Lesson 2.1:What an AI Assistant Actually Is (and Isn't)

“AI” is a broad label, and legal tech has used it for years: predictive coding to sort documents in e-discovery, accepted by US courts since 2012, or clause extraction in contract review tools. Those systems sort, classify and flag. They don't write.

What arrived with ChatGPT at the end of 2022 is different: generative AI. ChatGPT and the assistants that followed, such as Claude and Gemini, produce new text on request. That is what most lawyers mean by “AI” today, and what this guide is about.

If you've used one for real work, you've likely seen two things. The output is often useful: a workable first draft of a clause, a decent summary of a long agreement. And it is sometimes wrong in ways that are hard to spot, because a wrong answer reads exactly like a right one. A summary can skip the one provision that matters; a case citation can look real and not exist.

Both come from the same place. Once you understand what an AI assistant is doing under the hood, you'll know when to rely on it and when to check its work.

It predicts text; it doesn't look things up

The engine behind every AI assistant is a large language model (LLM). An LLM is trained on an enormous amount of text, including books, websites, legislation, commentary and code. From that training it learns the patterns of language: which words, arguments and structures tend to follow which.

When you ask it a question, it doesn't search a database for the answer. It generates a response one piece at a time, each time producing what is most plausible given everything before it.

That is why it writes so well. It has absorbed the style of millions of well-drafted documents, including many contracts, memos and judgments.

It is also why it can be wrong without knowing it. A plausible sentence and a true sentence are not the same thing. Ask for a case supporting your argument and the model will produce something that looks exactly like a case supporting your argument: a realistic name, a court, a year, a pinpoint citation, like Harlow v. Meridian Holdings, 758 F.3d 412, 419 (2d Cir. 2014)Invented. Sometimes the case is real. Sometimes it has been assembled from patterns and never existed. This is called a hallucination, and the model's tone gives you no clue which one you're looking at.

How it learned: reading, then coaching

A model is built in two stages. The first, pre-training, is reading on a scale no person could match: the model works through a vast slice of the written world, guessing the next piece of text billions of times and adjusting itself after every miss. OpenAI, Anthropic and Google don't say how much text that is. DeepSeek, a Chinese AI lab whose models are open-source, free for anyone to download, does: it pre-trained its V4 models on more than 32 trillion tokens DeepSeek · V4 model card, Apr 2026 (opens in a new tab). At eight hours a day, that is over half a million years of reading.

Nothing is filed away. The model doesn't keep copies of what it read; it keeps patterns, the way a lawyer who has read a thousand leases knows what a break clause looks like without recalling any single one. So what it saw often, it knows well; what it saw rarely, it reconstructs. It can explain a landmark provision accurately and still garble the citation of a minor judgment.

A model fresh from pre-training only continues text: ask it a question and it may answer with more questions. The second stage, post-training, coaches it on examples of good answers and on ratings from human reviewers, until it follows instructions, answers helpfully and declines some requests. That coaching is what turns a text predictor into an assistant.

Many assistants now also “think” before answering: they draft their reasoning step by step, then write the reply. It helps on problems with several steps. It is still prediction, one piece at a time.

Its knowledge stops at a date

What it can see: the context window

In a conversation, the model works only from what is in front of it: the app's standing instructions, the documents you attach, the conversation so far and your new message. That space is its context window, measured in tokens.

It has grown fast. In October 2026, leading models from Anthropic, OpenAI and Google each accept about 1 million tokens through their developer API OpenAI · Models (opens in a new tab), roughly 1,500 pages of dense text.

The chatbot on the web often gets less. There, the window depends on your plan. Paid Claude plans and Google AI Pro now offer the full million Anthropic · Context window on paid plans (opens in a new tab); ChatGPT's plans and free tiers offer a fraction of it. Check yours before pasting in a whole bundle.

Two caveats. Fitting is not the same as reading carefully. Models get less reliable as inputs grow, and details in the middle of a long input are the easiest to miss Chroma · Context Rot, July 2025 (opens in a new tab). And when material doesn't fit, the tool has to pick extracts, so the answer is only as good as the extracts it picked. Give it the documents that matter, not everything you have.

One more thing happens behind the scenes. Each time you send a message, the app sends the conversation so far back to the model, along with its own instructions and the tools it allows. That packaging is the harness, and it is why the same model can behave quite differently in two products. Chapter 5 looks at that.

What this means for your work

Put together, an AI assistant is a remarkably well-read writer that works from patterns, sees only what is put in front of it, and stops learning at a fixed date. That makes it strong on language and unreliable on facts it can't see.

The rest of this chapter turns to practice: asking a question versus delegating a task, how skills and connectors plug an assistant into your tools, and what lawyers are actually using it for. Chapter 3 then asks how far to trust the results.