In 2016, nearly two decades after IBM’s Deep Blue defeated the chess grandmaster Garry Kasparov, a computer program named AlphaGo beat Lee Sedol, the reigning world champion in Go. Go is one of the most complex and strategic games in the world. Its possible combinations exceed the number of particles in the visible universe. This means that no software would be able to win by simply memorizing games. It was a feat that marked a new era in artificial intelligence. Three years later, Sedol retired citing AI as the cause. The best human player of the most complicated game had lost to an AI. But the AI did not remain undefeated. Six years later, at the end of 2022, while the world marveled at the release of ChatGPT, an amateur Go player took back the title. How? Well, it turns out AI is super dumb.
The term “artificial intelligence” is a total misnomer. When you make a request to a large language model (LLM) like ChatGPT, what it does next is select the most appropriate word to follow your prompt. That’s it. Once it has selected the most appropriate next word, it repeats the process until the response is completed. What the “brains” of a LLM is, how it is trained, and what the actual process of deciding what is the best next word is complicated and interesting, but the result is that they do not “know” anything. Ask ChatGPT to produce a list of social media post suggestions or to write a blog post, and it has no knowledge of what social media is, what a blog is, or even what writing is. It merely is picking the next word that should follow your request and then picking the next word after that, repeating until there is no next word. LLMs are extremely good at this task and the results look remarkably similar to intelligence, but it isn’t. That’s the secret that allowed a human amateur to reliably beat the AI that trounced the Go world champion. The AI has no understanding of what “Go” is, what “groups of pieces” are or even what a “game” is and the strategy took advantage of that lack of understanding. It was a strategy so simple that any human player is able to defend against it, so it is never used in real games. It turns out that the AI didn’t “know” how to play Go, it was simply picking what the next best move was based off its training. Understanding this limitation and its implications can teach us a lot about how to use this revolutionary technology. What does this mean for your use of AI within your practice? It’s simple.
It’s very easy to attribute undeserved “intelligence” to LLMs like ChatGPT. After all, the results they produce can be astonishing, but treating LLMs as if they are intelligent will lower the quality of the responses they generate and may even get you in trouble if you use those responses without proofing them. If you want to get the best results from LLMs, remember that they do not “understand” or “know” anything. They don’t know what marketing is, who your audience is, or what the law is. They are simply extremely good at picking the next word from your prompt. With that perspective, ask yourself how would you manage an employee that is incredibly talented at writing, but is otherwise clueless? A brand-new employee on their first day of the job. What information would you need to provide to them before giving them an assignment?
Tips for writing an AI prompt
- Identify your perspective. LLMs don’t know who you are, what you do, or what your industry is. When you are writing a prompt for AI, identifying yourself will always lead to better results. Try starting it with, “We are an estate planning firm based in San Diego that focuses on providing exceptional client service.”
- Be specific in what you are asking. When writing your prompt, be extremely specific about any information you want contained in its response. Include any details you want it to include, the length it should be, the recommendations it should make. For example, “Write a two-page blog post encouraging people to update their estate plans after major life events. Use a metaphor of seasons changing and mention the fact that we are approaching the start of the fall season.”
- Identify your audience. Your LLM will not know who your target demographic is, what their interests are, or even what a “person” is. You must provide that information. For example, “The audience are married couples above the age of 60 who are retired.”
- Identify your tone. Should the writing be formal, technical, or casual? Specifying what you want your language to sound like is incredibly important. For example, “The tone should be casual yet respectful.”
- Give feedback and ask it to do it again. This last tip is likely the one that will make the biggest difference. Sophisticated LLMs like ChatGPT “remember” the previous parts of the conversation. More specifically, they use the conversation’s history to aid in the prediction of what the best next word is. So just like any work product from any employee, you want to review the results, provide feedback, and ask for another draft. A couple rounds of back and forth will make an enormous difference. Say things like, “This is a good start, but make the introduction more concise, remove paragraph two, and expand on the impact of an outdated trust.”
Here’s what a full prompt would look like:
“We are an estate planning firm based in San Diego that focuses on providing exceptional client service. Write a two-page blog post encouraging people to update their estate plans after major life events. Use a metaphor of seasons changing and mention the fact that we are approaching the start of the fall season. The audience are married couples above the age of 60 who are retired. The tone should be casual yet respectful.”
Once the LLM has produced its response, proof it and then provide feedback. Continue doing this until you are happy with the results. The process should be no different than if you were training a new hire. The benefit of doing it this way is that you can then continue using that same conversation to make new requests and the LLM will produce better results because you will have “trained” it.
The Correct Task for the Correct AI
Would you ask a productive and capable document production paralegal, who has no interpersonal skills and no client communication training, to answer your clients’ calls? Obviously not! Understanding the strengths and weaknesses of your team allows you to be an effective leader and allows your team to do their best work. You wouldn’t assign a task to an employee without considering their skillset, so don’t assign a task to an LLM without considering what it is trained to do. That mistake is exactly what led one law firm into massive legal troubles after they treated ChatGPT as if it was a legal research tool.
The different tasks that can be done by AI is growing exponentially day by day, and the best AI for each task changes too fast for me to make a recommendation on which tools to consider. That said, I have used many different LLMs to do many different things and it is important to know what the potential uses are. Here are a few of the tools I’ve explored and their uses:
- ChatGPT: Text generation, assist in writing blog posts, produce lists of suggestions, edit or proof documents, create checklists, refine corporate values
- MidJourney: Image generation, marketing images, social media images, creation of logos, creation of avatar photos. The photo for this blog post was created using this
- Otter.ai: Meeting notes, voice to text transcription
- D-ID: Text to voice generation which will also animate your face and create videos
- Tableau: Data Analysis
- Descript: Video editing by editing the written transcript
Take Interest in AI’s Development
Every employee becomes more successful after their supervisor takes an interest in their development. If you want to be able to benefit from this revolutionary technology and ensure that your firm is not left behind, you need to stay informed on its development. Ignorance towards AI today is the equivalent of ignorance towards the internet in the mid-nineties. Engaging with AI, understanding its capabilities and limitations, and integrating it thoughtfully into your practice will keep your firm at the forefront of innovation, reaping the rewards of efficiency, creativity, and competitiveness that this incredible technology has to offer.
Be sure to register for Rita Chaires’ How AI Can Help You Win at Online Marketing presentation at the upcoming Fall Summit. Not yet a Member? Talk to one of our Membership Consultants about joining us this Fall in beautiful San Diego, California. Call (800) 846-1555 or email email@example.com to learn more.
Practice Building Coach
American Academy of Estate Planning Attorneys, Inc.
9444 Balboa Avenue, Suite 300
San Diego, California 92123
Phone: (858) 453-2128
- Your Dim Yet Brilliant Employee: How To Get The Best Out Of AI - August 10, 2023
- A Robot Wrote My Blog - February 16, 2023
- Are You Really Bad at Your Job? - September 15, 2022