My legal practice, eDiscovery, is steeped in technology, and I am ambivalent about artificial intelligence (AI). On one hand, it facilitates incredible time savings for certain tasks, which translates into significant cost savings for clients. On the other hand, by relinquishing work to AI programs, we run the risk of distancing ourselves from our own expertise.
Online discourse and salesmanship suggest that AI integration is ubiquitous and omnipresent, but I have difficulty mapping that perspective to my own reality. I see specific use cases where there is high value in AI utilization, but I also see a lot of square pegs.
The term “artificial intelligence” is ill defined, and the marketing power of the term has long surpassed its utility. Even though, on plain meaning, “artificial intelligence” would be an insult if ascribed to a person, the term is applied to imbue a product or tool with horizonless possibility and appeal.
As ascribed to myriad technological “solutions”—which often work in fundamentally disparate ways (for example, large language models versus traditional machine learning tools)—“Artificial intelligence takes the work out of work and allows people to focus on higher level thinking” is the general thrust of the pitch. But often, the work itself is the point, and pushing through the more tedious aspects of the work facilitates and crystalizes the higher-level thinking. Analyzing, reflecting, and reconsidering are all critical facets of a finished project, whether it is a piece of correspondence, a memo, or a database project.
As I’ve considered the use of AI tools in my practice, my key takeaway is there is value in applying it to jobs that take a long time to complete and not necessarily to jobs that are “difficult” to complete. If AI can save me tens of hours on a single project, I’m more inclined to use it. If not, I’m inclined to eschew it and focus on my own practice. The clearest examples of both sides of the coin are eDiscovery review and writing.
eDiscovery involves categorization of enormous data sets and was borne out of the idea that we can leverage technology to solve the difficulties presented by other technology. Electronic communications are ubiquitous now, leaving millions of communications and writings potentially relevant to litigation. We use technology to corral those communications into manageable tranches of data and deliver them to the court.
eDiscovery attorneys leverage machine learning programs to classify data and then apply statistical analysis to validate the results. Computers are very good at ingesting pages and pages of data and spitting them out into various buckets based on the criteria we provide. The computer runs a pre-defined routine that can become more complex and account for more variables farther along the decision tree as computer processor technology improves, but ultimately it is prescribed. The attorney then validates the results of this undertaking via established best practices, including manual review of a statistical sample, and confirmation that the results are generally correct.
For this workflow, AI can save lot of attorney time, and a lot of client money, when the goal is primarily to narrow the scope of review prior to in-depth analysis. One million documents can be put into a technology-assisted review workflow in which ultimately only 500,000 are reviewed, after the computer has flagged that all remaining documents are likely to be irrelevant to this inquiry. The attorney validates those results and can then certify that obligations to opposing counsel and the court are satisfied. Note that this workflow still involves a high level of oversight by the attorney and a significant amount of work to account for each step along the way.
Continued developments in generative AI may eventually eliminate first-level human review entirely, so that the only documents reviewed by attorneys are the relevant documents, prioritized and bundled with short AI-generated summaries for case preparation. It would take a team of 20 attorneys working full time about two months to complete 500,000 documents. Now, there are technology solutions available that can complete the categorization via AI in less than a few days. Attorney time required to set up, test, and confirm the prompts, plus time required to quality-check and validate the output, pales in comparison to the time saved on manual review. AI tools are quite adept at cutting through a first-pass review.
On the other hand, I can’t think of a task less suited for relinquishment to AI than organizing my own thoughts in an email or memo to communicate a position or make an argument or to summarize and use the broad categorizations resulting from a first-level review. I’ve spoken to many people who use the tool to get past the “empty page,” but if writing on a blank page is difficult, the only way to make it easier is to do it over and over. It’s a practice, and in practice, we improve. A piece of writing is a communication and statement of the author. I don’t understand the benefit of offloading that work, and I think we’re worse off for it.
As attorneys, we have many resources to assist us in the initial steps of research and writing. Templates and brief banks are standard attorney tools, but these tools still require a level of expertise, experience, and understanding to use them properly. A lawyer’s work is to make clear arguments, and templates and brief banks streamline proving support for those arguments. As they hone their legal practice, attorneys will eventually hold the key cases for their practice area in their head, at which point they become a critical resource for the next generation of lawyers.
Offloading this research and writing to a computer means that we’re continually returning to the well every time we need to draft an argument. That approach may work as long as the well remains in place (which is an open question itself), but I don’t believe it is as efficient as putting in the initial work and ultimately carrying our own water.