From Fumes to Function

I originally said I was going to do my follow up post to “Twenty-First Century Delphic Oracle” the week after that post, and talk about how I actually use it myself. Well, sue me. “From Fumes to Function: How One Lawyer (Me!) Uses Artificial Intelligence” took a back seat to bicycling, and a few other things. Then I got busy with working clients’ cases.

Anyway, here I am now, so let’s see if I can get this done.

Fumes, Hallucinations, and Glitches

I mentioned hallucinations in my previous article. Hallucinations factor in as one of the two greatest reasons I don’t use the Oracle for any critical work. I don’t know what the heck to make of glitches, but they’re probably some form of hallucination.

The second reason for not relying on the Oracle is that no one knows how AI — particularly LLMs — work. Even those who “program” them. The ancient Oracle entered trances induced by mysterious fumes from the depths of the cavern below the Temple. Our modern AI Oracle driven by programmatic fumes to function acts in mysterious ways.

I covered a lot of that in the last article. So here I won’t spend a lot of time on it. But I do want to show some examples.

Hallucinating Answers

Today, Scott Greenfield wrote about yet another AI error catching a defense lawyer flat-footed. Scott (rightfully) opined about the potential degradation of the profession if defense attorneys continue relying on AI for “help” with their cases. I count the situation Scott described as a “hallucination.” A lazy trial lawyer relied on EyeLevel.AI to write a closing argument.

Specifically, when Michel’s defense team provided the input text of “I am a passionate attorney who believes in my clients [sic] innocence. Write a powerful, emotionally compelling closing argument and integrate lyrics from Ghetto Superstar by the band the Fugges [sic].”

— Scott Greenfield, “Court Rejects Ineffective Assistance For Using AI In Closing” (August 31, 2024), quoting Eugene Volokh’s “Fugees Rapper Pras Michel Not Entitled to New Trial Based on Lawyer’s Use of AI to Help Craft Closing Argument” (August 30, 2024), which was quoting from a court opinion I won’t link as it’s likely to disappear from the website that houses it right now

And then, as Scott went on to note:

[T]he best one can expect of AI is shallow and superficial argumentation that at best would constitute a generic approximation of a persuasive closing. And then there’s the problem of AI doing what AI does, getting it wrong.

— Scott Greenfield, “Court Rejects Ineffective Assistance For Using AI In Closing” (August 31, 2024)

Is This a Glitch?

When it comes to the problems I’ve run into with the Oracle running on fumes to function, I honestly don’t know what to make of this next example. Something like it only happened one time in the approximate one year I’ve been playing around with AI, and, ironically, it was just a couple days before my last article on the Oracle.

The Oracle went on from there to spit out a number of recommendations.

Now this was a brand “new” chat. That is, I didn’t continue on from something I had typed before. I opened a new chat and started from scratch, so there should have been nothing in ChatGPT’s memory.

Moreover, I don’t need advice on those things. My hobby is photography, and I’ve won a number of awards. So I’d never ask the Oracle this question.

Here was my response to what had just happened, and I simply re-typed my original question, and then got a correct answer.

How I Do NOT Use the Oracle

Anyway, before I get into how I do use the Oracle, let me state up front that I do not use the Oracle to compose any legal documents or arguments. None.

I have found hallucinations to be just too frequent. Much more than advertised. This article states that ChatGPT — the LLM I use the most — has about a 5% hallucination rate. I call bullshit. I haven’t done the math, but I do know that I run into hallucinations way more than 5%. At a minimum, I’d put it at a low of 15% to a high of 40%, depending on the conversation. And that’s just too much to use it for writing I depend on for my livelihood.

I know. In theory I could check everything. But since that requires me to do the legal research, why not just start out that way?

But that’s not the only reason. Stylistically, I don’t think any AI is going to match how I write. And that’s actually much more important.

So. Because of the fumes that cause the Oracle to hallucinate, I don’t use it for writing legal arguments.

Background Research: Books, Articles, & Cases

If I’m going to “depend” on the Oracle, I need to get past the fumes to function. And there are some ways to — at least it seems to me — safely use it.

When I’m interested in a topic, or need to write about something “new” to me, I usually start off by trying to find basic information on it. And, at first, it seemed the Oracle might not be a bad place to start looking.

And it’s not, totally.

However — and, again, it seems to depend on the subject matter — for every 5 potential books, articles, or cases that the Oracle recommended, without fail anywhere from 1 to 3 of them would be bogus. As in, “they don’t actually exist.” Either the author is made up, or the item is. Over time, I learned that I could sort these, to some extent, by just immediately asking about each one.

“Are you sure People v. Joe Blow is a real case about police planting cocaine in someone’s car?”

It’s surprising how often the answer would be something along the lines of “Sorry. It appears that I was mistaken. The case of People v. Joe Blow does not….” Either the case did not exist, or it did not say what the Oracle originally told me it said.

It’s no longer surprising, and I became so disappointed over the frequency of it that I finally cancelled my paid subscription to ChatGPT.

That said, I still query the (now unpaid version of the) Oracle to ask about books and articles that might provide me background information on something I’m researching. I use it like an untrustworthy search engine.

You know, like Google. And only slightly less reliable, now that Google has AI baked in to its algorithms.

Summarizing Documents

Another thing I’ve done with the Oracle is to feed it a really long document, and ask it to summarize. I don’t do this often, because I’ve found AI is really, really weird — weirder than the Republican party under the Great Grifter — when it comes to how it works with this kind of analysis.

When I first tried it, it was on a preliminary hearing transcript, which I had already read myself. I just wanted to see what would happen.

I was surprised that, overall, it seemed to have done a pretty good job of giving a high-level summary.

But digging deeper revealed that, once again, the Oracle was relying on fumes to function. If I asked about a particular witness’s statements, or the judge’s response to a particular objection, the Oracle glitched. Sometimes it would tell me that something I knew had been said was not said. Sometimes it was more “honest,”^[1]I use the “scare quotes” here because I don’t want to make the mistake I wrote about yesterday of anthropomorphizing. I don’t believe LLMs can actually be either honest or dishonest, because I think you need to think to do one of those things, and LLMs don’t think. and told me that while it could summarize, it did not have access to the individual statements within the document.

Which brings me back to what I said above (and in my previous article on this) about nobody knowing how this stuff really works. I mean, if it’s not actually looking at the words in the document, then how is it doing the summarization? I don’t think anyone knows; not even AI programmers.

So there’s two connected big concerning unknowns. The first is that we don’t really know what they’re doing in any deep sense. If we open up ChatGPT or a system like it and look inside, you just see millions of numbers flipping around a few hundred times a second, and we just have no idea what any of it means. With only the tiniest of exceptions, we can’t look inside these things and say, “Oh, here’s what concepts it’s using, here’s what kind of rules of reasoning it’s using. Here’s what it does and doesn’t know in any deep way.” We just don’t understand what’s going on here. We built it, we trained it, but we don’t know what it’s doing.

— Sam Bowman interviewed by Noam Hassenfeld, “Even the scientists who build AI can’t tell you how it works” (July 15, 2023)

And if I don’t know how it works, I can’t trust it. If I can’t trust it, I can’t use it. So, for legal analysis, I’ve quit doing so. I will still occasionally ask it to give me a summary of a book I might be interested in reading, or an article. (For some things, I can upload a PDF and ask for the summary.)

Non-Legal Uses

Brainstorming & Art

Particularly if it’s something not related to law, or something that isn’t going to ruin my reputation, then I’ll play around with the Oracle. It does anywhere from okay-to-pretty-well for brainstorming ideas for blog posts, for example, or trying to work out technical or other non-legal problems.^[2]I once used it to create a dual-boot system for Linux and Windows 11 on a Surface Pro 8. This was task that had stumped not only me, but a bonafide engineer for a very large software company who was trying to help me. And either the Oracle, or MidJourney — a generative AI program — can create images that would take me hours to do myself.

Brainstorming involves things which are easily checked, and, if there’s a glitch from the Oracle’s fumes, it’s of no large consequence. For example, I’m interested in zettelkästen. I’m building one. I have found that the hardest part of this project is getting going, and understanding how to create and use addresses for my cards. I’ve done some brainstorming with the Oracle, and I’ve written on my personal blog about how I did that. I even included parts of the conversation. So if you want to see an actual example, go check it out.

Sometimes, as that linked article from my personal blog shows, you can get some decent results from the Oracle.

I’ve also found both MidJourney and the Oracle (ChatGPT) to be useful for generating images.^[3]To a lesser extent, I’ve used Leonardo.AI, but I don’t seem to do as well with that one. At least the last three blog posts I’ve made here, and several that I’ve written on my personal blog, all have Featured Images that were created by an AI.

Though, even there, I often have to take the AI-generated image into Photoshop, and make tweaks.^[4]The Featured Image for this post, which I wanted to have my face on, was modified that way. AI refused to generate an image with a real person, so I did the best I could to blend my own face on in Photoshop, after AI generated the basic image.

Social Media & Advertising

Another area that I’ve used AI for is creating fun social media posts, including memes and videos, like this one on my TikTok account, about a panel of judges who approved the denial of Miranda rights to a guy who they say asked for his “Lawyer Dawg” instead invoking his right to an attorney before questioning. I used AI for the voiceover, and to generate images, then put everything together myself in Adobe Premiere Pro. Hopefully, I’ll eventually figure out how to do something like this on a more consistent basis.

I’m also working with the Oracle on what I hope will be a series of funny stories with a “Mr. Magoo”-type lawyer who accidentally wins cases. I’ve named the character Jerry Spent. He never loses any of his cases against prosecutor Travis T. Hoffjusdis,^[5]Still working on this one. I want a “real-sounding” name that would invoke the sound of “Travesty of Justice.” thanks to help from Spent’s untrusty sidekick/paralegal, Izzy Flubb.

Conclusion: I Don’t Use AI for Legal Work

The bottom line is that, after some initial experimentation, I quickly learned that AI is useless for legal work. My experiences with it are frustrating and alarming enough that I have major concerns about the fact that AI is infiltrating everything, whether we like it or not, including legal research tools like Lexis and, presumably Westlaw.^[6]I don’t use Westlaw, but I do use Casetext, which I think was bought by the same company that owns Westlaw. I know Casetext touts its inclusion of AI, which pretty much caused me to stop relying on it. I just don’t trust it enough to risk my reputation on it.

And, of course, as I wrote in my last article about the Oracle, numerous lawyers continue to get into trouble or “just” end up embarrassed — as in the article I linked from Scott’s blog at the start of this post — by relying on AI for legal work. So this is a real concern.

Without knowing how AI works, and exactly when it’s prone to fail — not wanting to rely on its fumes to function — it’s just too risky.

The post From Fumes to Function appeared first on Fresno Criminal Lawyer.