AI vocabulary explained: prompts, training data, and model layers
Free. Watch without an account. Sign in to save progress and take the exam.
Up next
The governing principle: applying old discovery rules to new AI dataTranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.
00:00:00So I don't know, didn't know if that's getting that way.
00:00:02I think that was what I was, and then I think, Mari, you kind of alluded toward training and it's distinct from the use in your earlier comment. I was wondering if you wanted to sort of expand on sort of what that means for lawyers.
00:00:17Yeah, and this ties into the layering point that Karen was just talking about. I guess the way I think about it and the way it sort of overlays with Discovery is you have, you know, the frontier model.
00:00:30Like what we all think about is chat GPT was products, entropics products, sort of what all of this is based on, at least for most of us in the sort of the enterprise tools that we're using.
00:00:40And then the next layer is if you have a third party vendor that is using some of those models to provide a particular service, or if in-house you're doing some sort of fine tuning, there's sort of a layer after that.
00:00:53And then there's, you know, sort of an additional set of instructions that can be based on, you know, your department, your group, your particular system administrator, and then also you as an individual.
00:01:04So there are these multiple layers of, you know, call it context training. Those aren't interchangeable, but they're sort of different steps in the process and different layers that sort of shift what output someone is going to get when they try to get information from any one of these tools.
00:01:22So I think that's maybe another way of saying what Karen was getting at. And I don't know, Daniel, if you had further thoughts.
00:01:29No, I do. I think the one thing to remember is, like, we won't go into the commands of like slash clear and all the different ways to do this, but at a practical level, what you have to understand is that it is a model that was created that then learns from what you give it.
00:01:47Important caveat.
00:01:49If it's corpus, meaning you're, let's say you're using it for decide discovery. If your corpus is ever like you point it to the entire computer rather than just the documents.
00:02:02And there's a real example, right? So let's say you say look for anything relating to soft, okay? Soft ice cream or whatever, okay? We use just to test it. We use to prove a point.
00:02:14Now, what people don't realize is when you upload an entire computer hard drive inside, it may almost every single file's metadata is the word Microsoft for PC.
00:02:24Okay, and the word Microsoft is the word soft. Okay, so your AI will see a whole lot of the word soft within the word Microsoft without having the context of not knowing that has nothing to do with soft ice cream.
00:02:38It can create downstream problems or issues because you're giving it a computer that may have 5 million files on it, of which may be only 150,000 or user generated content files.
00:02:50But all of a sudden, your data set is so skewed that how it's training and learning won't necessarily give you the outputs you're seeking at the level or the function you're looking to do it, just as an example.
00:03:02So that training that Mari is talking about is very important to understand how they work. And as Karen alluded to, it's even more important to understand how these layers work together, right?
00:03:12Because they all work in lockstep, right? One doesn't just do with the other. They all kind of work together. And, you know, you have to have some visibility into it.
00:03:23So there's some terms I know other teaching associates, depending on the firm prompts or engaging with, they call it different things.
00:03:32I mean, I've heard all sorts of names about how to engage or interact with the system and then the output and, you know, an activity log.
00:03:42Now, who used which tool went in with what settings? That's really important for reproducibility sake for you as an attorney, right?
00:03:52And certainly for your discovery, if you're using it for ESI. And then you have, there are different ones where you have system and model terms.
00:04:00So if you're hosting or involved with your own, you have system prompts, then you have training and fine tuning your data.
00:04:07It's way more laborious than you think. And then you have model version or identifiers.
00:04:14There's all sorts of models and all sorts of flavors of models and tweaks to them.
00:04:18And so, and all of these are important words. They all mean different things.
00:04:21And retrieval context embeddings are the documents fed to the model at runtime.
00:04:27But, you know, you have to be very, a lot of organizations don't know all the data they have, which then runs a very large risk to your model.
00:04:36Because if you don't know what data you're feeding it, you think you know what data you're feeding it, but you haven't verified the data you're feeding it, you may be very upset with some of the outputs that you get.
00:04:47We've sort of come to our attention. I don't know, Mari, is there anything you want to add or a definition of availability of these terms or Karen, Mari, you're up.
00:04:58No definitions to add. The only thought I might share is if or when you get to discovery and you're using these terms, I think definitions are crucial because, like Daniel said, so many people are using these words in different ways.
00:05:14And so just like now when we have document requests that we send that have 20 pages of definitions before you actually get to the request, I think those pages of definitions are going to grow exponentially as we all try to define these.
00:05:26But I think that's an important piece of this.
00:05:28Yeah, but you should use like, like one great takeaway for today just occurred to me because maybe lawyers don't even think of this.
00:05:36If you actually are involved in an AI dispute, you should define what these terms mean. Like prompt engineering means something much different to a software developer versus a user versus different people use these terms radically differently.
00:05:52And those definitions to Mari's point, if you don't define them, you're going to be paddling upstream, really unhappy for a very long time. I don't care anything you want to add.
00:06:04No, I would just double click on that, which is that we have two problems. One is that people are using the same words to mean different things and also that people are using different words to mean the same thing.
00:06:15So if you don't, if you're not level setting like what you mean when you're asking for a thing or when you're producing a thing, for that matter.
00:06:25It's just you're going to, there's going to be a lot of confusion downstream. And we see that in the governance world all the time.
00:06:32Yeah, I was just thinking governance, like literally that must happen all the time because you tell a software engineer to come up with governance.
00:06:40They're thinking about a model and nothing to do with like governance in the traditional sense of what lawyers associate governance with about controls.
00:06:52We're inventing all these new words like assurance or we're using words we know like assurance to sort of describe a category of governance, but nobody's really clear on the boundaries of that.
00:07:02And so I think lawyers have to just resist the urge to think we know what is meant when even familiar words are being used to describe this stuff.
00:07:12We may get a point when we do any sort of development to define the core terms for any project.
00:07:18Yeah, we do.
00:07:21Yeah, and you don't want, you don't want the person, you know, ultimately deciding that, you know, a judge potentially in federal or state court to be walking into deciding a dispute without knowing how you want the term to find and needing to decide sort of on their own in a black box.
00:07:36So another reason.
00:07:38I mean, their clerks will help.
00:07:40Yeah.
00:07:44Sorry, I love law clerks, they're great. But I mean, they may be practice law never, or maybe for a year or two, and have no real exposure to some of these concepts. So it's a great point.
00:07:56And equally importantly, they're probably not technical, right? So things that it sounds like it's reasonable to ask for, or demand, or won't be.
00:08:07Right. And so you're missing it on both ends.
00:08:10Yeah, 100%. We see as an arbitrator, I see requests for like, I often find it entertaining because they come in and say we'd like all of the weighted scores for a model.
00:08:22And I'm like, you don't have enough, you can't afford to host the data, like, let alone, like, just the math, like, what do you mean? I'm like, that's about 400 terabytes of data that you're going to have to host.
00:08:36Yeah, let alone utilize.
00:08:39Use, figure out how it was used, filtered, processed, or anything else. There's just this huge disconnect.
00:08:47Well, think about all the, I mean, all the times we've said all documents related to, right?
Automatically generated; it may contain errors.
Study materials
Notes, flashcards and practice questions for this lesson, reviewed before publication.
Study notes
Why discovery lawyers need precise AI vocabulary
The speakers agree that AI-related terms are used inconsistently across people and organizations. Two distinct problems were identified: people using the same words to mean different things, and people using different words to mean the same thing. Without level-setting on definitions when requesting or producing information, confusion results downstream. This is compared to how discovery requests already carry many pages of definitions before the actual requests begin, and the speakers expect these definitional sections to grow substantially as parties try to define AI-related terms.
Layers of AI models from frontier model to individual use
Mari Grace describes AI systems as built in layers. At the base is the frontier model, such as the products associated with ChatGPT or Anthropic, which underlies most enterprise tools in use today. The next layer involves a third-party vendor that uses those frontier models to provide a particular service, or an in-house fine-tuning process. Beyond that is an additional layer of instructions that can be set by a department, a group, a particular system administrator, and finally the individual user. These layers represent different steps of what can be called context training. They are not interchangeable, but each layer shifts what output a user ultimately receives when seeking information from a tool.
How training data can get corrupted: the metadata example
Daniel B. Garrie explains that a model is created and then learns from whatever corpus it is given. If, in a discovery context, someone points an AI tool to an entire computer rather than to just the relevant documents, problems can arise. As an example, a search was run for "soft ice cream." Because nearly every file's metadata on a PC contains the word "Microsoft," which itself contains the string "soft," the AI encountered the word "soft" constantly without the context that it had nothing to do with ice cream. On a drive with millions of files, of which only a fraction may be user-generated content relevant to the task, the data set can become so skewed that the model's training and learning no longer produce the outputs sought at the level or function intended.
Core technical terms: prompts, system prompts, fine-tuning, model versions, and retrieval context embeddings
Daniel B. Garrie notes that different teams and firms use different names for how a user engages with an AI system, for the output, and for the activity log that records who used which tool with what settings. That activity log matters for reproducibility and for discovery involving electronically stored information. For organizations hosting or building their own systems, relevant terms include system prompts, training and fine-tuning of data (described as far more laborious than people expect), and model version or identifiers, since there are many models and many variations or tweaks of them. Retrieval context embeddings refers to the documents fed to the model at runtime. A related risk is that many organizations do not actually know all the data they hold, which creates risk because feeding unverified data into a model can produce outputs the organization did not expect or want.
Why discovery requests need explicit definitions for AI terms
Mari Grace states that once these terms come up in discovery, definitions are crucial because so many people use the same words in different ways. She compares this to current document requests that already include many pages of definitions before the substantive requests, predicting that these definitional sections will grow substantially as parties try to define AI-related terminology. Daniel B. Garrie adds that a key takeaway is that parties involved in an AI dispute should affirmatively define what these terms mean, since a term like prompt engineering can mean something very different to a software developer than to a user or to other different people, and failing to define such terms early creates significant downstream difficulty.
The risk of same words, different meanings, and different words, same meaning
Karen Silverman frames this as two separate problems: people using the same words to mean different things, and people using different words to mean the same thing. If parties are not level-setting on what a term means when asking for or producing something, confusion results. This pattern is also observed in the governance space, where a software engineer asked to build "governance" may think of model-level governance, while a lawyer may think of governance in terms of traditional controls, and new or borrowed words such as "assurance" are used without clear agreement on their boundaries.
Judges, clerks, and the risk of deciding AI disputes without defined terms
The speakers caution that a judge deciding a dispute, whether in federal or state court, should not be left to interpret undefined AI terminology on their own. Law clerks may have practiced law only briefly, with little exposure to these technical concepts, and are also likely not technical themselves, creating a gap on both the legal and technical sides of understanding a dispute.
The gap between what parties request and what is technically feasible
One example given involves an arbitration request asking for all of the weighted scores of a model. The response noted that such a request would involve an enormous volume of data, roughly 400 terabytes, well beyond what most parties could host or afford, let alone use, filter, or process. This illustrates a broader disconnect between the scope of what is requested in discovery, such as "all documents related to," and what is actually feasible given the scale of AI model data.
Flashcards
13 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 13
Practice quiz
Practice — not graded, not the final exam