative AI - but only the kind that grounds its output in cited sources you can check. Many matters need both, in sequence.
You can see both approaches in one workspace with Judicio's 7-day free trial - 500 credits, no credit card - and run research, drafting, review, and matrix extraction on your own files. Professional plans are $200 per month for 5,000 credits, or contact us for a walkthrough. Whatever you adopt, let the kind of AI determine the kind of checking, and keep your judgment in the loop.
Judicio's outputs are research and drafting aids, not legal advice; the lawyer retains judgment over every result.