How AI Works (The Non-Technical Version)

The Partner Who Thought AI Was "Just Autocomplete"

Margaret Chen had practiced corporate law for twenty-three years. She had negotiated mergers across four continents, structured deals that reshaped entire industries, and mentored dozens of associates who went on to become partners themselves. When her firm's innovation committee began pushing for AI adoption in late 2024, she was skeptical -- politely, professionally skeptical, the way a seasoned litigator is skeptical of an opposing expert's credentials.

"It's autocomplete," she told her managing partner over coffee. "Fancy autocomplete. I've seen the demos. It predicts the next word. My paralegal does better work."

Then a client dropped a four-hundred-page multi-jurisdictional contract on her desk -- a cross-border acquisition involving entities in Delaware, the Cayman Islands, Luxembourg, and Singapore. The client wanted a preliminary risk analysis by Friday. It was Wednesday afternoon. Margaret's two best associates were buried in due diligence on another deal.

Out of sheer desperation, she opened the AI tool her firm had recently licensed. She uploaded the contract. She typed a prompt she later described as "embarrassingly simple": Identify the key risk provisions in this agreement, flag any inconsistencies between jurisdictional requirements, and highlight unusual indemnification language.

Fourteen minutes later, she was staring at a structured analysis that identified thirty-seven risk provisions across the four jurisdictions, flagged two genuine inconsistencies between the Luxembourg and Singapore governing law clauses, and highlighted an indemnification cap that was unusually favorable to the buyer relative to deals of comparable size.

Margaret spent the next two hours verifying every finding. The AI had gotten thirty-five of thirty-seven provisions exactly right. One flagged provision was actually standard for Cayman structures, and another was a formatting artifact rather than a substantive risk. The two jurisdictional inconsistencies were real -- and her team had missed both of them on a similar deal six months earlier.

"It's not autocomplete," she told the innovation committee the following Monday. "And it's not a junior associate either. It's something we don't have a category for yet. And that's exactly why we need to understand what it actually is."

Margaret's experience captures the journey that thousands of legal professionals are now undertaking. The AI tools available to lawyers today are genuinely powerful -- and genuinely limited in ways that matter enormously in legal practice. Understanding what AI actually is, how it works at a conceptual level, and where its boundaries lie is not an academic exercise. It is a matter of professional competence.

---

What AI Actually Is: Pattern Recognition at Scale

To understand AI as a legal professional, you do not need to study machine learning papers or learn to code. You need one core insight, and the rest follows from it.

The Precedent Analogy

Think about how you learned to practice law. In law school, you read thousands of cases. You began to recognize patterns -- how courts reason about negligence, how contractual interpretation varies by jurisdiction, how regulatory frameworks interact. Nobody gave you an explicit rule for every situation. Instead, you developed judgment by absorbing vast quantities of legal reasoning and learning to recognize the structures within it.

Large language models -- the technology behind tools like ChatGPT, Claude, Google Gemini, and the legal-specific platforms you will encounter throughout this course -- learn in a conceptually similar way. They are trained on enormous datasets of text: books, articles, legal opinions, contracts, regulations, correspondence, academic papers, and much more. Through this training, they develop statistical representations of how language works -- how words relate to each other, how arguments are constructed, how different types of documents are structured.

When you ask an AI a question or give it a task, it generates a response by drawing on these learned patterns. It is, at its core, an extraordinarily sophisticated pattern-recognition system.

Where the Analogy Holds

This analogy is useful because it helps explain what AI does well in legal contexts:

Document Analysis. Just as an experienced attorney can scan a contract and quickly identify unusual provisions because they have seen hundreds of similar contracts, an AI can analyze documents against patterns it has learned from millions of examples. It can identify non-standard indemnification language, flag missing boilerplate, and compare clause structures across agreements -- often faster than any human reader.

Legal Research. When you research a legal question, you draw on your knowledge of how courts have approached similar issues. AI can perform analogous pattern matching across vast quantities of legal text, identifying relevant precedents, statutory provisions, and regulatory frameworks. It excels at finding connections between disparate areas of law that a human researcher might miss simply because no single person can hold the entirety of legal knowledge in their head.

Drafting and Summarization. Years of reading well-drafted legal documents teach you what good legal writing looks like. AI has absorbed the patterns of effective legal drafting from millions of examples. It can generate first drafts of memos, letters, contract provisions, and briefs that follow the structural and stylistic conventions of legal writing with remarkable consistency.

Issue Spotting. Perhaps most valuably for practicing lawyers, AI can perform a kind of systematic issue spotting -- reviewing a set of facts or a document against learned patterns to identify potential legal issues, risks, or opportunities. It does this not through legal reasoning as we understand it, but through pattern matching against its training data.

Where the Analogy Breaks Down

And here is where the analogy fails -- and where the failure matters enormously for legal professionals.

When you reason about a legal question, you are not merely matching patterns. You are applying judgment. You understand that a case from the Second Circuit carries different weight than a case from the Ninth Circuit when you are advising a client in New York. You know that a contractual provision that is standard in one industry may be unusual in another. You can assess whether a novel legal argument is likely to persuade a particular judge based on that judge's prior decisions and judicial philosophy.

AI does none of this. It recognizes patterns in text and generates responses that are statistically likely to be appropriate based on those patterns. It does not understand legal concepts in the way that you do. It does not know what a contract is in any meaningful sense -- it knows how contracts are typically structured and what language typically appears in them.

This distinction is not philosophical hairsplitting. It has concrete, practical consequences that every lawyer using AI must internalize.

---

The Hallucination Problem: Why Verification Is Non-Negotiable

The single most important concept for any lawyer using AI is hallucination. This term describes the phenomenon where an AI generates information that is plausible, well-formatted, and confidently stated -- but factually wrong.

Why AI Hallucinates

Hallucination is not a bug that will be fixed in the next software update. It is a structural feature of how large language models work. Because these systems generate text by predicting what words are statistically likely to follow other words, they can -- and regularly do -- generate text that follows the pattern of accurate information without containing accurate information.

In practical terms, this means:

Fabricated case citations. AI can generate case names, citation formats, and even detailed holdings that follow the exact pattern of real legal citations but refer to cases that do not exist. The citation will look perfect -- correct reporter abbreviation, plausible volume and page numbers, a case name that sounds like a real case. But when you look it up in Westlaw or Lexis, there is nothing there.

Invented statutory provisions. AI can cite statute sections with plausible numbering and language that sounds like real statutory text but is actually fabricated. It may cite a real statute but attribute to it a provision that does not exist or describe a provision inaccurately.

Distorted holdings. Even when AI cites a real case, it may describe the holding inaccurately. It might correctly identify Smith v. Jones as a real case but misstate what the court actually held, combine elements from different cases, or attribute to the majority opinion reasoning that actually appeared in the dissent.

Confident fabrication. Unlike a human who might hedge when uncertain -- "I believe the rule is..." or "If I recall correctly..." -- AI typically presents fabricated information with the same confident tone it uses for accurate information. There is no reliable indicator in the AI's output that distinguishes verified facts from hallucinated ones.

The Mata v. Avianca Wake-Up Call

The legal profession learned about hallucination the hard way. In Mata v. Avianca, Inc., a case filed in the Southern District of New York, an attorney used ChatGPT to prepare a legal brief and submitted it to the court containing citations to six cases that did not exist. The AI had generated complete fabricated citations -- with case names, reporters, page numbers, and detailed factual and legal summaries -- for cases that had never been decided by any court.

When opposing counsel could not locate the cited cases and brought this to the court's attention, the attorney asked ChatGPT to verify the citations. The AI confirmed they were real (they were not). The court ultimately imposed sanctions and referred the attorneys for disciplinary proceedings.

This case was not an anomaly. It was the inevitable result of using AI-generated legal content without verification. And it will not be the last such case.

---

What AI Can and Cannot Do in Legal Practice

With the foundational concepts in place, let us be specific about what AI can and cannot do for you as a legal professional. This is not a theoretical exercise -- it is a practical framework you will use every day.

What AI Excels At

First-draft generation. AI can produce solid first drafts of many types of legal documents: research memos, client letters, contract provisions, discovery responses, demand letters, and more. These drafts will generally follow correct structure and conventions, but they require substantive review and refinement by a qualified attorney.

Document review and analysis. AI can review large volumes of documents far faster than any human team, identifying relevant provisions, flagging anomalies, and summarizing content. This is particularly valuable in due diligence, contract review, and discovery contexts.

Legal research acceleration. AI can help identify relevant legal authorities, suggest research directions, and synthesize information across multiple sources. Used properly -- meaning with rigorous verification -- it can dramatically accelerate the research process.

Plain-language translation. AI excels at translating complex legal language into plain English for clients, and vice versa -- taking a client's description of their situation and identifying the legal concepts at play.

Brainstorming and issue spotting. AI can serve as a systematic thinking partner, helping you identify issues you might not have considered, suggesting alternative arguments, and stress-testing your analysis.

Routine correspondence. Status updates, scheduling communications, standard acknowledgment letters, and other routine correspondence can be drafted quickly with AI assistance.

What AI Cannot Do

Exercise professional judgment. AI cannot assess whether a particular legal strategy is appropriate for a specific client's situation, considering their risk tolerance, business objectives, relationships, and the full context of their circumstances.

Guarantee accuracy. No current AI system can guarantee that its output is factually correct. Every piece of AI-generated content requires human verification.

Maintain privilege. AI systems are not covered by attorney-client privilege. Information shared with AI tools may not be protected, depending on the tool's data handling practices and the jurisdiction's privilege rules.

Understand context. AI processes text. It does not understand the human dynamics of a negotiation, the political considerations within a corporate board, the emotional state of a client going through a divorce, or the unspoken norms of practice before a particular judge.

Apply ethical judgment. AI cannot assess conflicts of interest, determine when disclosure obligations are triggered, or make the nuanced ethical judgments that are central to legal practice.

Stay current. AI models are trained on data up to a certain point in time. They may not reflect recent statutory amendments, new case law, regulatory changes, or evolving professional standards.

---

Confidentiality from Day One: Protecting Client Data

Before you type a single word into any AI tool, you must understand how that tool handles data. This is not an abstract concern -- it is a concrete professional obligation under Rule 1.6 of the ABA Model Rules of Professional Conduct and its state equivalents.

The Data Handling Spectrum

AI tools handle your data in fundamentally different ways, and understanding these differences is critical:

Consumer AI tools (free versions of ChatGPT, free Gemini, etc.) typically use your input to train and improve their models. This means that client information you enter may be incorporated into the AI's training data and could theoretically influence responses given to other users. For legal work involving any client information, these tools are generally inappropriate.

Professional/Enterprise AI tools (ChatGPT Enterprise or Team, Claude for Business, Microsoft Copilot for Microsoft 365 with enterprise data protection) typically commit to not using your data for model training, provide data encryption, and may offer additional security certifications. These are more appropriate for legal work but still require careful evaluation.

Legal-specific AI platforms (Harvey, CoCounsel, Casetext, and similar tools discussed in the next lesson) are designed specifically for legal professionals and typically offer the strongest data protection guarantees, including compliance with legal industry standards and, in some cases, SOC 2 Type II certification, data residency controls, and explicit contractual provisions regarding confidentiality.

Practical Data Protection Steps

Regardless of which tools you use, adopt these practices from day one:

1. Read the terms of service -- specifically the sections on data retention, data use for training, and data sharing. If you would not send client information to a third party without these protections, you should not send it to an AI tool without them either.

2. Anonymize when possible -- when using AI for tasks that do not require specific client details, strip out names, dates, identifying information, and case-specific details. You can get excellent results from AI by describing a legal scenario in generic terms.

3. Use enterprise-grade tools for client work -- if your firm has licensed an enterprise AI tool or a legal-specific platform, use it for client-related work rather than consumer alternatives.

4. Understand your firm's policy -- most firms are developing or have developed AI usage policies. Know what tools are approved, what types of information can be shared with AI, and what documentation is required.

5. When in doubt, anonymize -- if you are unsure whether sharing specific information with an AI tool is appropriate, anonymize it. You can always add specifics back into the AI's output after the fact.

---

The Current Legal Technology Landscape

AI in legal practice is not a single tool or a single capability. It is an ecosystem that is evolving rapidly. As of early 2026, the landscape includes:

General-purpose AI assistants like ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google) that can perform a wide range of legal tasks but are not specifically designed for legal work and require careful handling of confidentiality.

Legal-specific AI platforms like Harvey (built on large language models but trained specifically for legal work), CoCounsel by Thomson Reuters (integrated with Westlaw), and Casetext (now part of Thomson Reuters) that offer AI capabilities designed specifically for legal professionals with stronger data protection.

AI-enhanced legal research tools like Westlaw Precision with CoCounsel, Lexis+ AI, and vLex Vincent AI that integrate AI capabilities into existing legal research platforms.

Document review and contract analysis platforms like Kira Systems, Luminance, Ironclad, and Evisort that use AI to review, analyze, and manage legal documents at scale.

Practice management AI integrated into platforms like Clio, PracticePanther, and MyCase that automate administrative tasks, client communications, and workflow management.

We will explore each of these categories in detail in the next lesson. For now, the key point is that the legal AI landscape is diverse, and different tools are appropriate for different tasks and different practice contexts.

---

Applying What You Have Learned

---

Reflection: Where Does AI Fit in Your Practice?

Before moving to the next lesson, take a moment to consider these questions:

1. What tasks consume the most time in your practice? Research? Drafting? Document review? Client communication? Identify the two or three tasks where AI might have the greatest impact.

2. What is your current comfort level with technology? Be honest with yourself. If you are starting from a low base, that is fine -- this course is designed for you. If you are already using AI tools, think about whether you are using them as effectively and safely as you could be.

3. What concerns you most about AI in legal practice? Accuracy? Confidentiality? Job displacement? Ethics? Name your specific concerns so you can watch for how this course addresses them.

4. What would "success" look like for you? At the end of this course, what do you want to be able to do that you cannot do now?

These are not abstract questions. Your answers will help you focus your attention on the lessons that matter most to your practice and your professional development.