ChatGPT is a large language model (LLM) that takes text prompts from users and produces responses that mimic human writing. It is not a search engine, nor is it a database. In addition to generating text, it is a powerful tool for summarizing, annotating, analyzing, translating, categorizing, and interpreting language. In this guide we focus on ChatGPT, though other generative AI tools like Bing, Bard, Microsoft Co-Pilot, and Claude are other LLMs gaining in popularity.
Background: Artificial intelligence has origins as early as the 1950s and 60s, originally trying to answer the question of how to make something “smart." One approach was to constantly create millions flow charts and code them so that then the computer would be “smart." The other approach was the neural net (connectivist) approach which sought to emulate the brain. Although the neural net approach had legitimacy, it only began to gain steam as computer processing speed improved in 1970s and 80s. ChatGPT is based on the neural network approach.
While AI made great strides in decision-making tasks like self driving cars, gameplay/chess, and modeling, other areas like language proved more difficult to conquer. To emulate language, initially terms were coded in metadata one term at a time, in something like a large spreadsheet. Language models improved on this by assigning words to unique vectors. AI assigns vectors again and again and tries to play the game to get better.
Example: The man went to the store to buy a gallon of [______]. The first time it plays the game it guesses the word [elephant] and self-supervises itself by evaluating the response. Eventually it realizes the correct response [milk] and remembers that vector choice.
In 2017, a team led by Google developed the modern day transformer. A transformer essentially boosts the speed of these models, allowing for a language model that uses millions of dimensions and billions of words/vectors in multiple language. This is known as a large language. ChatGPT stands for Chat Generative Pre-trained Transformer and is an example of a large language model.
It is crucial to note that although ChatGPT may appear to "know" something about a particular law, it does not actually know about its legislative intent, statute of limitations, or application in case law. Its powerful algorithms and ability to mimic human language can give the appearance of knowledge and understanding.
Example: Jack and Jill went up the hill to fetch a pail of [_______]. In this instance, ChatGPT will likely respond with [water] due to the popular English nursery rhyme. Factually though, this makes little sense. Water flows downward. Jack and Jill shouldn't be heading up a hill to fetch a pail of water at all. ChatGPT does not understand gravity or physics but does know that the word "water" often follows that text prompt. ChatGPT and large language models in general are trained to simply guess the next word, without any concern for accuracy.
Release of ChatGPT-3.5: Open AI launched several versions of ChatGPT beginning in 2018 and fine-tuned it by adding elements like Reinforcement Learning form Human Feedback (RLFH). RLFH uses human feedback in the training loop to minimize untruthful, offensive, biased, or heavily nuanced outputs (like our Jack and Jill example above). ChatGPT-3.5 was released in December 2022 to much fanfare. However, at the time OpenAI CEO Sam Altman cautioned that ChatGPT really shouldn't be used "for anything important right now."
For more on what exactly ChatGPT and other large language models are doing:
ChatGPT-3.5 is the free version offered by OpenAI, released in December 2022. ChatGPT-4 is the newest publicly available version of ChatGPT. It costs $20/month. Although they were released within just six months of each other, ChatGPT-4 is significantly more sophisticated. GPT-4 correctly articulates the laws that are relevant and successfully orients the student/lawyer to the correct elements and terms of art. GPT-4 is also better able to handle longer and more sophisticated fact patterns.
How to interact with GPT-4 without a subscription
Strengths
Weaknesses
