What generative AI actually is and how it produces mutable outputs
Free. Watch without an account. Sign in to save progress and take the exam.
Up next
AI vocabulary explained: prompts, training data, and model layersTranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.
00:00:00Some recent decisions and then some best practices.
00:00:03Please follow along using supplemental materials in the slides.
00:00:07If you do have any questions, feel free to email us if they relate to the presentation.
00:00:12If they relate to testing your platform, don't ask me to check it out.
00:00:16But maybe my panelists aren't so inclined. I'm not going to respond.
00:00:20But if you have questions about the material, I'm happy to take any questions.
00:00:24So what is generative? So I'm going to just focus the conversation.
00:00:29There's all kinds of AI. We're going to talk right now about generative AI.
00:00:36I'll let maybe Mari, you want to kick off and then Karen fall on and I will round it out.
00:00:41Mari?
00:00:44So I guess just high level, if we're covering the basics in terms of what generative AI is, like Daniel said, it's more than what we're all accustomed to.
00:00:55It's not just searching and retrieving documents.
00:00:58It's not searching through a set of sources and just pointing to you to what already exists.
00:01:03It truly is generative. It is coming up with new content using patterns that it has found across its training model.
00:01:11So that's at a high level. I don't know, Daniel, how much detail you want to share.
00:01:15I don't think it's, Karen, anything you want to add to that?
00:01:18No, I just think that that's both the magic and the mischief of it, right?
00:01:24Which is that because it's generating unique content each time, even the same prompt or query asked in succession can yield different results, right?
00:01:34And so from a lawyer's perspective, and we're talking about things that are fixed or known, generative AI by itself without any of the kind of apparatus that I think we'll talk about is highly mutable and by design, right?
00:01:51And so what do we do with that from a perspective of evidence?
00:01:54Yeah, totally. And one other thing I might add for those who haven't, you know, are new to this is that it really isn't going and finding a correct, discreet answer, even though it might seem in any particular tool when text is showing up and giving you something that is beautifully written.
00:02:12It actually is just identifying patterns. So in the huge volume of text and data that it has been trained on, you know, what is the most likely next word in this sentence?
00:02:25It is not giving you source of truth. This is accurate, which I think is a distinction that is hard to draw when, especially now, what the models put out is so well written.
00:02:36Yeah, and I think that's a great, great segue into, and it's very dangerous if your associates or people you task with this process don't understand that distinction.
00:02:47It is a distinction with a big difference for your end result. As Karen said, you can run the same prompt, and you will get two different results possibly, I mean, more likely than not, without the right harness and other things around it.
00:03:02It's like, I think, Mari, you hit on the base. I mean, statistical model training is distinct from its use. So I think, Karen, you were sort of talking about it, and, you know, and you already mentioned the outputs aren't reproducible by default.
00:03:17And you also mentioned enterprise AI is layered. So maybe, Karen, you want to explain a little more what you meant by layered, because I think lawyers are, I don't want to say mystified by this construct, but I think would be a good place to sort of touch on a little bit more.
00:03:31Yeah, so I think it's, I mean, one way, one thing that might be a little bit useful is that a lot of these generative AI tools are more in the nature of, we think of them as software, that's okay.
00:03:44But if they are software, they're more like a word, right? Which is that it only is valuable in relation to the other things you're doing with this tool, right?
00:03:55It doesn't come as a fully functioning app or solution. It's a paint set. And so what are you going to do with the paint set? And what purpose are you going to put it to?
00:04:09And as we were talking about earlier, because it's essentially, it's not only recognizing patterns and producing, you know, unique content, it's doing that on the basis of probabilities, right?
00:04:24So it's heavily influenced by its training set and what was dominant in the training set in relation to the question, and you're now asking it in some new context.
00:04:37And so to control for this variability, there are layers of technologies and methods and kind of techniques that are placed around it.
00:04:49Daniel, you mentioned harness, there's a bunch of different definitions of harness, but this AI, this generative AI tech stack is starting to include tools to control what the technology can access while it's generating new content that control the tethering of the outputs of this content to known truths and reality.
00:05:13Ideally, a reality that matters and is relevant and contextual. There are tools that check provenance of data that you're using to make sure that your inputs are fit for purpose for the tool and for the output you're seeking.
00:05:32There's this whole stack and set of layers of, I really think of them almost like techniques to try to take this general paint set and make it more reliable for its use in legal contexts.
Automatically generated; it may contain errors.
Study materials
Notes, flashcards and practice questions for this lesson, reviewed before publication.
Study notes
What generative AI actually is, as explained by Mari Grace and Karen Silverman
Mari Grace describes generative AI as something more than the search-and-retrieve tools people are used to. It does not search through a set of existing sources and point users to what already exists. Instead, it is truly generative: it creates new content by using patterns it has found across its training model.
Karen Silverman adds that generative AI is not going out and finding a single correct, discrete answer. Even when a tool produces text that looks beautifully written, it is really just identifying patterns in the huge volume of text and data it was trained on, and predicting the most likely next word in a sentence. It is not giving a source of truth or a guarantee of accuracy. Karen Silverman notes this distinction is hard to draw precisely because the output from current models is so well written that it can seem authoritative.
Why generative AI outputs are mutable, per Karen Silverman
Karen Silverman frames this as both the magic and the mischief of generative AI: because it generates unique content each time, even the same prompt or query asked in succession can yield different results. From a lawyer's perspective, where the focus is usually on things that are fixed or known, generative AI by itself, without additional apparatus around it, is highly mutable by design.
This raises a direct evidentiary question that Karen Silverman poses to the group: what do we do with that mutability from the perspective of evidence?
Daniel B. Garrie on the risk of not understanding mutability
Daniel B. Garrie states that it is dangerous if associates or others tasked with using generative AI tools do not understand the distinction between pattern-based generation and a fixed, accurate answer. He describes this as a distinction with a big difference for the end result: running the same prompt can produce two different results, more likely than not, without what he refers to as "the right harness and other things around it."
Daniel B. Garrie also summarizes a key point from Mari Grace's explanation: statistical model training is distinct from its use. He notes that outputs are not reproducible by default, and that enterprise AI is layered, setting up the next part of the discussion.
Karen Silverman's explanation of layered generative AI tools
Karen Silverman explains the idea of "layered" enterprise AI by comparing generative AI tools to paint rather than to a finished piece of software. These tools are only valuable in relation to what else is being done with them; they do not arrive as a fully functioning app or solution. The relevant question becomes what purpose the user puts the "paint set" to.
Because the underlying system recognizes patterns and produces unique content based on probabilities, it is heavily influenced by whatever was dominant in its training set in relation to the question being asked in a new context.
Layers used to control variability, according to Karen Silverman
To control for this variability, Karen Silverman describes a generative AI tech stack that includes several categories of tools and techniques:
- Tools that control what the technology can access while generating new content.
- Tools that tether outputs to known truths and reality, ideally a reality that is relevant and contextual to the task.
- Tools that check the provenance of the data being used, to confirm that inputs are fit for purpose for the tool and the output being sought.
Karen Silverman describes this whole stack as a set of layers or techniques meant to take the general-purpose "paint set" of generative AI and make it more reliable for use in legal contexts.
How this discussion sets up the evidentiary question
Taken together, the speakers describe generative AI as a pattern-based, probability-driven technology that produces new content rather than retrieving existing sources, that is mutable by design absent additional controls, and that requires layered tools around it before its outputs can be treated as reliable enough for legal use. This framing is presented as the foundation for the lesson's later discussion of recent decisions and best practices.
Flashcards
12 cards. Test your recall, then flip to check.
Select the card, or press Space, to flip it. Left and right arrow keys move between cards.
1 of 12
Practice quiz
Practice — not graded, not the final exam