Like many, I came to Chicago for the 2023 ABA TechShow knowing that ChatGPT was going to be a buzzword in the hallways, sessions, and panels. That was no surprise as it has been trending (including in legal tech circles) since its public launch in November 2022.
My goal as a TechShow attendee was two-fold: (1) Gain a better understanding of how ChatGPT works, and (2) attempt to put aside the hype to objectively evaluate its potential impact on how lawyers work.
Here is what I discovered and why I am excited about what ChatGPT can do for legal services.
How does ChatGPT do its thing?
ChatGPT is a large language model (LLM) trained by OpenAI that uses learning techniques to generate human-like responses to natural language inputs, called prompts.
At its core, ChatGPT is a neural network that has been trained on vast amounts of text data to predict the language most likely to follow a given sequence of words.
In the TechShow session “ChatGPT for Lawyers,” which was facilitated by Tom Martin of LawDroid and Carolyn Elefant of MyShingle (a former speaker at the Commission’s Future Is Now: Legal Services conference), Martin helped us understand how ChatGPT creates generative responses rather than automated outputs.
In other words, when forming an answer, ChatGPT adapts to the relationships of the parameters given rather than following a predetermined decision tree.
Martin reviewed three key components of ChatGPT: vectors, tokens, and temperature.
Vectors – ChatGPT uses vectors to represent words. Think of a vector as a mathematical arrow that has both a magnitude and a direction. In the case of ChatGPT, each word in its vocabulary is represented by a vector in a high-dimensional space.
These vectors are learned during training using a technique called word embedding, which maps words to dense numerical vectors that capture the meaning and relationships between words.
Tokens – A token is a sequence of characters that represents a unit of meaning, such as a word or a punctuation mark. For example, the prompt “What is the meaning of life?” has seven tokens, one for each word and one for the question mark.
In ChatGPT, the input text is tokenized into individual tokens that are then fed into the neural network. The model uses these tokens to generate a probability distribution for the next word in the sequence, based on the context provided by the preceding words.
A single prompt is limited to 1,000 tokens (~1.5 pages of content), so you cannot paste an entire appellate brief into ChatGPT and ask for an analysis…yet.
Temperature – ChatGPT uses a technique called temperature sampling to control the level of randomness in its responses. Temperature is a hyperparameter that determines the degree of randomness in the output generated by the model.
Think of a dial between zero and two. Turning the temperature up toward two leads to more diverse and unpredictable responses while turning it down to zero leads to more conservative and predictable responses.
In practice, ChatGPT works by taking in a sequence of tokens and generating a probability distribution over the next possible tokens. The model then samples from this distribution to generate the next token in the sequence, and the process is repeated until a stopping criterion is reached.
When ChatGPT was released on November 30, 2022, it was working with data containing 175 billion parameters and limited knowledge of events after 2021. The next version, ChatGPT 4.0, is expected to be released in 2023. It will have 100 trillion parameters that are expected to enhance its abilities, including creativity.
The output generated by ChatGPT is often remarkably human-like, and the model has proven to be a powerful tool for a wide range of natural language processing tasks, including language translation, question-answering, and chatbot development.
Putting ChatGPT to work for legal services
During TechShow’s LegalTech Visionaries panel, Erin Levine of Hello Divorce (and another former Future Is Now speaker) put it bluntly when she said, “ChatGPT is not going to replace [lawyers], it’s going to make us look like superheroes.”
Although she described writing prompts for ChatGPT “like talking at a 10-year-old,” she told attendees that she is not only using the platform to help draft social media and marketing content but is exploring where it could fit in servicing her clients.
Visionaries panelist Jazz Hampton of TurnSignl described ChatGPT as a “great launching point” for content and admitted he often uses it to draft LinkedIn posts. Once you recognize its capabilities, he said, you may ask it to modify content as well as tone or style. For example, asking ChatGPT to edit a LinkedIn post to be more cheerful.
In the TechShow session “Using AI and Data Analytics in Litigation,” Steve Embry of TechLaw Crossroads queued up a discussion for Pablo Arredondo of Casetext by asking a question that is probably on the minds of many lawyers: “So what?”
In short, Embry answered his own question. He said that ChatGPT has already proved itself as a tool for lawyers seeking to deliver legal services more efficiently (faster) and effectively (lawyers doing lawyer work).
From assisting with legal writing to streamlining client communications, ChatGPT offers a range of benefits that can enhance lawyers’ ability to provide quality legal representation, Embry said.
For example, with its open API, ChatGPT may enhance law firms’ abilities to respond to client inquiries, schedule appointments, and provide updates on case progress.
This can help lawyers free up valuable time that would otherwise be spent on administrative tasks, allowing them to focus more on the legal aspects of their practice.
Legal prompts for ChatGPT
Below are prompts Martin and Elefant shared that lawyers may want to try entering into ChatGPT. As a reminder, a prompt is language supplied by a human that is the starting point for ChatGPT’s response.
Prompts can range from short questions to complex legal situations. The better and more specific a prompt the higher the likelihood that you will receive an accurate response.
Prompts for assisting legal writing
- Draft a demand letter for an auto accident case [with specific details for injuries and dollar amounts]
- Draft interrogatories and requests to produce documents for [specific cause of action and facts from a Complaint]
Prompts for legal marketing
- Draft a polite response to this negative online review: [paste review]
- Write sample YouTube titles with short descriptions for videos [about a topic]
- Write an outline for an eBook about the probate administration process in California, include a short summary for each step of the outline, and use language suitable for an 8th-grade reading level
- Also for this eBook, develop a legal disclaimer and color scheme and propose a cover photo and a title
- Write a 5-day email plan for client marketing [with specific details about the topic]
Prompts for administrative work
- Convert information from this paragraph into an organized spreadsheet: [paste paragraph]
While ChatGPT may allow lawyers to focus more of their time on legal work, I would like to share a notable caveat from my research: when asking ChatGPT legal research questions, be mindful that it can get confused about what is real and what is fiction.
For example, when I asked it to provide caselaw on a topic, it cited and described three cases from three different jurisdictions that were all relevant to my prompt.
However, while they sounded very convincing, the cases were fake. So, I followed up my prompt by asking for “actual cases.” This caused it to first discuss a “real case” and then to say that the real case wasn’t a real case when I asked for the citation.
Confused? You can see the screenshots in this Twitter thread.
Arredondo called these errors “hallucinations” that can nevertheless be very convincing. When asked if the newer versions of ChatGPT would be any better at deciphering fact from fiction, he said, “No, in fact, they will only be more convincing!”
However, when the OpenAI model is integrated with a data set (for example, cases, documents, etc.), it can draw upon reliable data to produce accurate, factual answers that are supported by real citations.
Basically, when the model is applied to data, it can improve accuracy (though caution should still be applied). For example, Docket Alarm uses GPT-3 artificial intelligence to summarize case filings without having to open them.
We still need attorneys ‘at the helm’
While ChatGPT might help lawyers improve their writing skills or act as a virtual legal assistant in some respects, it will not replace the critical thinking, legal expertise, judgment, and empathy that human lawyers bring to their practice. ChatGPT should be viewed as a tool to enhance the legal practice, rather than a replacement for it.
Arredondo put it well when he said, “it still needs the attorney at the helm.”
As the legal industry continues to evolve, ChatGPT and other LLMs will likely be used more in legal practice. Lawyers must stay informed on these technologies, so they can separate reality from hype and meet the changing needs of their clients more efficiently and effectively than ever before.
We will be discussing legal innovation and technology in your practice at our virtual Future Is Now: Legal Services conference. Register today and join us on April 20! CLE is available.