We welcome you to the weird world of GenLaw with our 1600+ legal GPT store. Now, ponder: Can we build Artificial Legal Intelligence (ALI)?
Subtext: Philosophically, we should build Artificial Legal Intelligence. Paradoxically, the question is: Will we ever be allowed to use it? Previously in Agents, we discussed the theoretical possibilities. Shifting focus, in this segment we’ll present the practical applications.
Three years ago, we made a video that foreshadowed the struggle. The video below shows how much money is invested to influence the law compared to how little to improve it. In 2020, we already pointed at OpenAI, among others.
Academic Angle
In the previous post, we stated that legal data has the complete DNA to create autonomous legal services. So by adding legal data and aligning Large Language Models (LLM), we should already have Artificial Legal Intelligence. However, we have yet to see ALI. Notably, here are some academic observations as to why we haven’t:
- Domain-specific large language models tend to underperform;
- LLMs perform worse on legal queries;
- LLMs perform worse when continually fine-tuning (catastrophic forgetting – CF);
- Excessive prompting causes ‘Butterfly Effect‘ and degrades the outputs.
I shared my views in the first and second study on LinkedIn and The Legal Tech StartUp Focus. Most importantly, my friend John Barker made another observation on GPT and Claude: they have been getting worse over time on specific legal queries. In this interview, Sam Altman acknowledges that out of the 10,000 answers an LLM will offer, maybe one of the answers will be perfect. GenAI relies on embedding vector databases, which results in a different output every time. Having 10,000 slight variations on the same topic will be a big deal in some areas like Health.
In Legal, we need consistency on every output. So using GenAI to dispense accurate legal answers every time will be challenging. Baffling to me: why can GenAI generate programming code consistently? Undoubtedly, progress will address these concerns.
Apps Aligned
Considering the above, have you tested any of the Copilots on Legalpioneer? I would love to hear your thoughts. Here are some, that may never get funded but matter to our everyday lives:
Contrary to the above, here are some topics that seem obvious candidates for funding:
- Security
- Human Resources (HR)
- Anti-Money Laundering (AML)
- Know Your Customer (KYC)
Of course, there is a large set of copilots looking at contracts in spaces like:
- Mergers
- Music
- Real Estate
We’ve seen legal copilots with multi-modal support and graphic outputs to help you visualize laws or patents. Some might ask: where are the apps for lawyers? We have research and drafting apps, but again, do we dare to use them?
APEX Apps
This brings us to the future. At the end of 2022, we told Law360 about the increase in consumer-facing legal. Now we’re seeing these client-centric apps proliferate. We will also see the emergence of apex apps. These apps aim to sit on top of the food chain. Perplexity AI and ARC Search are two examples.
Let’s recap: There is a blueprint for Artificial Legal Intelligence and a roadmap. We’ve posted the ‘road signs’ on LinkedIn. Here are five evolutionary stages we’ve seen, with the date we posted them:
- GenAI is an operating system (OS) – January 11, 2023
- The types of legal copilots – November 13, 2023
- Likely most popular legal copilots – December 15, 2023
- The power of Decentralized ActionAI – January 14, 2024
- Likely most dominant copilots on law (Apex) – January 31, 2024
Point #3 had the second most views on my feed ever, and point #4 had one of the fewest. The above reminded me of this famous quote:
The future is already here – it’s just not evenly distributed
William Gibson
In closing: I made the video below on a beach in Portugal during my summer break in 2023. This blueprint was part of our September 23 webinar. It looked weird when I made it, it makes more sense to me now.
Do you have thoughts or concerns about building ALI or Copilots? Let me know on LinkedIn.
Image Courtesy: DALL·E 3 by OpenAI based on the content of this analysis.