Scott Stevenson, Co-Founder & CEO of Spellbook

Scott Stevenson, is Co-Founder & CEO of Spellbook, a tool to automate legal work that is built on OpenAI’s GPT-4 and other large language models (LLMs). It has been trained on a massive dataset of 42 terabytes of text from the Internet as a whole, contracts, books and Wikipedia. Spellbook is further tuning the model using proprietary legal datasets.

What initially attracted you to computer engineering?

I loved video games as a kid, and was inspired to learn how to make them as a teen–that set me on the course of becoming a software engineer. I’m drawn to the profession’s inherent creativity and also appreciate the hardware aspect intertwined in computer engineering.

Can you discuss how your experience with GitHub Copilot was the initial inspiration for Spellbook?

We had been working with lawyers for years, trying to help them automate the drafting of routine contracts using advanced templates. They would often say the same thing: “templates are great, but my work is too bespoke for them.” 

GitHub Copilot was the first generative AI assistant for software engineers–you can start writing code and it will “think ahead” of you, suggesting large chunks of code that you might want to write next. We immediately saw how this could help lawyers draft bespoke agreements, while also helping them intelligently “auto-complete” contracts.

How does Spellbook suggest language for legal contracts?

In the first version of our product, we offered a sophisticated auto-complete feature, similar to Github Copilot. Now we have a number of other mechanisms:

1. Spellbook Reviews can take an instruction like “aggressively negotiate this agreement for my client” and suggest changes across an entire agreement.
2. Spellbook Insights automatically finds risks and suggested clauses across an agreement.

Spellbook also reviews contracts, what type of insight does it offer legal professionals?

Spellbook offers a variety of insights during contract reviews for legal professionals. These insights can be tailored using different “Lenses.” We provide default lenses for tasks like contract negotiations, but lawyers can also provide custom instructions, such as “Review this contract to ensure it complies with California customer requirements.”

Spellbook can uncover potential risks, identify oversights, pinpoint inconsistencies, and receive valuable suggestions for improving and enhancing contracts.

Can you describe how Spellbook overcomes the token size limits that are offered by LLMs?

This is a significant part of what sets us apart and constitutes our unique approach. Managing lengthy contracts that can be more than hundreds of pages can put a strain on an attorney’s bandwidth, but Spellbook’s technology excels in handling them efficiently. While we won’t delve into the specifics of our methods at the moment, this is where our expertise truly shines.

How is the data sourced to train the AI models?

We have availed of public datasets like EDGAR, as well as proprietary contract data sets we built during our company’s first phase at www.rallylegal.com. However, we think that RAG-based approaches are the best way to incorporate accurate legal data into generated text. RAG allows many data sources, such as a client’s own documents, to be referenced.

Laws and regulations change rapidly, how does the AI keep current with the latest news and developments?

We are finding that retrieval-augmented generation (RAG) approaches are extremely effective for this. We think of language models more as a “human reasoning” technology. We generally shouldn’t treat LLMs as “databases”, and instead allow them to retrieve reliable information from trusted sources.

How does Spellbook mitigate or reduce AI hallucinations?

We have relentlessly tuned every feature in Spellbook to provide the best results for lawyers. As mentioned above, RAG also helps keep results relevant and up-to-date. Lastly, our approach to AI is called “Assistive AI”: we always keep the lawyer in the driver’s seat, and they need to review any suggestions before they are acted upon. This is central to everything we do.

At the moment contract drafting and review is the primary use case, what are some additional use cases that Spellbook plans on offering?

We are quite focused on being the best tool for commercial/contracting lawyers right now. One natural extension of that is helping lawyers with legal diligence during a complex transaction. Often law firms will put together a deal room containing every substantial legal document in an organization, reviewing for risks and discrepancies across the corpus. Spellbook is working towards implementing this use case!

What is your vision for the future of AI in the legal profession?

Our “Assistive AI” vision is for every lawyer to have an “electric bicycle” which helps them do their job much faster while producing higher quality work and spending more time on adding strategic value to clients rather than copying and pasting. We think AI should come to lawyers and be a “wind at their back” without requiring much habit change. We think every lawyer will soon have an AI switched “on” during every hour of their work, whether they are in Word, emailing or in a client meeting.

This ultimately means that the 70% of potential legal clients, who cannot afford legal services, will finally be able to be serviced. We’re really excited about that too.

Thank you for the great interview, readers who wish to learn more should visit Spellbook.