Chapter 3: Generative AI Models and Tools

Chapter Overview

In Chapter 2, we dove into the inner workings of generative AI to see how neural networks, large language models, and the transformer architecture collaborate to produce human-like text. We discovered how attention mechanisms enable these models to handle context, and why data quality, along with thoughtful human oversight, matters so much. We also explored where generative AI excels, such as piecing together known patterns, and where it falls short, like truly novel scenarios. Ultimately, we learned that understanding the “nuts and bolts” of AI helps us gauge when and how to rely on these tools responsibly in legal practice.

In this chapter, we will focus more concretely on how to apply generative AI in the real world and what’s out there in the world of generative AI tools. You will:

By the end of this chapter, you should be able to confidently explain what each of these generative AI tools does, decide when a particular solution is best suited for the task at hand, and identify the practical considerations, like cost, security, and ease of integration, that should influence your choice.

Proprietary vs. Open-Source Large Language Models (LLMs)

Imagine you’re choosing between using a commercial legal research platform that has all the bells and whistles or building a custom research tool from scratch. On one hand, the commercial platform might be easier to adopt and offer ready-made features. On the other hand, building your own system could give you greater flexibility and customization. This same dynamic plays out in the world of large language models.

Proprietary LLMs

Proprietary LLMs (like ChatGPT or Anthropic Claude) are owned and maintained by a single organization. These companies manage every aspect of the model, from how it’s trained to how users connect with it, often with user-friendly interfaces that many people can use without deep technical knowledge.

However, proprietary models do have some limitations:

Open-Source LLMs

Open-source LLMs (like certain versions of LLaMA or other community-supported projects) make their code or model weights publicly available. This transparency allows anyone with sufficient resources and expertise to adapt the model for their needs.

Open-source models do have disadvantages:

Callout: Key Term – LLM (Large Language Model)
A Large Language Model is a type of AI system trained on vast amounts of text. It learns statistical patterns of language and can generate coherent, contextually relevant responses to prompts. Think of it as a supercharged version of your phone’s autocomplete, but capable of much more nuanced understanding and creation of text.

Predictive AI Models vs. Reasoning Models

As AI continues to evolve, you may hear terms like “predictive model” or “reasoning model.” These terms help describe how an AI system processes information and the kinds of tasks it performs well. Let’s break down what each type of model does and introduce three specific examples you may encounter: GPT-4o, o1, and o3.

GPT-4o: A Predictive AI Model

GPT-4o is an example of a predictive AI model. Introduced in May 2024, it can process and generate text, images, audio, and video. That’s what we mean when we say it’s “multimodal.” Because it works with so many media types, GPT-4o can:

Many lawyers appreciate GPT-4o for its speed. It offers faster response times and reduced operational costs compared to earlier versions. This speed is especially useful for tasks such as real-time content updates or code editing. If you need quick turnarounds, say, drafting a motion in court under time pressure, GPT-4o might be the right choice.

o1: A Reasoning AI Model

Released in September 2024, o1 shifts focus from pure prediction to reasoning. Think of it as an AI designed to “think aloud” (internally) before providing an answer. This feature, often described as a “chain of thought,” helps it tackle complex tasks more accurately than a purely predictive model.

Because o1 excels at analytical thinking, you might use it for:

o3: Advanced Reasoning Model

o3, introduced in December 2024, is the successor to o1. It refines the reasoning approach even further, using an internal “private chain of thought” that allows it to map out a plan and reason through tasks before arriving at an answer.

In testing, o3 demonstrated impressive performance in challenging settings:

The key takeaway? o3 is built for sophisticated problem-solving. If your legal work demands advanced reasoning, such as analyzing complex procedural rules or dissecting intricate regulatory frameworks, o3 might be the ideal tool.

Example Scenario

Context: You’re working on a patent infringement case involving cutting-edge biotech.

Predictive Approach: GPT-4o might quickly summarize the patents and relevant court rulings, or generate a draft argument.

Reasoning Approach: o1 or o3 could delve deeper into the technical details, offering a line-by-line analysis that highlights nuances in the claim language or prior art references.

In practice, you might use both. GPT-4o for speed and summarization, and o3 for deep analytical insights.

Key Differences

In sum, if you need quick content generation, across text, audio, and video, GPT-4o is often your best bet. If your task demands careful, methodical thinking (e.g., analyzing complicated legal issues or performing detailed code debugging), then a reasoning model like o1 or o3 is a stronger choice.

Practice Pointer

When you’re deciding between a predictive or reasoning model, consider the complexity of the legal question. Does it demand in-depth analysis (reasoning) or a rapid overview (prediction)? Think about the time constraints you have, the degree of risk in getting an answer wrong, and whether the project involves tasks that benefit from the AI “thinking” through multiple steps.

Copilots vs. Agents

Generative AI tools generally fall into two categories: copilots and agents. While these terms aren’t always used consistently in every product description, they represent helpful concepts for understanding how much autonomy the AI has in carrying out tasks.

What Is a Copilot?

Copilots function as assistants or extensions of a human operator. They enhance your workflow but still require human oversight. For example, a copilot might:

A copilot is like a junior associate who comes prepared with a rough draft or background research for your review. You remain in control and bear responsibility for the final work product. This oversight is crucial, especially in the legal profession, where your name (and potentially your license) is on the line.

What Is an Agent?

Agents are designed to work semi-autonomously or sometimes even fully autonomously. Agents combine an LLM with a goal and instructions, knowledge, and access to tools. They can perform tasks end-to-end with minimal human intervention. In legal settings, an agent might:

While agents can save significant amounts of time, they also raise ethical considerations. If your agent is drafting a legal document for filing with a court, do you trust it to get everything right? If it makes a critical error, are you at risk of malpractice? For these reasons, many law firms still require a final human review, even if an agent does 90% of the legwork.

Agent - GenAI Book Illustrations.png

Callout: Key Term – Autonomous

When we say an AI tool is autonomous, we mean it can act on its own without constant human input. In law, this autonomy can be helpful for routine tasks but also risky if the AI makes flawed decisions. Always assess the appropriate level of human supervision.

Practical Considerations Before Choosing an AI Tool

Picking the right AI tool for your legal work isn’t just about which has the best performance metrics or the coolest features. You also need to account for the real-world practicalities of practicing law.

1. Usability

2. Integration

3. Cost-Effectiveness

4. Security and Privacy

Practice Pointer

Before integrating a new AI tool, draft a concise checklist of must-have features and constraints. Include factors like data security, ease of learning, and cost. Having a clear list ensures you choose a tool that aligns with your firm’s unique needs.

With the fundamentals in mind, predictive vs. reasoning, copilots vs. agents, proprietary vs. open-source, let’s now explore some specific generative AI tools making waves in the legal industry. Each tool has its strengths, weaknesses, and ideal use cases.

Benefits of AI Tools for Lawyers

Regardless of the tool you choose, here are some broad benefits you can expect from generative AI:

ChatGPT 4o (and o1)

ChatGPT 4o, developed by OpenAI, is a multimodal AI model capable of handling text, audio, and visual inputs. Its speed is a standout feature: it can respond in near-real time, facilitating natural back-and-forth conversations.

Key Features

  1. Multimodal Input and Output

    • Process and respond to text, audio, and images.
    • Potential to transform client interaction by allowing multiple input types (e.g., voice commands).

  2. Fast Response Time

    • Responds to audio inputs in as little as a few hundred milliseconds, close to human conversational speed.

  3. Vision Mode

    • Interprets visual content, which can be applied to analyzing images, diagrams, or even videos.

  4. Reasoning Capability with o1

    • o1 can “think before speaking,” adding a layer of deeper analysis for complex tasks.

  5. Advanced Data Analysis

    • Handles spreadsheets, helping lawyers track case details or financial records more efficiently.

  6. Text Analysis

    • Can review legal documents, check grammar, extract key information, and even translate content.

  7. File Compatibility

    • Supports multiple file types (PDF, .docx, .txt, .rtf, .xlsx) commonly used in legal practice.

Additional OpenAI Features

Pricing

Technical Aspects

Pros and Cons of Implementation

Pros Cons
Increased Efficiency: Automates tasks like document review. Potential for Bias: AI models can mirror biases in training.
Improved Accuracy: Reduces errors in contracts and memos. Dependence on Technology: Over-reliance can impede skill growth.
Enhanced Client Communication: Summaries in plain language. Data Security: Sharing sensitive info may raise confidentiality issues.
Accessibility: Text-to-speech can help users with disabilities. Limited Creativity: AI lacks the strategic nuance of human insight.
Cost-Effectiveness: Long-term savings by freeing attorney hours. Need for Human Oversight: AI-generated work still requires review.

Anthropic Sonnet 3.5 (Claude)

Anthropic Sonnet 3.5, often referred to simply as Claude, aims to be helpful, harmless, and honest. As an ethical AI system, Claude incorporates an approach called Constitutional AI to align the model with human values.

Key Features

  1. Constitutional AI

    • A technique to reduce harmful or biased outputs by aligning the model’s decisions with a set of agreed-upon principles.

  2. Strong Reasoning and Comprehension

    • Useful for analyzing legal documents, writing research memos, and summarizing nuanced information.

  3. Natural Language Processing

    • Helps produce clearer, more concise text, benefiting legal drafting and client communications.

  4. Context Window (200,000 tokens)

    • Can handle lengthy documents without losing track, making it suitable for e-discovery or large contract reviews.

Unique Anthropic Features

Pricing

Technical Aspects

Pros and Cons of Implementation

Pros Cons
Ethical AI: Reduces biased or harmful outputs. Limited Transparency: Exact model details are undisclosed.
Strong Reasoning and Comprehension Cost: Still requires paid plans for extended features.
Natural Language Processing Potential for Over-Reliance: Always verify AI outputs.
Large Context Window

Google Gemini

Google Gemini is a multimodal AI model developed by Google AI. Like GPT-4o, it can process and generate text, images, audio, and even code.

Key Features

  1. Multimodal Capabilities

    • Can handle various information types, useful for analyzing scanned PDFs, images from case evidence, or audio interviews.

  2. Long Context Window

    • Gemini 1.5 Pro can handle up to 2 million tokens in one go, making it ideal for large-scale e-discovery.

  3. Advanced Reasoning

    • Built to excel at complex analysis and problem-solving tasks.

  4. Integration with Google Services

    • Syncs with Google Search, Google Docs, and other G-Suite tools, streamlining collaboration and research.

Pricing

Technical Aspects

Pros and Cons of Implementation

Pros Cons
Multimodal Capabilities: Handles text, images, audio. Cost: Premium plans may be prohibitive for some smaller firms.
Long Context Window: Suitable for massive document sets. Integration Challenges: May need technical support to integrate fully.
Advanced Reasoning: Good for complex legal questions. Dependence on Google: Not everyone prefers or trusts deep Google integration.

Google NotebookLM

Google NotebookLM is an AI-powered research and note-taking tool. By uploading PDFs, Google Docs, websites, or Google Slides, NotebookLM helps you synthesize and summarize complex information. It provides inline citations so you can verify exactly where the answers are coming from.

For law students, NotebookLM can be a handy tool to break down long cases or research papers into more digestible chunks.

Google Advanced with Deep Research

Google’s Gemini Advanced with Deep Research is designed to gather and synthesize information from the web autonomously. You input a research question, and it:

  1. Plans its approach to researching that question.
  2. Conducts iterative searches, refining results as it goes.
  3. Compiles a detailed report with source citations.

This is helpful when performing preliminary research on emerging legal issues. It can produce a well-organized overview, which you can then refine for briefs or internal memos.

Google Learn About

Google Learn About is an experimental AI-driven platform that provides deep, interactive learning on a wide range of topics. Built on the LearnLM model, it tailors responses to different learning levels and includes educational aids such as images, videos, and quizzes.

While not exclusively focused on law, it’s easy to imagine “Learn About Torts” or “Learn About Contract Drafting” modules that help new attorneys or clients grasp fundamental legal concepts in an engaging, interactive way.

Perplexity

Perplexity functions more like an AI-powered search engine with conversational abilities. It’s useful for research-driven tasks where you want quick answers plus source citations.

Key Features

Because of its emphasis on credible sources, Perplexity can help lawyers or law students gather references. However, it doesn’t specialize in drafting or analyzing legal text to the same extent as GPT-4o or Claude.

Benefits for Lawyers

Pricing

Generative AI Tools: A Comparative Analysis

Below is a side-by-side comparison of some of the tools we’ve discussed. While these details might change as the tools evolve, this table offers a snapshot of their unique strengths.

Feature ChatGPT 4o o1 Anthropic Claude 3.5 Google Gemini Perplexity
Input Modalities Text, audio, visual Text, visual Text Text, images, audio, code Text
Context Window 128,000 tokens 128,000 tokens 200,000 tokens Up to 2 million tokens Up to 32,000 (Pro)
Number of Parameters 1.76 trillion Not publicly disclosed Not publicly disclosed Varies (Nano to Pro models) Not publicly disclosed
Key Strengths Multimodal, fast Deep reasoning, complex problem-solving Ethical AI, strong reasoning Multimodal, large context Efficient search, citations
Pricing $20/mo (Plus) / $200/mo (Pro) Included with ChatGPT subscription $20/mo ~$19.99/mo + varied plans Free or $20/mo (Pro)

Example Scenario

Situation: A mid-sized law firm wants to adopt an AI tool to assist with e-discovery. They handle large litigation cases requiring review of millions of documents.

Ultimately, the firm might combine the strengths of multiple models, using Gemini or Claude for the heavy-lifting e-discovery and ChatGPT for drafting short client summaries.

Opportunities for Law Students

For law students, mastering at least a working knowledge of these tools is not optional, it’s increasingly expected. Clients today want efficient, cost-effective, and accurate legal services. When you can say, “I know how to use generative AI ethically and effectively,” you stand out in job interviews and on the job.

“Legal tech literacy” is becoming as fundamental as knowing how to format a brief or file a motion. It’s about understanding what tools exist, how they can help, and knowing enough to ask the right questions, even if you’re not a programmer.

Chapter Recap

We covered a lot in this chapter:

  1. Proprietary vs. Open-Source LLMs

    • Proprietary models offer convenience and user support, but come with closed systems and potential vendor lock-in.
    • Open-source models offer flexibility and transparency, but can be more challenging to maintain.

  2. Predictive vs. Reasoning Models

    • Predictive models (like GPT-4o) generate quick, versatile responses across multiple media types.
    • Reasoning models (like o1 and o3) engage in a deeper “chain-of-thought,” excelling in complex problem-solving.

  3. Copilots vs. Agents

    • Copilots assist humans with tasks but require oversight, ideal for drafting, summarizing, and editing.
    • Agents can handle tasks autonomously, which can be powerful but also raises ethical and accountability concerns.

  4. Practical Considerations

    • Usability, integration, cost, security, and privacy are crucial when selecting an AI tool.

  5. Popular Generative AI Tools

    • ChatGPT 4o (and o1): Multimodal, fast, with specialized reasoning capabilities in o1.
    • Anthropic Sonnet 3.5 (Claude): Ethical focus, large context window, strong reasoning.
    • Google Gemini: Multimodal with an extremely large context window, integrated into the Google ecosystem.
    • NotebookLM, Advanced with Deep Research, Learn About: Google’s suite of AI research and note-taking tools.
    • Perplexity: Search-centric model with conversational and citation features.

You’re now equipped to weigh the pros and cons of each major AI player in the legal space and have a framework for deciding whether you need a predictive or reasoning model, a copilot or an agent, and a proprietary or open-source solution.

Final Thoughts

Chapter 3 explored how generative AI models, both predictive and reasoning, are reshaping legal workflows. We looked at proprietary tools that offer ready-to-use solutions and open-source models that provide deeper customization. We also examined the difference between AI copilots, which enhance your work under human oversight, and more autonomous AI agents, which can tackle tasks with minimal input but present greater ethical and accountability concerns.

As you move forward, consider the following:

Generative AI offers the promise of faster drafting, more comprehensive research, and innovative client communications. Yet, caution is warranted: reliance on these systems without proper checks can jeopardize client trust and professional standards. By maintaining a balanced perspective, embracing AI’s efficiencies while upholding legal ethics, you can harness these tools to elevate your practice in a responsible way.

What’s Next?

In Chapter 4, we’ll look at legal-specific AI tools such as CoCounsel, Spellbook, AI.law, and Alexi. These tools directly integrate into law office tasks, like drafting briefs, analyzing contracts for risk, or synthesizing case law, often providing ready-made solutions fine-tuned for the legal profession. We’ll dive deeper into their specific features, how they integrate with existing workflows, and relevant ethical considerations. Stay tuned to learn how these specialized tools can streamline your day-to-day legal practice and transform how you approach client service.