Pragmatic AI with Matt Stauffer

Matt Stauffer:
Welcome back to Pragmatic AI, where we're working to bring a clear eyed look at how AI interacts with the real world. My guest today is Sarah O'Keefe, founder and CEO of Scriptorium. Sarah

Thank you so much for joining me. Can you say hi to the people and tell who are you and what do you do and what is Scriptorium?

Sarah O'Keefe:
Hello, people. my name is Sarah O'Keefe. I run a company called Scriptorium, which is based out of North Carolina. We have been around for almost 30 years, which is truly, truly horrifying.

Matt Stauffer:
Wow.

Sarah O'Keefe:
and our job is to understand how to bring technology into enabling content. So technical product, learning, support content, all the customer-facing stuff that helps people do their jobs.

And so we're interested in questions around how do you do that, how do you do it efficiently, how do you translate efficiently, how do you deal with variants at scale? Most of our customers are large to extremely large enterprises that have run into problems with content operations, with efficiency, with a merger, and now they have like 18 different content systems in the organization. What do we do? That's my word.

Matt Stauffer:
Okay.

And is your work primarily about the content or are you also doing technical integrations like are you doing CMSs and stuff, or is it more just about the workflows and processes for content itself?

Sarah O'Keefe:
So we look like technology consultants. We do a lot of work around content management and specifically component content management systems, topic

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
based content. And really at the end of the day, it's change management and organizational change development because the problem is always the people, not the tech. I mean the tech has plenty of problems, but it's very, very straightforward compared to the people.

Matt Stauffer:
Right. Yeah. Yeah. Sure. It's not the source. Yeah.

Okay. And you said you work largely with enterprise. Is it business, non profits, a little bit of everything? Basically, if they're big enough to need your help, you're there for it?

Sarah O'Keefe:
It's it's a little bit of everything. really the question is how much content do you have and are you running into problems scaling that content? So

Matt Stauffer:
Okay.

Sarah O'Keefe:
typically it's large enterprises, but we work with a number of not for profit organizations that also have just huge volumes, standards organizations, that kind of thing. Anybody that has lots

Matt Stauffer:
Yeah, okay.

Sarah O'Keefe:
of information.

Matt Stauffer:
Yeah, that makes a ton of sense. Okay, so I can imagine some ways, and also obviously,

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
spoiler, I've watched some of your content, but imagining I had not I could imagine some ways AI has come to touch on your world, but I want to ask you specifically, what is your broad level of interaction with AI, both from a personal and from a work perspective? Are you an early adopter? Are you a late bloomer? What's you know, how do you interact with AI?

Sarah O'Keefe:
so it's interesting because the information that we deal with is being fed into and pushed out of AI systems. So what I would say is that we are a an advisor trying to deal with the question of what's going to happen when we put this content into AI and

Matt Stauffer:
Yeah.

Sarah O'Keefe:
also what's going to happen when we pull it out of AI. So we see it mostly as a tool, just like anything else.

And the thing my number one position I think when it comes to AI right now is actually one of pure frustration because it can't do what people are trying to get it to do.

Matt Stauffer:
okay. Well, we need to go there and I d I didn't even know you were

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
gonna say that. I can imagine some frustration because I've so I had a hu a couple SEOs on the podcast and I know that SEO is very mer much this thing where you don't control the systems, right? And so like you want a certain output, but you don't own it, and so you're just left from the outside kind of throwing rocks at it, hopefully you can get it to do the thing you want. And I can imagine that you're usually primarily working in systems where again, like you said, you're mainly working on people processes.

But additionally, if they are doing content distribution, it's probably in their own internal platforms. So when you first said that, my thought was, you you're probably going to be frustrated with it because, you know, all of a sudden the distribution is now happening outside of something you can control because of some massive lab. But you're you s you were saying frustration because it's not doing what we're asking it to do. What are we asking it to do that it's not equipped to do well?

Sarah O'Keefe:
Okay, so it's you know, it's not alive and it's not sentient, notwithstanding all the hype. and LLM fundamentally and and when we say large language model, that's actually not exactly what the public facing bots are at this point because they've added on a bunch of technology and a bunch of additional pieces and parts to help. But

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
fundamentally, a large language model puts together words, right? It's it's autocomplete, but better.

Matt Stauffer:
Right. Mm-hmm.

Sarah O'Keefe:
And then you tack on things like retrieve log manage generation and various other things on the back end, and you can sort of put guardrails around that to help it do what it needs to do. But let's step back from that. Why am I frustrated? As a product and technical content enabling content producer, what is my job, right? Or what is the job of our clients? Their job is to produce the information that enables people to do.

Their job, right? To get the thing done. You pick up a piece of software, you can't figure out how to use it, you go read the reference content, you go read the API docs, you go read whatever. You go look at a video that says, and you say, I didn't see that button. Now I see how to do the thing. So fundamentally, enabling content is about helping other people to do what they need to do. Now, in that context, when you start talking about

certain kinds of content, medical device content, there are regulations. In the US, we have the FDA. and in Europe you have other regulatory agencies and all around the world. When you're dealing with regulated content, fundamentally the regulators frown on the idea that you are making things up.

Matt Stauffer:
So you're not off to a good start with AI.

Sarah O'Keefe:
So you're not right. Because what AI

Matt Stauffer:
Yeah.

Sarah O'Keefe:
does is it generates things that are fluently written and look plausible. Now, I love my marketing side colleagues. You're all great. on the marketing side, you generally have a little bit more leeway for, that's not quite accurate. We'll we'll say kindly. We're not gonna use the L word.

So there's a little bit more leeway there. In the content that I'm dealing with, there's no leeway. Now, if I take all my corpus of information, right, that my client has produced 40, 50, 80,000 pages of content in varying degrees of cleanliness, and then we take that and we feed it into some sort of a large language model, whether public or private, and then we say generate.

the updated version of this documentation, what it's going to do is going to be potentially not accurate. And so what has happened is that I, as the author now, have been we've injected a middleman. All right. We've injected the AI as the middleman between me and the end user, consumer of the information.

So we're producing all this synthetic content, which may or may not be accurate, and we can, you know, go around our butts to try and make it accurate. But maybe it's accurate, maybe it isn't. And if it's not, we're gonna be in big trouble.

Matt Stauffer:
Yeah.

Sarah O'Keefe:
So how do we make it accurate? And now we sort of circle back to how do we make sure that the content is actually clean, accurate,

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
up to date, properly versioned, and all the rest of it, because that gives us some fundamental chance. But what's happening with AI in

content broadly is that the people who are the publishers of the content, the owners of the content, are being displaced or pushed back such that the the machine, the AI, is getting in there and is injecting itself as a middleman between the people creating the content and the actual end deliverable. Now

Why do the consumers, the end consumers, find the AI potentially so appealing, right? Why do they prefer that to, hey, go read this carefully crafted document? That's a failure of the content that we've been producing, right? Because it's not it's not good enough, it's not appealing enough, it doesn't give you that same happy, I asked the AI and it gave me an answer right away, and it gave me exactly the answer that I asked for, and it

Matt Stauffer:
Right.

Sarah O'Keefe:
looks good.

Right. So there's a there's a very, very big disconnect between what is the job of somebody producing technical content and why why are you being it's not disintermediated, right? It's like re disintermediated.

Matt Stauffer:
Yeah. so this is a conversation that's happening actually today in my section of the world. And I know it's not a perfect overlap with yours, but I I want to get your insight

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
into what we're experiencing. the founder of a technology that I work with, just put out a tweet literally this morning that said, I used to write our documentation primarily facing on the human being, and I wanted to make sure that I wasn't overloading them with way too much, you know, explanations, whatever, because I needed to make sure they could have a journey through this explanation.

but the downside was that I wasn't always able to put all the specifics that I think might actually have been helpful. I'm now writing primarily for AIs because people are not going to the docs. They're going to the AIs. So I'm making much more expansive documentation so that the AI can interact with it. So when I saw that, I was definitely curious about what he was thinking there. But as I'm thinking about what you're saying, I think to me the number one reason why people are going to AI, for example, for our technical docs, is because they're like, I could read.

you know, 300 pages of everything like that. I can try and Google and hope Google's gonna help me find it or use the built-in search feature. But getting down to the one specific thing I need has been historically relatively difficult. Whereas the AI, if I say, here's my exact and specific scenario, can you give me the the place to find or the answer right there? So I I definitely think that it's ease of getting to your very specific answer in that moment. And one thing that you mentioned was you said it's a it's a failing of the original content.

So is there a vision that you have or would have had in the past for how we should differently either write or present content that would have not necessitated the AI replacement? Or like what's the solution to that problem?

Sarah O'Keefe:
Right.

It's it's a great question. So first, if you ask the AI a super specific question,

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
then what you want is for it to return the exact answer. And if you set things up correctly and if you have, you know, kind of the right guardrails around your AI, you can set it up in such a way that when you ask that, I have X and then I configured Y and then Z happened, and it says, here you go.

Okay, and it gives you exactly the right answer and you're happy and you fix your problem and everybody, you know, walks away. But the better question is what happens when you ask a scenario like that that actually does not exist in the docs? Because

Matt Stauffer:
Mm-hmm. Right.

Sarah O'Keefe:
what it's going to do is it's going to invent an answer. Unless you, again, go to great lengths, it is actually quite difficult to get the AI to say, I don't have that information. That information

Matt Stauffer:
Yeah, yeah.

Sarah O'Keefe:
doesn't exist, that scenario doesn't exist.

So what it's gonna do is invent an answer, right?

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
Because it's designed to produce content. That's what it does. So

Matt Stauffer:
Yeah.

Sarah O'Keefe:
if it's not there, it's gonna synthesize something together. And if you're very, very lucky, you will get an answer that actually works. But there's no telling, right? If you're

Matt Stauffer:
Yeah. Yeah.

Sarah O'Keefe:
the first person to run into that problem, there's a decent chance that it's just gonna that's where we talk about hallucination, which I don't

Matt Stauffer:
Gonna make something up, yeah. Yeah.

Sarah O'Keefe:
I don't much like the term, but

That's what it is. It's basically trying to bridge those gaps and fill in the information. Now, backing up, you know, we're talking about the 300 page document. So first of all, nobody should be shipping 300 page PDFs only.

Matt Stauffer:
Right. A hundred percent. Yeah.

Sarah O'Keefe:
Right? If I mean if you want to do that or if you need to do it for regulatory purposes or various other things, that's fine. But fundamentally the answer to this is break things down into small bite-sized chunks and deliver

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
those, whether it's HTML or something else.

So that when you ask for the answer that you're looking for, it can say, here you go, here's that chunk, that topic only, not please go search this 300-page PDF. hmm. So now we back up and we talk about how content is being ingested into AI. Because the way this is broadly working right now is that

You know, on the technical content side, we have, and most of my work is around structured content, topic-based content, breaking things down into topics into bite-sized chunks, and then rendering those chunks as HTML, as portals,

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
as web pages, and also PDF, but not only, right? Like there's there's valid reasons for doing 1500 page PDFs, but providing p as an information access source there.

Awful, right?

Matt Stauffer:
Yeah. Yeah, a hundred

Sarah O'Keefe:
Okay, so we have our 1500 page PDF, and we have our nice bite-sized HTML. And in my world, all of that was derived from the original source, which is structured and organized and managed and has a pile of metadata sitting on it. Okay.

And what's happening for for AI LLM training purposes nearly always is that the AI team goes off, finds a bunch of PDFs and dumps them into the LLM

And the tech writers are sitting in the back corner going, you know, we have this really nice database full of content that's versioned and managed

Matt Stauffer:
Yeah.

Sarah O'Keefe:
and has approval dates on it. And you just grabbed what was actually not the final PDF off some

Matt Stauffer:
Yeah.

Sarah O'Keefe:
SharePoint drive. You know, you grabbed eight different versions of this PDF that were in review that have like review comments sitting on them, and you dumped them into the LLM.

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
So

Actually this conversation isn't happening because like the LLM thing is happening, but even trying to get the content creators into that conversation has been very, very difficult.

Matt Stauffer:
Right. Yeah.

Sarah O'Keefe:
why doesn't the AI perform? well, it's a magical mystery. Who could have known? So yeah, frustration. So in yeah, in that context, I sat down with a colleague of mine, Carlos Avia, who's a

professor, actually dean now at Virginia Tech. And he and I wrote a book about how to do it right. And we are

Matt Stauffer:
Okay.

Sarah O'Keefe:
done ish. that's

Matt Stauffer:
Congratulations.

Sarah O'Keefe:
coming out yeah, it's been it's been a fun summer. that's coming out in probably in the next couple of weeks. Yeah. Mm-hmm.

Matt Stauffer:
Okay. So we will make

sure that when it is out, we update the notes, the show notes that they're in there. Actually,

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
you know what? By the time this releases, it might be it might be live. So we'll make sure it's in the show notes.

Sarah O'Keefe:
It's probably gonna be out.

Matt Stauffer:
Yeah. Yeah. so let's talk a little bit about what you guys cover in the book. I imagine, and I want to make sure I'm understanding correctly, that you're primarily talking about organizations of the sort you're working with, how to make sure that the AIs that they control, whether it is a big lab AI that they're putting specific projects or specific, you know, company specific uses.

Or trained AIs in those contexts. So you're not talking about as much about how to influence how a generic chat GPT is going to answer for someone. You're talking about AIs that you control what they're being loading with, how to load it well.

Sarah O'Keefe:
So there are two fundamental I think principles that we've that we've structured the book around. And the first one really really is is Carlos's side of this and I'll do my best to explain what what he came up with or what he's discussing. But when you talk about communication and communication theory, right, there's a a speaker and a recipient, broadly, and a message.

And so if I say something and you hear it, or I write it down and you read it, either way, there is what I was trying to say, and there's what you actually heard. Right? And those two things are not necessarily the same. I mean, you know, infamously there's games of telephone where things get mangled, but also there's a cultural context, and maybe I'm speaking a language that I don't speak that well, such as English, and and

And you know, and your English comprehension is, you know, not the greatest for whatever reason, or, you know, we're in a second language, or I'm speaking very academically and you're not, or vice versa, right? But the the distance between what I intended and what you heard, there's there's differences there. And so there's some theory around this where you talk about the author's intent, but the reader also has intent. The person

consuming the message comes to it annoyed and frustrated because their software isn't working, or they're

Matt Stauffer:
Yeah.

Sarah O'Keefe:
trying to install a r a roof rack on their car and they can't figure out how, right? So

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
emotionally they're annoyed and predisposed to say this is terrible just because. or I really hate this company, I don't like this layout, you know, I'm just angry. So there's a whole bunch of stuff like that. So my intent

as the writer plays into the message and your intent as the reader plays into the message and the the message itself or the work plays into this as well. So if you think for a second about like a wedding invitation, when you look at a wedding invitation, you have certain kinds of fonts and certain kinds of colors and they say something about how formal the wedding is. I mean, in addition to the actual

Matt Stauffer:
Mm, mm-hmm.

Sarah O'Keefe:
words that say formal or semi formal or

garden casual or whatever. so the reader, the author, and the work, the actual document, all affect how information is consumed or perceived. That's

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
the intent. So I have intent, you have intent, and the work itself has intent. I mean it's not sentient obviously, but it has intent. And the fundamental premise that we wrote about is that we now have an uninvited author.

Which is the AI, which also

Matt Stauffer:
Hm.

Sarah O'Keefe:
has intent because the algorithm itself has intent. Again, it's not sentient, but it has it's a database and it has stuff in it, and that stuff has relationships. And based on the training data and everything else, there's potentially bias there. It will perform better in English than it will in a different language, that type of thing. So as you attempt to deliver information that will

enable customers, enable people to succeed at the task at hand. You have these four intents, ba four now. It used to be three, but you have

Matt Stauffer:
Yeah.

Sarah O'Keefe:
these four intents now basically in competition with each other. And sometimes in conflict. So that's fundamentally the first principle of this book is that you have a fourth intention and you have to deal

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
with that and you have to make sure that the information that you're feeding it is set up in such a way that it actually works, that it actually

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
Can deliver what you want it to deliver, and that your intent as the author survives the trip through the AI processing. So

Matt Stauffer:
Yeah.

Sarah O'Keefe:
that's kind of like part A. Part B is that we looked pretty carefully at the question of synthetic content, information produced by an AI. That's a little different from using AI as a retrieval engine, right? If I ask it a question and it says,

You know, I I the other day I said, What day was so and so born? And it told

Matt Stauffer:
Right.

Sarah O'Keefe:
me. And I said, Where did you get that? And it said, I don't have a source, this is just general knowledge. Mm, okay. you know, and then I pushed it some more and eventually admitted that it was maybe Wikipedia. Okay.

Matt Stauffer:
Yeah.

Sarah O'Keefe:
So, but the AI itself is producing content, synthetic content.

And in some cases it's producing information that doesn't exist anywhere else. So if I ask it a question and then say, please translate this into French or German or Italian, that's new content, right? That's new synthetic content. You can do things at a very basic level, if you have a podcast and then you ask the AI to produce a transcript, that is synthetic content and it's changing media, right? It's going from audio to text.

So there are transformational things that synthetic content can do. And there are and then of course it can create what amounts to net new content from its database. When you think about synthetic content and what it means to produce synthetic content, you have to think about the risk of how that content is being produced and the gap. So circling back to your question, you know, I asked a really specific question and it gave me an answer. Okay.

But did it actually have the information to produce that answer, or is it inferring or inventing that answer? when you synthesize or summarize or translate, translate languages, translate media, then it is generally going to perform better than when you ask it to extrapolate. And so for me.

As a technical content producer, that distinction really, really, really matters.

Matt Stauffer:
Yeah.

Okay, so you said synthesize, translate, and what was the third thing you said it did d did better? Transform, okay.

Sarah O'Keefe:
like transform, like i the

audio to text or text to audio, that type of thing, where there's existing content and I'm just

Matt Stauffer:
Yes. Yes.

Sarah O'Keefe:
changing the medium that it lives in.

Matt Stauffer:
And I don't know if this is the reason why, but I have certainly like because we one of the things we do for our clients is often build AI-based systems. And if the

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
system can have a an input that it can check its output against, then it's much more likely to be able to say, Yep, I did that right, versus if the input it's checking its output against is inference based on hundreds of thousands of training documents, it there's no there's no check. And what you even mentioned before, you know, we were talking before we started recording about the fact that programming is having an easier time because it is like there's a right or a wrong.

answer in a way that there is not with with with writing, with written you know, with content. yeah.

Sarah O'Keefe:
Programming is a language, right?

But it is a it's an artificial language and it is limited in the sense that it

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
has a limited number of words that you can put together in a somewhat limited I mean, limited compared to natural language. And so the problem set is

Matt Stauffer:
Yeah. Yes, yeah. And it it either compiles or doesn't.

Sarah O'Keefe:
yeah, it either compiles or it doesn't. and the problem set is relatively limited compared to any actual language on earth.

Matt Stauffer:
Yeah. Okay. So you you kind of give us those two foundational ways or like conceits that are going to be talked about in the book. Who is the book for? Who should be considering reading this? And who you who are you guys targeting and what problems are they having that you're hoping to address with this book?

Sarah O'Keefe:
So I think there's there's a whole bunch of audiences, but broadly it's people working in this space who are dealing

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
with these problems. but mostly it's their managers, right? And their directors. Because

Matt Stauffer:
Okay. huh.

Sarah O'Keefe:
what we're trying to do is explain this problem in such a way that the executive management who are making decisions around AI and who are investing in

Matt Stauffer:
Yeah.

Sarah O'Keefe:
AI and who are saying, Hey, tech writers

Why can't we replace all of you with actually that's the good scenario. The good scenario is that they

Matt Stauffer:
Yeah, right.

Sarah O'Keefe:
say to the tech writers, Why can't we replace you with AI? The bad scenario is that they've already laid them off and have replaced them with AI and they can't figure out why it's not working.

Matt Stauffer:
Yeah. So I haven't heard you say anything either here or publicly that makes me think that your proposal is throw AI entirely out of the system. So I imagine that your proposal has some form of collaboration. what is the ideal outcome? Is it still using AI but better trained I AI in partnership with actual human beings versus firing other humans? Like in a system like the the clients that you work with, what do you think the ideal situation for them to set up would be?

Sarah O'Keefe:
So again, we the it gets weird because we're we're talking about AI both as a how do we use this as a productivity tool and then also how do we deal with it when we have to deliver, right?

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
So going back to the the productivity tool, the point of we need to create a bunch of content, like net new content, there there are some fairly obvious things that you can do, right?

We can use AI to generate summaries of a particular topic that are useful. We can use it to generate alt text. We can use

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
it to support writing initial outlines, or we can use it to say, go through this JIRA database and give me all the updates since you know a month ago, or all the updates from this sprint and produce a draft of the docs or of the release notes or whatever we need to do in here.

But the first thing you really have to do when you start thinking about AI as a productivity tool, and I think there's a lot of value there, a lot of things that you can potentially do with it, is you have to ask the question of what is the risk of deploying AI here? Right. Now, you know, I have a 1500 page document and I need it to be 500. I'm sorry, 1500 word document, and I need it to be 500 words. That's a great use case for AI. Like, you know, synthesize this down to 500 words. Cool.

now read it and make sure it didn't omit anything really, really critical, but yeah, start there. Or find it's really good at finding patterns.

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
Which the inverse of that is it's really good at finding outliers. So you can use it for spelling support and grammar support and just like, you know, a spell checker, but you know, high powered. You can use it to generate

or actually downsample things, summarize down, synthesize down.

Matt Stauffer:
Mm.

Sarah O'Keefe:
It's when you extrapolate up that things start to get weird. So you want to be careful when you start doing that. So fundamentally, as a content creator, there are a whole bunch of things that I can do with it that'll help me be more efficient with my time instead of you know going through things by hand, which as humans we're not that great at.

Matt Stauffer:
Yeah.

Sarah O'Keefe:
So then we kind of move downstream from there and we think about all the other things that we can do with it. And you have to start thinking about what is the risk of of of using AI in this context. Now, I've talked a lot about regulatory, right? Or you get a procedure wrong and somebody gets injured, which is, you know, like a serious problem. but there's also cultural issues here. And the example I always, always use is video games.

There's been huge pushback on use of AI in video games.

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
And the reason is that, you know, we play video games for fun. They're supposed to be fun. And

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
AI in general is not fun. It's it's sort of flat and just ugh. It's yeah, it's fundamentally

Matt Stauffer:
Yes, it's the average, not the unique. Yeah.

Sarah O'Keefe:
the average. And so if your business model depends on producing content,

that is interesting and unique and fun and entertaining, there's huge risk around using AI in that in that scenario. Now, I personally think that if, you know, you're using AI to like fill in backgrounds or something, that's probably not the worst thing in the world. But but many, many people who buy video games have decided that that is not accepta they won't buy anything that has any AI in it because they see it as a slippery slope and

Matt Stauffer:
Right.

Sarah O'Keefe:
They're probably right.

Matt Stauffer:
Yeah.

Sarah O'Keefe:
So that's kind of the back end, right? You know, deploy it, use it for productivity, think carefully about balancing the risk of it getting something wrong versus the

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
benefit of being more productive. one of the things I really push on is like alt text. We should be doing alt text for all of our images and we don't. And AI will get us to like 80 or 90 percent, which is pretty solid.

So we should go down that road. if you look carefully at the translation industry, localization, AI has pretty much taken over there for most use cases. And you can kind of see where it's going where machine translation broadly, which has been around for thirty years, is has reached a state where for a lot of content it's good enough. It's not perfect, but it's good enough.

And so we use it because it's good enough and we reserve human, you know, linguistic translation for literary fiction, for maybe high-stakes legal work,

Matt Stauffer:
Right.

Sarah O'Keefe:
for you know, super technical biochemistry analysis, you know, things where you know you need really specialized expertise. And that

That makes a lot of sense, right? And so you then kind of apply that model to content and you you end up with something very similar. Now, downstream with the AI

Ultimately, at the end of the day, this is where we bring in the discussion of deterministic scripting versus probabilistic AI. So scripting, whether it's Python or JavaScript or any other traditional programming that you do, is deterministic, which means that when you run it, it always does the same thing. Now, it may or may not work, right? I mean, if you if you're me, there's a bug in there, and you did something wrong.

But deterministic is like one plus one is always two. Or actually, if you screwed up your your deterministic script, one plus one is always three. It's just that it's always three, and you can go

Matt Stauffer:
Right. But it's always three. Yeah. Yeah.

Sarah O'Keefe:
in there and you can look at the code and you can say, crap, and you can fix it. probabilistic, which is AI, machine learning, and all the you know, broadly that whole bucket of things, one plus one is sometimes a different number. Usually it's two, but maybe not.

And so at the end of the day, the question is, how deterministic does your content need to be? How

Matt Stauffer:
Hm.

Sarah O'Keefe:
predictable do you need it to be? And what are the use cases where predictability is critical? And what are the use cases where having it be more probabilistic and giving you the average of what's in the database is acceptable? That's the determination that you have to make.

And so there's a lot of, you know, really fascinating questions about how you deal with this and what you do and what are your standards. Now, you know, I think the brutal reality is that for product and technical content, generally the standard has not been that high. Like a lot of this content's not very good.

Matt Stauffer:
Right.

Sarah O'Keefe:
And so just getting it to average feels like an improvement potentially. So

Matt Stauffer:
Yeah. Hmm.

Sarah O'Keefe:
here we are.

Matt Stauffer:
Yeah. so I'm gonna that was really, really, really helpful context. I'm gonna re ask one a piece of my question because I'm still curious. Do you feel like your clients and the the authors or the readers of this book have any control over how I don't know what the right phrase is, but like Chat GPT, Claude, the the the public facing A AI tooling, do they have any control over that? Or is their control mainly happening in the context of

what the company owned, the nonprofit owned, you know, AI. And you mentioned AI in two spaces, right? AI is a productivity tool and AI is a

Sarah O'Keefe:
Mm-hmm, mm-hmm. Yep.

Matt Stauffer:
distribution medium. But from a distribution medium, do you even worry about trying to like SEOs do, trying to tweak how the how chat chat P T G P T is gonna do something? You're like, that's outside of our control. We're really talking about our tools.

Sarah O'Keefe:
So first it's it's I mean, it's obviously true that if you build your own internal private processor, private chat bot, that you can control that. You can tell

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
it what information it has access to, you can exclude things, you can tell it not to ask entertain questions like what is the weather tomorrow? And you can glue on a whole bunch of resources like a product specification database or you know other kinds of things like that.

So that's yeah, part A is clearly if it's private and it's on the inside, you have way more control over what's gonna happen and you can test in a semi-controlled situation. on the public-facing chatbot side, it is surprisingly not that difficult to

To do the kinds of things that you that we used to do with SEO for a public chatbot. Now,

Matt Stauffer:
Mm-hmm. Okay.

Sarah O'Keefe:
at an extreme, if you look up some of the stuff that's out there that will poison the AI, you know, the research says that it takes as few as 50 or 100 documents to poison a public-facing AI with bad information.

Matt Stauffer:
Yeah. Mm-hmm.

Sarah O'Keefe:
So the implication is I should be able to get good information in there, right?

Matt Stauffer:
Right.

Sarah O'Keefe:
We'll turn that to good. but it's a lot of work and it's very hard to control. but the the big picture answer to this is what it's always been and what it was for SEO. You deliver good information that ranks highly because it's accurate. You make sure that it is chunked up properly. Again, not your 300 page, if only it was only 300 pages document,

Matt Stauffer:
Right. Yeah.

Sarah O'Keefe:
but rather, you know, bite-sized topics.

with a bunch of metadata attached to them. And then there are some ways of telling the chatbot how to deal with that information. So MCP model context protocol allows us to say, this is what I'm delivering to you or this is what you're reading here. And then the implication is that we can potentially get the public facing chatbots to do better. Maybe.

Matt Stauffer:
Got it. Okay.

So we're not talking just Jason LD or the more

traditional forms of structured data. We're also talking about kind of like new AI targeted things where you're trying to hopefully shepherd them

to to

understand more than we could have traditionally taught in a search engine.

Sarah O'Keefe:
So the search engine or sorry, the AI engines like

Tagged information that is semantically useful. And as it happens, that is structured content, which typically is encoded in something like XML with validation. That doesn't

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
mean you feed XML to the AI necessarily because it's not the most efficient

Matt Stauffer:
Yeah.

Sarah O'Keefe:
way of doing it, but that's probably fundamentally the starting point on the authoring backend for the writers, for the content creators.

Then you think about, okay, well, I'm authoring it this way, but how do I deliver it? And what's the best way to optimize it for delivery to the AI, which might be different from the best way to optimize it for delivery to the human. But

Matt Stauffer:
Yeah.

Sarah O'Keefe:
we've been doing multi channel for, you know, twenty five years. It's really not, maybe more. It's not rocket science to say I need two outputs here. So that should not

Matt Stauffer:
Yeah, yeah, yeah. Yeah.

Sarah O'Keefe:
be a problem.

The problem comes when you start ingesting fifteen hundred page page PDFs because that's what you found when you went scraping the internal SharePoint database.

Matt Stauffer:
Yeah, company internal. Yeah,

a hundred percent. Okay. and so obviously in a scenario where people are saying, Hey, look, we fired the whole content team and we just uploaded a whole bunch of PDFs and nothing's working, the first thing to do is obviously to hire Sarah O'Keefe, right? Like that's that's the easy one. Hire

Sarah O'Keefe:
I mean clearly.

Matt Stauffer:
scriptorium, please. are there people who are cause you mentioned that the target market is not even necessarily people who are doing this work, but it's their bosses. So there is a

message that they are receiving, which is that you

can just and that's not working. And we've all gotten that, right? Like you can just fire all your programmers and

replace them with one guy with clawed code and you know we're seeing what happens. Or, you know, all these big companies that fire everybody and then they realize the downside and have to hire them back. So I know that quite a bit of what you're you're fighting here is the you can just of get rid of the humans, feed everything to AI. And a lot of what you're I'm hearing you say is it requires a understanding how to use the AI well

But also be understanding its limitations, which goes back to requiring humans in the first place. If there were one output of the book, if your book has all the success you want it to have in the world, what would you describe that output as being? Or that outcome, I guess.

Sarah O'Keefe:
AI is a tool. And tools should be used in the context that they are good at doing. I mean, this is this is the the old if everything you know if you have a hammer, everything looks like a nail. I mean, it's

Matt Stauffer:
Exactly. I was gonna say the same thing. Yeah.

Sarah O'Keefe:
just so dumb, but of course it's true. fundamentally, in order to get the AI to produce useful content, you have to feed it useful content.

And so if the input is garbage, the output I mean, this is just all cliches, right? But people decided that what they could do was take a bunch of unmanaged, unstructured, unvetted, unupdated, unaccurate information and dump it into an AI and it would magically come out pristine on the other end. Now, here's the key thing. It comes out looking pristine.

Matt Stauffer:
Yes.

Sarah O'Keefe:
It it's very lovely. Looks great.

And the problem is, this this is the fundamental issue that we have with AI. I think cognitively, when you see something that looks pretty, you assume it's accurate. Or when it reads fluently, it's grammatically correct. Yeah.

Matt Stauffer:
Yes. The grammar is good, the writing is structured well. Yes, a hundred percent. Yeah.

Sarah O'Keefe:
All of those things used to be proxies for somebody put some effort into this content, right?

Matt Stauffer:
Yes. Yes.

Sarah O'Keefe:
And now the AI has I'm gonna say fixed.

those things, right? It's always going

Matt Stauffer:
Mm-hmm. Yeah.

Sarah O'Keefe:
to produce grammatically correct unless you tell it officially not to, but by default, it's going to produce grammatically correct content that is fluent, that is mechanically correct. And so we have we've broken the connection between this looks good and this is good.

Matt Stauffer:
That's so good. And we've had the same problem in coding, right? If if

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
it's syntactically correct, if you know, these are good signs. I'd love that you said a proxy. It's a proxy for somebody who thought about this. They're an expert. They understand how to do this well. I just did a hiring round. I read 680 something people's applications. It is my first time doing a round of hiring since AI. And the markers that I used to look for are different because they're feeding it at AI and a lot of those markers are not there anymore. And I actually found myself when I would see a typo being like, huh, what a relief.

You know what I mean? So yeah, you're 100% right there. Yeah. Which is apparently that's a thing that people are doing too. And the good news is

Sarah O'Keefe:
'Cause they're probably real. Except probably they told the AI to carefully put in one typo. Mm-hmm.

Matt Stauffer:
I could tell usually. We we put some and you don't know this, but we put some questions in there. One of them being, what's your favorite movie? And you know, I can tell when an AI is answering that question. but yeah, 100% the proxy of of quality is is very different. Okay.

So I know that we need to get over to our speed round in just a second. Before we do that, is there anything else you want to share either about any of these topics, anything you're thinking about, or about your book or about Scriptorium that we haven't got a chance to cover today?

Sarah O'Keefe:
No, I think that you know, I'm actually, well, Carlos is optimistic and I'm somewhere between not so optimistic and just resigned to the inevitable, because what I'm looking at is social media,

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
which was our kind of our last big, hey, we're gonna have this cool thing happen. And we didn't handle that very well.

Matt Stauffer:
Mm.

Sarah O'Keefe:
So I'm looking at AI and thinking about how are we going to manage this and how are we going to handle it, and it's it's quite concerning.

the one thing I will say for for those of you that are looking at these issues and, you know, thinking about going into it wholesale, I want to touch on the global issues rather than, you know, we've been talking about the very kind of micro how do I feed my AI? But let's step back for a second. there are a number of organiz or well, the EU specifically is regulating AI.

Matt Stauffer:
Mm-hmm.

Mm-hmm.

Sarah O'Keefe:
They passed

the EU AI Act. And if you deliver into the European Union, then you are subject to the EU AI Act. It's very similar to GDPR, the data privacy stuff, in terms of its reach. So a workflow that in the US is fundamentally unregulated and based on a bunch of AI stuff is going to run into problems if you then take that and just

transfer or deliver wholesale into the European Union. I don't hear enough people talking about that yet because they

Matt Stauffer:
Yeah.

Sarah O'Keefe:
seem to think that it doesn't matter because we did it in the US or in North America. So that's that's problem A. And problem B is digital digital sovereignty, which I have trouble saying and even more trouble spelling. This is the question of who owns your systems. Now, you know, we talked about

Big AI. So clearly a public facing chatbot you don't own and you don't have sovereignty over because you've got all these issues. But if you bring it in-house, so you have an enterprise license of one of these things and you bring it in-house, well, then you kind of own it, maybe, right? But there's these really interesting questions about what level of control do you need over your AI systems? And as a big enterprise, are you going to

license one of the public systems? Are you going to build your own on an open source or open weight, which are not the same thing system? and then you have to worry about, well, what if that system gets banned in my jurisdiction? So you're in the US, you decide to use a Chinese model, it gets banned, or vice versa, right? So if you're a global organization, you have to think pretty carefully

about what your AI infrastructure looks like because you could get yourself in all sorts of trouble, both on the output side and on the input side if you don't understand the regulatory nightmares that we're dealing with. And I should I should close this by saying I am not a lawyer and you should definitely talk to your IP lawyers about this.

Matt Stauffer:
Yeah.

Yeah. Well, I appreciate you saying that. I mean, we build systems for our our our clients and still I don't think this has come up with any of them yet. So I'm like, I should make sure we dig into that with my lawyer. So

I appreciate that. Okay. Are you ready for your speed round?

Sarah O'Keefe:
I don't think so, but go ahead.

Matt Stauffer:
As ready as can be.

All right. First question is what is your AI tool of choice?

Sarah O'Keefe:
honestly, whichever one is lying in front of me, I don't care, they're all terrible.

Matt Stauffer:
Okay. Wow. That you are the first person who's answered that way. I'm not supposed to comment on your answers here as a speed round, but wow, I'm fascinated. All right. Second question. Have you ever cussed out an AI?

Sarah O'Keefe:
no, I I point and laugh a lot and take screenshots and share them. you know, when you ask the AI a question, this is the experiment I would ask people to do. Ask the AI a question about something that you are an expert in. maybe it's knitting, maybe it's video games, I don't know, something, right? Ask it a question. Nearly always the first answer you get is pretty good, and then ask it for more details, and as you go down that rabbit hole it'll get worse and inaccurate.

Matt Stauffer:
Uh-huh.

Sarah O'Keefe:
Okay?

And then ask it a question about something you don't know anything about and do the same thing and notice that as you go down the rabbit hole you think it looks okay. And then

Matt Stauffer:
Yeah. Yeah.

Sarah O'Keefe:
think about that carefully.

Matt Stauffer:
That's so good. I'm gonna make my kids do that. Thankfully my kids don't do a lot of AI, but still I'm like I need the baseline understanding, you know, about that. All right. What is one pro usage of AI more people should be using?

Sarah O'Keefe:
I think that you can use it to say find the deficiencies in this document. find, you know, I was talking about contracts and getting lawyers to actually review things, but as a first step, you know, tell the AI, you know, read this I do a lot of this, read this SOW and tell me where the problems are for me. Now, do

Matt Stauffer:
Hm. Okay.

Sarah O'Keefe:
I then also send it to my attorney? Yes, I do. But

Matt Stauffer:
Yes, yeah.

Sarah O'Keefe:
fundamentally it will find the really obvious bad stuff, and that's a good that's a good start. I also use it a lot for various kinds of translations, but I don't think that's necessarily a a pro use. That's just a really obvious usage. If you speak multiple languages, ask it the same question in multiple languages and notice that or ask it to translate, or not translate, but answer in that language, and notice that the answers are different.

Matt Stauffer:
Fascinating. I've never All right. is AI going to ruin the world?

Sarah O'Keefe:
so it is September twenty twenty six. So far no. Ask again next week or maybe later this afternoon. I think that the LLMs no. Lar large language models are just large language models. They are not that big a deal, and it's not that difficult to manage a an autocomplete engine.

the people that are operating them that are using them in irresponsible ways are a big problem. should you glue a you know, an inference engine onto like a bioweapons lab? No, you should not. Are they going to? Almost certainly yes. Am I happy about it? Mm no.

Matt Stauffer:
Yep. Yep. All right. And your last question is what is the most controversial take you have about AI?

Sarah O'Keefe:
I think that people really wanted an easy button and they turned off their brains when AI came out. There's like, this is great, this will save me a pile of money. And now because of the way publicly traded companies work, it is a it's not quite a Ponzi scheme because there's definitely some value in there, but nobody wants to be the first lemming to step back from the cliff.

Matt Stauffer:
Really shot myself in the foot not being supposed to answer when you guys give these because it's so it's some of the best stuff. First, lemming the step back from the cliff.

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
Wow. Okay. well, I'm not allowed to. Okay, there you go.

Sarah O'Keefe:
It'll all be fine. I swear, maybe.

Matt Stauffer:
So if people

have loved everything

you've said and they

want to buy your book and they want to

Sarah O'Keefe:
Mm-hmm.

Matt Stauffer:
follow everything you say, what are the best ways for them to do that? And of course, this will go in the show notes as

Sarah O'Keefe:
Yeah, so the title of the book, well, first of all, you can find me at scriptorium.com and there's contact information and I'm on LinkedIn and I have a not common enough name that you should be able to find me there without too much trouble. the book title is The Uninvited Author, AI, Intent, and the Future of Content. And we've tried to cover off those those three things in

Matt Stauffer:
Mm-hmm.

Sarah O'Keefe:
the context of our uninvited author. I'll also say one thing.

You know, I've written several books over the years. This is the first sort of post AI book. And I've also found this in, you know, articles that I'm writing. The the interesting thing about AI is that it has made the content that I write less professional because I get to say things now that previously would have gotten edited out for being too over the top. And now we leave them in because if we take them out, everybody just assumes it was AI.

Matt Stauffer:
Yeah. Wow.

Sarah O'Keefe:
Right. So we have to be

more interesting and fundamentally less professional as writers in order to prove that we're real human beings. And and that made it fun, actually, because it

Matt Stauffer:
Wow. Like you said, the the the fake typos. I like that.

Sarah O'Keefe:
was it's fairly demented, really, at the end of the day. So

Matt Stauffer:
Yeah. Yeah.

Well, all of these will be links in the show notes. Your LinkedIn, Scriptorium, the Uninvited Author page. we're pretty sure that it's gonna be live by the time this comes out. So go take a look at all these things. And I'd find you on LinkedIn, so I I know you're prolific there. That's your main social network, right? Okay, so

Sarah O'Keefe:
yeah. Mm-hmm. Yep.

Matt Stauffer:
Sarah, I am so grateful you spent some time to teach us today. Thank you so much for helping people in your section of the world

obviously understand how to interact with AI, but also helping the rest of us learn from you and everything you've kind of learned and put together over the years.

Sarah O'Keefe:
Yeah, well thank you. This is super fun and we'll talk again.

Matt Stauffer:
Sounds good. And the rest of you, we will see you all next time.

Sarah O'Keefe:
Thanks everybody.