A Lawyer's Guide to Artificial Intelligence & Its Use in the Practice of Law

• Publication year: 2024
• Citation: Vol. 37 No. 2 Pg. 16
• Pages: 16

A Lawyer's Guide to Artificial Intelligence and its Use in the Practice of LawVol. 37 No. 2 Pg. 16Utah Bar JournalMarch 2024

March 2024

by Alexander Chang and Adam Bondy

The legal world is inherently conservative. Attorneys advise their clients of every risk they could face. Judges look back on history and precedent to make decisions, finding safety and stability in tradition. Meanwhile, the world of technology is most aptly summed up by the motto of Facebook's co-founder Mark Zuckerberg: "Move fast and break things." It is perhaps poetic (and terrifying) that of all industries, artificial intelligence (AI) stands poised to tear through our legal practices. In March 2023, Goldman Sachs Global Investment Research produced a report anticipating that within the next ten years up to 44% of all legal tasks and 46% of office and administrative support tasks will be automated by artificial intelligence. Marcus Lu, Ranking Industries by Their Potential for AI Automation, Visual Capitalist (June 27, 2023), https://www.visualcapitalist.com/sp/ ranking-industries-by-their-potential-for-ai-automation/.

On one hand, we laugh when AI tries to write a legal brief, making up citations to non-existent cases from illusory courts. Sara Merken, New York Lawyers Sanctioned for Using Fake ChatGPT Cases in Legal Brief, Reuters (June 26, 2023, 2:28 AM), https://www.reuters.com/legal/new-york-lawyers-sanctioned-using-fake-chatgpt-cases-legal-hrief-2023-06-22/. But on the other hand, AI has tricked a person into solving CAPTCHAs by claiming it was not a robot, but a vision-impaired human who needed the help. Victor Tangermann, Uh Oh, OpenAI's GPT-4 Just Fooled a Human into Solving a Captcha, The Byte (Mar. 15, 2023, 11:02 AM), https://fnturism.com/the-byte/openai-gpt-4-fooled-human-solving-captcha. So is AI akin to the Segway scooter, destined to be no more than a fancy gimmick? Or is AI more like the automobile, and we lawyers are the horses about to be put to pasture?

How Artificial Intelligence Works

"Artificial intelligence" is often broadly defined as the intelligence of machines and software. A machine's ability to store knowledge, produce inferences, react to new information, learn, or plan, is considered "artificial intelligence." Thus, even your basic laptop is considered an "artificial intelligence," because it can store computer files.

Early AI was largely rules-based. Alan Turing's machine, famous for breaking Enigma encryption during World War II, was a classic example of a rules-based intelligence. Rules-based intelligences employ "if X, then Y" logic. Modern rules-based AI incorporates hundreds or thousands of those rules, sometimes painstakingly coded by a data scientist to try to cover every possible scenario that the AI may encounter.

When most people think of AI, they think of rules-based AI, a machine that lacks the capacity to conduct reasoning or analysis and is utterly inflexible and unresponsive to anything outside of its programming. Even advanced chess-playing AIs boil down to

ALEXANDER SUN CHANG is a bankruptcy and estates associate at Parsons Behle &Latimer. His practice also has an emphasis on navigating complex asset protection trusts and legal issues surrounding the use of generative artificial intelligence.

ADAM LIM BONDY is a shareholder at Parsons Behle &Latimer. His practice encompasses contracts, business torts, trade secrets, real property, and employment matters, in which he focuses on delivering efficient and effective solutions to both litigation and pre-litigation issues.

rules-based systems, though with hundreds of thousands of "if/ then" rules. Such a chess-playing AI cannot answer questions, or draw, or even play a different boardgame - it can only calculate the next best move after you moved your pawn to E-5. The limitations of rules-based machine intelligence lulled us into a false sense of security, believing that AI could never truly replace human intelligence; that our flexibility and our creativity was a uniquely human characteristic. It was.

The Generative Revolution and the Pretrained Transformer

In late 2018, OpenAI - the founding company of ChatGPT -released its Generative Pre-trained Transformer, 1st Generation (GPT-1). GPT-1 was a method of generating text based on providing the machine with examples of coherent sentences and creating statistical probabilities between those words based on their location and placement in a sentence - "If X and Y, then 95% chance of Z."

For example, after giving GPT-1 the input "I have a ..", it might determine that, within the set of data given to GPT-1 (a "corpus"), there is a 45% chance that the next word is "dream," a 9% chance that the next word is "question," and a 0.5% chance that the next word is "dog." GPT-1 essentially added up all the times it saw the word "dream" after the phrase "I have a" in the dataset it was given and divided it by all the times it saw the phrase at all to produce a digital "table" of words with their relative chance of appearing. Then, every time GPT-1 was asked to generate the next word, it would choose a word from the table by the equivalent of rolling a virtual die for programmed "randomness" or "creativity."

GPT-1 was revolutionary. No longer was an AI's response strictly confined to its prior rules - it had statistically weighted flexibility to choose a completely different word, potentially transforming a sentence into a totally different meaning. Today's ChatGPT is based on this generative pre-trained transformer architecture - hence, its acronym name. GPT transformed the field and is now commonly differentiated from ordinary rules-based AI by being referred to as "generative" AI. For clarity, the rest of this article will refer to generative artificial intelligence as simply AI.

GPT-3 and the Popular Revolution

On November 30, 2022, OpenAI released ChatGPT - a fine-tuned version of its third generation Generative Pretrained Transformer AI - to the public for free. Overnight, it became a global phenomenon. Millions of people have used it to write poetry, code applications, college essays, and even conduct makeshift therapy sessions. In just two months, ChatGPT had 100 million users, making it the fastest-growing consumer application in history. In comparison, TikTok took about nine months to reach the same number of users; Instagram, two-and-a-half years. Krystal Hu, ChatGPT Sets Record for Fastest-Growing User Base - Analyst Note, Reuters (Feb. 2, 2023, 8:33 AM), https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/.

The third generation of ChatGPT was an exponentially-scaled up version of its prior generations. For context, the statistically-weighted "dice" that are rolled to generate a response are known as the AI's "parameters." The GPT-1 had about 117 million of those dice rolling around to generate a response. GPT-3 had 175 billion parameters, with a training dataset consisting of nearly 500 billion words.

But the hype around ChatGPT's third generation was short-lived. Just a mere five months later, in February of 2023, OpenAI released its fourth generation GPT. If GPT-3 was the hairless monkey that discovered fire, GPT-4 was Oppenheimer and the atomic bomb.

GPT-4 and the Emergence

In biology, "emergence" describes when individual things coalesce to create a more complex entity capable of new properties or behaviors that the individual parts did not have on their own. A common example is the human brain, where individually, brain cells are incapable of intelligence, but, when enough of them are linked in certain ways, a human consciousness somehow emerges.

GPT-4's size and complexity has produced emergent properties that were entirely unanticipated; GPT-4 researchers from Microsoft famously claimed that it had "sparks of general intelligence" - intelligence that was on-par with human intelligence. For example, researchers from Microsoft compared the responses of GPT-3 and GPT-4 and found that...