Generative AI for lawyers

The Generative AI landscape

The list below provides a diverse set of currently available generative AI systems, but it is not exhaustive. These are not always end-user products, but more the underlying tools which are used to build end user products. There are numerous other systems and models in each category, each with its unique capabilities and applications.

1. Natural Language Processing (NLP)

GPT: short for Generative Pre-trained Transformers, are a family of neural network models which are trained to generate text (or other content, such as images) in response to a prompt.

BERT: Understands the context and meaning of words in a sentence.

Word2Vec: This is a collection of algorithms which can be used to converts words into numerical vectors for NLP tasks. Vectors capture information about the meaning of the word based on the surrounding words.

Seq2Seq: this model ranslates one sequence of text to another and is used for tasks such as machine translation, text summarization, conversational modeling and image captioning.

2. Computer Vision

GANs: short for Generative Adversarial Networks, these can generate realistic images by training two neural networks, a generator and discriminator network. The discriminator network tries to discriminate between real and fake images, and its output is provided as feedback to the generator network to improve its output.

Pix2Pix: a generative adversarial network (GAN) capable of transforming images from one representation to another, such as converting semantic labels to realistic images, transforming sketches to photorealistic renderings, and enabling image-to-image translation tasks.

3. Music and Audio Generation

WaveGAN: a Generative Adversarial Network to generate waveform audio.

MuseNet: Composes original music in various styles and genres.

Lyrebird: Mimics voices by learning from a few seconds of audio.

DeepComposer: Generates original music compositions using different musical parameters.

4. Video and Animation

Vid2Vid: Transforms segmentation maps into photorealistic videos.

Toonify: Converts real-life videos into cartoon-like animations.

DALL·E: Generates images from textual descriptions.

5. Reinforcement Learning and Game Design

AlphaGo: Plays the game of Go at a world-class level.

OpenAI Five: Competes against human players in complex video games, like StarCraft and Dota.

DeepMind Lab: Provides an a suite of challenging 3D navigation and puzzle-solving tasks for learning agents.

MuZero: The successor project to AlphaGo, this framework learns to play board games, such as chess and shogi, without prior knowledge.

6. Generative Design and Creativity

DeepDream: Enhances images by amplifying patterns and features.

DeepArt: Converts photos into artistic styles and paintings.

StyleGAN: Generates novel and realistic images with control over visual attributes.

DeepMind Lab: Provides an environment for developing and testing reinforcement learning algorithms.

Generative AI tools for lawyers

ChatGPT

ChatGPT is the most well known large language model neural network (LLM). The free version is ChatGPT 3.5, and ChatGPT 4 costs US$20 per month for a subscription. The difference between the two is significant. For example, it has been reported that ChatGPT 3.5 passed some sections of the Uniform Bar Exam in the US, but came in the lowest 10th percentile overall. ChatGPT 4 however not only passed the exam, but came in the top 10%.

Claude

Claude is an AI assistant developed by Anthropic, PBC. Created using a technique called Constitutional AI, Claude claims to have been trained to be helpful, harmless, and honest using natural language feedback. Claude has a personality and can conduct complex conversations, but it was designed by Anthropic researchers and engineers to be an ethical and trustworthy assistant that provides truthful responses without deceiving people or causing unintended harm. Claude aims to demonstrate the potential benefits of aligned AI if it is developed responsibly and for the good of humanity.

Gemini

Gemini is Google’s offering in this space. Gemini Nano, the entry level model, has either 1.8 or 3.75 billion parameters. Google claims to abide by certain AI Principles in the development of this model. It is also trained on a variety of unpublished data sources. It can accept up to 20Mb of data through the web interface, or larger amounts through its text API.

Prompt engineering

Prompt engineering is a new term which describes how to write the most efficient prompts for use in generative AI systems, like ChatGPT. Here are a few resources which might help you get started.

Prompt databases

These are examples of prompts which other people have created, that might help you explore (a) how to write better prompts, and (b) what sort of things you might be able to get ChatGPT and the other to do for you.

Prompts.chat

https://prompts.chat

These are a bunch of general purpose (i.e. not legal) prompts which users have submitted. They are good examples of how to write clear and concise prompts for various tasks.

Learn Prompt

https://www.learnprompt.org/chat-gpt-prompts-for-lawyers/

A range of basic prompts for various legal tasks that provide a starting point for using ChatGPT for a range of legal tasks. They are only a starting point though - don’t forget that ChatGPT is iterative!