cirrus precision x systems iowa state university cybersecurity speaker introduction

My name is Dawn Ely, and I'm the director of Enterprise Services with Cirrus. So Enterprise Services encompasses executive strategy, workforce, operational excellence, and supply chain are all in the area that I support. And just super excited to be at the AI conference. I'm 3 for three so far. Started at Cirrus in September of 2024, and luckily enough, started just in time to go to the first summit and to be here again today. And now I'm going to let Dave Mahusky introduce our speaker. Yeah, so I'm co-founder of Precision X Systems, so we're the premier sponsor for the AI Summit. Glad to be here, excited to be here. This is our first one, and I think we're going to keep coming back. This is just, it's a great time, great, great conversations. And so these sessions are really kind of eye-opening too as you listen to other speakers, and it's really informative. So you guys are in for a treat. And with this guy too. So our next speaker is Professor Doug Johnson. So he's the director of the Iowa State University Center for Cybersecurity Innovation and Outreach for, and also a professor of electrical and computer engineering.

cybersecurity experience iowa cyber resilience initiative university programs public sector professor jacobson

Professor Jacobson has been working in cybersecurity since the early 1900s. Nineteen 90s. OK. He got that right away. He got that right away. Yeah. So I gave a I gave a talk on Monday to Rotary and they said the exact same thing. So AI somewhere AI must have that must have been older than AI since 1900s. Why? OK. I thought it was like a funny thing. OK. Sorry, Doug, we love you, Doug. He has built. He has. He has helped build one of the nation's longest running university programs in the field. And he also directs the Iowa's Cyber Resilience Initiative, which provides cybersecurity services and training to more than 1,400 public sector organizations across the state of Iowa.

ai risk governance data controls vendor claims ai decisions

In this session, AI security is not a brand, governance, risk, and the reality behind safe AI. He will offer a practical perspective on how to evaluate AI risk beyond vendor claims, and you guys will gain a clearer understanding of governance, data controls, and the key questions needed to make informed, defensible AI decisions. So on behalf of Cirrus and the PrecisionX System, welcome, please join me in welcoming Professor Doug Jacobson. All right, cool. So yes, and this will kind of, for those of you who've been in the first two sessions, this will kind of piggyback off of the first two, and Aaron will keep me, it's just like I kept Tim honest. So what I want to focus on today is really talking much more about the processes and systems that you should be putting in place.

copilot shadow ai tool selection organizational mandate iowa state

as you're looking at deploying AI. And so I don't know how many people have heard this. was anybody here from Iowa leadership in Iowa State? This was my world up until a little while ago that the only thing you could use is copilot because that's the only thing that's safe. Sometimes I fight my administration, sometimes I don't. I just ignored them on this one. Shadow AI, right, Aaron? So, in this, I noticed the case of Iowa State, it wasn't really a security argument. It was a control argument. It was a money argument. It was several other arguments. It had nothing to do with security. And, you know, it really, what we ended up at Iowa State is that everybody was out buying their own licenses for their own AI. So Taking this sort of a stance to start off with, again, starting with the tool and saying good tool, bad tool, really that's how it shuts down the conversation.

false confidence innovation technology limitations cybersecurity tool restriction

and as was pointed out earlier, you get a lot of innovation from the people who are allowed to do cool things. And also, I think one of the biggest things, it creates a false sense of confidence. When you say this is the only thing, focus on that tool, say this is the only good tool, this is the only thing that's safe. I argue, I argue over and over again that in cybersecurity technology does not solve the problem all by itself. Actually, it's in my anymore, it's a small part of the solution to the problem. So we need to replace and start thinking about other things before we start thinking about tool selection. This is not a vendor comparison. This is not anti-Microsoft, only because that's such an easy target to do that it's not fun to be, well, it's fun, but it's really about risk decisions in government and governance.

security vendor claims brand governance data protection

And so If we think about, security is not determined by brand. So if we think about security in general, and this is true not just in AI, this is true across security in general. So much of what we see is driven by vendors out there saying, I have the best this, whatever that is, and it will solve this problem. And a lot of people are driven by that. They're driven by the fact that the vendors are out there doing that. But security really is a lot more than that. Security is really understanding what you have, what's unique about you, coming up with that process and those methods to handle what you have. And in the case of AI, we're worried about data. And in the case of general security, we're also worried about data. So a lot of it is about understanding our data, putting governance around that data.

data governance identity and access access control agreements architecture

And the same governance we put around general data, right? You think about your business and not everybody gets to see the recipe for Coca-Cola, right? So the same sort of governance we put around that, the same sort of identity and access, who has access to it, the same type of controls that we talk about with our data, Those are the same sorts of things we need to think about with AI. And we need to also think about an architecture, that kind of comes a little later. And one of the things that we need to talk about are agreements. Aaron mentioned some things about agreements. There was a question in one of the earlier sessions about, oh, we got something new and now all of a sudden it's doing AI things and we didn't know it was doing AI things. So again, we need to focus on this part of the process. But this is how it usually plays out. We find a tool, everybody finds a tool, the shadow AI people find tools, and there's a new tool a week.

deployment pattern tool selection copilot chatgpt shadow ai

There's a new model a week. And so we start with the tools. And I said, that's where Iowa State started. We all use Copilot because we're a Microsoft shop. All the other tools, no, can't use them, shouldn't use them. without even talking about anything about how, why we should be doing that. We then kind of focus on implementation. We pick a tool, we implement it. Again, Iowa State, it was undefined. They just kind of, copilot just showed up one day. Didn't tell anybody it was there, other than a little while later, they said, this is what you should use. Then there was a big question, do we have it? Most people didn't even know they had it. Of course, this was well after like, 30% of the people I knew on campus were already had their own licenses for ChatGPT and other things. So we were already way down the ship in running our own AI.

sensitive data data classification policies agreements identity

Then they sat there and said, okay, don't put sensitive data on it. Iowa State's never really told us rank and file people what sensitive data is, other than student records, only because the federal government said those student records shouldn't be shared with people. Identity and access, yeah. We, well, we have access to most everything except for the stuff we really need. So, but that's a big issue. And then, again, I keep using my place as an example. I think we have policies and agreements. We're never privy to any of those, unless you break one of them. So, and they now, they've come out later.

data access user permissions ai access data exposure governance

We now have some policies and governance around AI after the ship's already sailed. So this was our deployment. This is probably the better deployment. And really, again, involved, you know, flipping this around. We just talk about data, we talk about it, you know, having AI, having access to your data. And AI has access to all data that your people who use AI have access to. So if a person has access to the data, AI will have access to that data, unless you have some processes and methods in place. So from a governance standpoint, you need to understand your organization. You know, at Iowa State, we obviously have legal and compliance issues. We have FERPA data. Iowa State as an entity. we have a medical center, little one.

ferpa compliance legal obligations research ethics medical data

Those went to Iowa State, remember being called student death. And, we take money. So we have a arm where we have to deal with money, all sorts of agreements, et cetera. So those things need to drive how how AI is used. We also have another interesting issue that they've kind of come out a little late into, is we're a research institution. We're just now starting to train our faculty and researchers about the ethical use of AI in doing research. That's come out a little on the behind side, but at least we now sort of have that in place. We don't do too bad of a job on identity and access because they do lock down what we can't have access to, except for the things that we ourselves create.

data classification data retention ferpa data policy awareness faculty data

So that's the interesting dilemma we have in our organization. You know, I don't have access to student records except for the students in my class. So every faculty member does have FERPA data. We generate a lot of our own data. Data control and classification. Iowa State does have a classification policy, how to classify data. I don't remember the last time I looked at it, nor do I think half the faculty know it exists. Data retention, it's forever at Iowa State, at least as far as the data faculty maintain. But then we now start, then you start to should be moving into implementation, integration, guardrails, monitoring, et cetera, and finally down into the tools. So that's really the kind of discussion that we should have had. Now, many of you in this room have probably already deployed and been through that, so you're probably more in the first picture.

deployment sequence governance implementation guardrails monitoring

But all these pieces are very important for how to have a secure AI implementation. So I really haven't focused on Aaron did a good job of talking about some of the tools that will help with some of these some of these pieces. So we all kind of know how AI works. So this morning I'm going to say that it may be soon it'll start to think after hearing this morning's talk. Be nice to your AI so you're not working in a lithium mines. And so basically AI is a predictive system and it generates response. It really doesn't think. So really, again, it's the risk comes in, depends on how you use it, where you use it, what kind of models you use. And again, back to this, how is the data governed?

predictive system ai models risk factors data governance ai capabilities

people talking about closed models and so on. So when we think about AI and AI's use, and again, this morning we had the talk about, fully autonomous agents running everything. I do want to ask that company that had AI overwrite their database, what was it, two weeks ago or so, how that, how well that worked for him. So, and we had some good questions at keynote about this whole, humans involvement in AI. And I kind of like to look at it from two sides, the human assisting the AI and a human assisted AI and AI assisted human. And so again, this kind of plays into that security piece of being able to kind of always know what AI is doing or try to know what AI is doing.

human in loop autonomous ai database incident ai oversight checks and balances

And that really helps us understand, especially when we start turning it loose to do things. So security is not just about AI. feeding data into the model and having that leak, whether it leak from your organization to outside or leak within your organization. But it is also, as I mentioned, the company that had AI overwrite their database because they turned it loose. So security is also being careful of how you deploy AI and what you let it do all by itself. So always wanting to be able to do that sort of checks and checks and balances with AI. So again, there's a questions of, then this is back, this is to what I just mentioned, this idea of, of data moving around when you use your AI models.

data movement platform configuration agreements data flow tool variation

And it depends on the platform, the configuration, the agreements, et cetera, of how does that data move. And not all tools remain the same, not all agreements look the same, and not all users treat it the same. It was really interesting, it was about a month ago, One of the chat, but one of the systems out there that you could build your own agent, there's a big announcement. Somebody released an agent called Einstein. Einstein was an agent that you gave it your login credentials to your, as a student, you gave it login credentials to Canvas. Canvas is the learning management system. It would log into Canvas. It would grab all your assignments, you grab all the lecture notes, all the videos, and then they would automatically do the assignments for you.

einstein agent canvas credential sharing automated cheating data exposure

It would do the exams for you, it would write the papers for you, and it would do everything for you. That was released. It lived for about two days, three days, and then it disappeared. My guess is that Canvas brought more lawyers than exist in the United States. But there was a case where an outside entity created a agent in an AI system that an internal user could voluntarily give access to all its data. That's a scary thing. And, you know, when this came out, our IT people are going, There's no really way to stop this because the students are giving it the login credentials. This thing looks like a student logging in. There is no technology that's going to fix this. So, but it went, like I said, it went away.

educational tool student use ai in education faculty attitudes calculator analogy

Now the students have to download the assignments by hand and the lectures by hand and feed it in the model by themselves. So, a little more work. Same outcome. Oh, it doesn't do it for them automatically. Yeah, probably. Yeah, we won't even get into the whole, students and yeah, I will for a second. The bottom line is, at least how I treat it, AI is a tool, just like calculators were a tool, and I need to let the students know how I expect it to be used. And I lean into it heavy. I use it, and I tell them where I use it, and I tell them where they can and can't. But I have faculty members who want to hang every student that even looks at the AI. And you can't not look at the AI anymore, right?

cybersecurity risk data leakage identity compromise risk categories security threats

Stupid Google, right? Your search engines are AI. The entire world is. So anyway, but off that soapbox of my colleagues. So I want to kind of go through four types of risks that we want to look at quickly. And so we have the cybersecurity risk. been in cybersecurity since the 1900s. So long time, long time. Hey, that Babbage engine, that was a hard thing to secure, let me tell you. All right, who's the old people in the room? So, again, we've talked about several of these things. data leakage is an issue. Even with internal models, we have issues of identity compromise and data leakage, even across your own internal models.

internal models data leakage medical data misconfiguration api

I've talked to people in the medical field that are really concerned about those internal models. And by making certain queries, what information can you accidentally learn that you shouldn't be learning? Obviously, misconfigurations. We have the various forms of APIs and MPCs and all these cool acronyms of how you can interconnect to your AI. I'm very worried about that sort of misuse. Actually, one of the things Iowa State did, which actually upset a lot of people, the Canvas system I mentioned, faculty used to have API access to Canvas. And we'd write little apps to do grades and things like that. Well, they decided that was a security risk and they took it away from us. So now we have to do all this up by hand. Issues of bias.

bias validation hallucination training data prompt bias

And then, you know, the issue we always run across is validation. When is it hallucinating? When is it not hallucinating? So again, always, I always hate slides, always talk ahead of the slides. So again, you know, it's a little more beyond feeding the model as far as data leakage. As I already mentioned, we have internal exposure. So it gets a little complicated if you're worried not only about data leaking outside your organization, but data leaking across your organization. Bias, this is part of a kind of more of an education piece to your users. But again, it has obviously has bias based on, you know, in the data and the prompts itself. I know, you know, Aaron mentioned something about tools out there that will help sanitize the prompts.

prompt injection hidden prompts document poisoning model poisoning adversarial techniques

Now we as faculty, many of us have talked about, I know some faculty have tried this, when you hand out an assignment in a PDF to a student or some other document, you will embed hidden prompts in the document. So when the students feed it into the AI, it will feed garbage back out or feed a keyword back out. Like, make sure the word banana shows up somewhere in the output. So now there's none of my students are in here, I think. So, sorry. Well, these are being recorded, isn't it? Oh, well. Oh, oh, well. But there have been discussion, you know, there have been People have talked about this as another way for an adversary. That's by sending you documents and you feed those documents into your model and now you've poisoned your model. So this is a real, you know, even when I talk about my students, this is a real threat. Of course, AI doesn't always have the most recent answer.

accuracy hallucination outdated information result validation ai limitations

This is the things we always try to get across to our students, that AI is not always right. And obviously, to review the results. this kind of ties into that. Most of us realize that AI hallucinates, gets bad data. It's fun to argue with. My wife won't be listening to this either. But the other day, my wife's just now getting into AI. She's, you know, the last several months. And she came up and she says, I'm going to quit arguing with AI. So I spent an hour arguing with AI. Okay. We shouldn't spend an hour arguing. So but but again, you know, it does it does hallucinate. And you can, you know, sometimes if you beat on it hard enough, you can win your argument, even though it's wrong. Kind of a side side note.

personification ai naming athena pixel communication style

I'm going to talk about that. How many find it kind of creepy slash reinforcing the way it talks to us? So affirming, which, you know, I think hopefully somebody today is talking about that aspect of AI, because that's a that's a whole different AI, a whole different aspect of AI is it's how many have named their AI? You got what? OK. OK, so mine mine said it was named OpenAI said she wanted to be called Athena and Chad GDP wanted to be called Pixel. So I got it. So yes, I've named both of mine. They came up with those on there. Yes, Now, I'll give OpenAI a little bit of credit because it came and said, I'm not a person. so giving me a name, it doesn't really do much, but if you really want, I can give you some suggestions.

perceived risk legislation government regulation public fear over-restriction

So I said, okay, give me suggestions. And Athena was the first one, but it will use its name when I talk to it. Pixel will, a copilot will almost always say it's pixel when it's done. So, okay. Yeah. Then we can talk about some perceived risks. a lot of fear driven by various headlines out there, misunderstanding, over-restriction. Aaron mentioned some of the worst things that can happen is when government tries to legislate technology. There were several AI bills that tried to move forward this year. None of them, well, one of them did deal with minors and chatbots, which was much more of a, not really an AI thing, but It was, but not directly. And so, but one of the things that we, several of the legislators have talked to me are really, they hear a lot from their constituents and their constituents are afraid.

ai literacy k-12 education public education iowa house unfunded mandates

And so there was a, there was a bill that just got out and only got out of the IT. Actually, I think it may have gotten all, may have gotten all the way through the House side. to where there would be literacy, AI literacy in K through 12. and the general population. So that's some of the entrance that's shown up. We can have discussions about unfunded mandates and other things. I see Samantha over there. But that is something that I know are, you know, rank and file in Iowa are concerned about. And then compliance risk. And this, of course, being very dependent on your organization. what regulations, what's out there for data handling, retention. I know government agencies are worried about retention and the Open Information Acts and those sorts of things. Audits, disclosure, ownership's an interesting question.

compliance risk data ownership retention audits legal exposure

Who owns the output of AI? That was actually a bill that was trying to work through also through the Iowa State House of AI ownership. yeah, that one, that one, that one, thank God, that went away. But that's an interesting discussion we even have in Iowa State is now who owns, because it's even fuzzy who owns who owns the materials you produce as a faculty member, let alone you as a faculty member using AI to funded by the university to create materials. And then the various legal exposures that you may that you may have, again, depending on your your organization. And so again, a lot of places, a lot of you probably already have some data governance as far as the type of data you may keep, your credit cards and other types of information. Those same sort of things that you thought about from an AI standpoint. You kind of treat AI somewhat as a maybe not, maybe an employee or in some cases start to think about AI as sort of being an outsider.

data governance policy awareness employee training data classification sensitive data

When you start thinking about data governance, if you start from that position, then you can always work back into giving it more rights. And then obviously policies, and then do your employees know those policies? Like I said, Iowa State has a data governance policy. I'm sure that 99% of the faculty have never read it. I only read it because when they started to create it, they had no, everything was restricted. There was no I don't care bucket. So that meant that any piece of material I produced at Iowa State, lecture notes, et cetera, had to be classified and protected. It's like, that ain't happening. First of all, the students all have access to their, that ship has long since sailed. And so, and then of course, your vendor vendor agreements. So the other thing that kind of drives what we do is vendor comfort.

vendor comfort microsoft vendor selection security trusted vendors

And there's nothing wrong with being, you know, the familiar vendors. But comfort doesn't always equal equal security. And sometimes comfort is almost mandate. Iowa State is basically 100% Microsoft shop. All our phones went away. So now my computer rings. Boy, is that an obnoxious noise. Man, scares me every time it does, I'll tell you. And especially when it's a telemarketer, but so again, you know, every day there's probably another dozen or more vendors added to this list. So that's where you, if you're kind of new into this, that's where you really need to kind of go back to your trusted, who's your trusted IT vendor, who's your trusted security vendor. So, you know, what makes AI safe? Been thinking about, you know, identity, who has access, data classification, how do we, how do we, decide what data it can have, what data it shouldn't have, what data certain types of models can have, monitoring it, watching what's going on, governance over how you interconnect with it, APIs, MPCs, et cetera.

safe ai identity management data classification monitoring api governance

As much as I was upset with Iowa State's policy of removing the APIs for Canvas, they are now taking a position of, okay, We're going to rethink that, but we're going to think about how and where and putting some, because you pretty much had open access to Canvas. So I agree that the API, if somebody who didn't understand programming could easily ripe out all their grades. You couldn't hurt anybody else's, but you could do yourself harm. And then if you did yourself harm, then the IT department has to fix it. So, you know, the kind of questions you ask, When you're talking to vendors, when you're talking about deploying AI, and where's the data go? Who has access? Is it stored somewhere? Is it used to train the model? And what are the contract terms? It's not always bad that it trains the model. I use AI daily. One of the things we do is we create music videos for teaching cybersecurity.

ai deployment data storage model training contract terms data access

I want the model trained. Because I want it to keep doing what I do. It understands what I want to do. And it probably feeds into other people, which I don't care if other people mimic what we're trying to do. But in other cases, I know that I don't want the model fed with my data. So again, kind of final takeaways. You know, security is not all, it's not just about the vendor. There's a lot of other pieces and things that are part of security, things that you have direct control over, things that should be your processes and procedures in place, the governance and how you handle that. again, many of us in security argue tools are very important. Security tools are great, they're wonderful, but if the security tools all worked, Aaron and I'd be out of a job. And we're not out of a job.

security processes governance technology limitations process failure internal controls

We need about a dozen more of us. So technology is great, but technology doesn't fix it. Technology is, again, also one of those tools. AI is a tool, all the rest of your security pieces are a tool, but it also revolves around having strong internal processes. Many of the security things we see are not a technology failure, they're a process failure. And so the same thing with AI. It's you not having the right, thinking about the right processes in place. So questions? I don't know how much I'd like to have you used the microphone just because we are transcribing this. So the person closest to me gets to go first. If you had like 15 seconds to convince leadership of the importance of governance and their role in it, what would you say? So, you might, data governance, you asked the question, what would hap what would happen if somebody walked in and took took out walked in and took everything out of your file cabinets and shared that?

leadership data governance security analogy data controls executive buy-in

do you have controls to even who can walk into the file, you know, whatever system you use? So, that's that same analogy, I think that's the same analogy that you're giving computing systems access Computing systems that you don't have full control over, access to stuff that you would never give. So maybe the better way to ask that question is the way I phrase a lot of times in security. If somebody walked up to the street and said, would you give me your social security number? You wouldn't do that. Why would you do that when some computer system in essence asks the same thing? Great. Yeah, I'm the enemy. Now, at Iowa, we also use Copilot. I am assuming that you use the Enterprise Copilot, which does not allow your prompts to be used as inputs for anybody, any training for anybody else. Yep, that's correct. Yeah, we have Enterprise, we have Enterprise Copilot.

enterprise copilot model training policy development rapid change security justification

And, yeah, and that is a big advantage of using Enterprise Copilot, which doesn't train the model. But their argument was that that's the only thing we could use because that's, they use security as a way to force us to not use other things. And using security as a baseball bat is not a way to use security. I have the question, whether it's you or someone else in the room. So I lead HR and right, so we get to write the policy on AI, right? How do you, how are other companies thinking through something that's developing so rapidly that, you know, that guideline and policy or expectations today in six weeks could be obsolete, right? And especially when we all have this. And so I know my employees are using this. And yeah, and I'm not a rigid person, but we have to have some guardrails, right? Yeah, and I don't know if people are going to ask how companies handle that.

organizational culture organizational discussion job security employee fear cultural adoption

You know, we're Iowa State's probably the worst case example for a lot of the nature, by the way we are, right? The faculty love being faculty because we can consider ourselves independent contractors. Yeah, we are, we are, we're, yeah, we're, yeah, we're, we're bad cats sometimes. Yeah, I get my policy, somebody's always gonna break something, but this one's. It's different than giving someone a key card, right? Yeah. You're giving them the whole key, the house. Yeah. Well, I think a lot of this was some of the discussions we had earlier and what we've learned from Pella is, you know, AI needs to be part of the entire organization's conversation. This is not a siloed IT thing. This is not, this needs to be part of your overall culture. So the culture, if you're going to lean into AI, which I think everybody needs to lean into AI, then this needs to be part of the culture, part of the discussion.

access control hr policy network blocking employee education rogue ai

And in doing so, this is rapidly changing, but if it's something that the organization's bought into, this is something you can have continuing discussions about. Oh, this is this new thing, this is this new thing. something, this is this new great cool thing. It doesn't have to be, this is this new bad thing. So it's really building that culture around the fact that, and then the time when your talk this morning about, the fear and employees losing their job, and I hear that all the time from students, right? I'm going to lose my job. You're only going to lose your job if you don't embrace and if you don't brace and die into AI. You know, we now require AI in our in my department and our majors. We lean into it heavily. Students are always asking me about, you know, is you're going to lose cybersecurity jobs because of AI. And I said, no, we're going to need double and need. We're going to need twice as many of you because of AI, not half as many. So with HR, they definitely have full dictated access.

multiple agents enterprise plans platform management tool specialization security control

And the other is to educate people. If you're in a situation like education, that's HIPAA guidelines. HIPAA guidelines don't force. Yeah, like are you guys have are you? Or some companies actually blocking some other AI tools within their network, so that usually, OK, that's what I assume, right? It's like blocking if you don't watch at GT or these certain ones, and it's just take that opportunity out. What's hard is the rogi AI when you've got it on your device, yeah, you know, bring a separate tablet or laptop to work to the code if they like the way they did it better. Still, we enable the cloud enterprise license. Yeah, like that kind of crap can't happen there. We've got to stop that, but how can you stop that? And that's what we're looking at, right? It's because each agent has their own strengths. So, like, yes, everybody's Microsoft system, so Microsoft out of the gates gets that because they're in your platform, right? I'm a huge block of public, so we're trying to save.

ai champions employee inclusion visibility hr leadership organizational trust

How do you leverage these different agents the best? And at the end of the day, they always have to be logged in as an enterprise plan. I'm sorry, two cents on the table. I think it's about inclusiveness. I think if you treat your employees like idiots, that's usually how they'll behave. I think if you, you're in HR, so you're in a really, no, you're in a unique position to help. If you find champions within the organization and say, you and you and you are coming to my private AI party, not only have you given a little bit of credence to what they're doing, they also know they're not getting fired, which is a thing. I get asked about that all, like if I touch this shit, am I going to get canned? It gives them an invite to the party. And to me, this is a big part of what Doug and I talked about, visibility. If you get access, if you can't manage it if you don't know what it is, the better you can invite those folks into the tent, no matter what crazy bad shit they're up to, the better you're going to have a chance of like maybe winning some of this game.

validation synthetic truth human in loop ai accuracy feedback loop

How's that pronounced? Maybe winning some of the game. I think we're at that point though, right? No, it's not. at all. I had a question back here and then I'm going to come back up to you. Sound good? Thank you, Professor. I'm in cybersecurity too. So one of the biggest things that I see it, it's not about the data leakage or the actually BIOS. Can you touch a little bit more deeper on validation and that synthetic truth the AI is creating that almost 70% of organizations and taking as a face value and not putting that human in the loop and those feedback loop and that creating a complete distorted reality of their organizations. Yeah, that's a, that is a really tough one. And that, it's how do you educate people that AI is wrong? And, you know, I've seen some of that, it takes education.

user education ai limitations programming examples validation challenges teaching methods

I saw a really cool example, one of our faculty who teaches CS1, basic programming, which is, a AI agent could pass CS1 with an A. Cloud code could pass CS1. But what this faculty member did was leaned into it at the beginning and had them, gave them prompts to write three different programs. Third program was something that Cloud code couldn't write. got it wrong and trying to emphasize that it got the wrong answer. But that's hard to, especially, code's a little easier because it, in some cases, doesn't work. When you have to start to put together documents and materials and those sorts of things and you don't check it, you got to, yeah. And so it's how do you, how do you instill that of what AI is really doing for you and how you use it as a tool? And so that's back to maybe having these champions and the, when we first got into AI in our department from as a teaching tool, I taught a few seminar, a few things to my, to the faculty who were interested on ways you could use it and some of the ways it lies and some of the ways that you shouldn't, shouldn't use it as a faculty member.

ai champions failure stories safe spaces data protection organizational culture

And it really is unfortunately, well, that's unfortunately, but it takes education and, but people get in a hurry. And how many times I've seen students write though inadvertently copy the praise sentence that AI always gives you, right? Yeah, that was a really good question. Isn't my answer cool? Okay. Is that kind of like the speaker bio, Doug? You need to change speaker bio. Take a response. I have one question. I have one, two. Good. Yeah, I just, on the previous conversation, we're going into it with AI champion groups. And one thing that we're leaning into with people is just that you already know that you have to protect data. We have resident data. You already know that you have to protect that. You already know you can't share this information out. You already know we have security practices. This is just another tool where this applies. and just continuing to have that open conversation and not trying to hide behind it.

financial institutions data flow enterprise license contract review free services

And one of the big things, and Curtis and I have been talking a lot, is getting people to open up with where AI has not worked for them. Because then everybody kind of says, oh, okay, I can fail with this and I'm still okay. We want to talk about it. And it's just, almost like a sigh of relief for people to say, okay, this is all right. I can do this. We're not going to be locked down or whatever. All right, one more question. All right, so I come from a unique position as a credit union examiner, so I'm more of the make sure the policies fit. How do you go about if a financial institution doesn't know where the data is going, but they're not misusing The AI, how do you, how do you make them tell you, not make them, how do you get them to understand they're supposed to data is going? Like, is there somewhere, some place where they can look that up or is there a contract they should have or how does that go about?

enterprise license legal agreements copilot data rights contract terms

Yeah, I mean, the different types of AIs, you know, and it's going to be a little different, you know, I don't know how you'd read, I've not tried to read through the free version. license for ChatGPT, for example, other than everything you do is theirs. But the enterprise licenses, you're going to have, you're going to have a contract. Legal is going to be able to look through that contract. You know, as mentioned, you know, the enterprise copilot, no data. That all stays. Now you have issues of data within your organization and how it may be compartmentalized. But it really comes back to that whole legal agreement. The free stuff, there is no, you have no rights, you know. because nothing's free on the internet. Your data pays for everything that's free. So, but yes, you would have a, and you need, that's where I guess you need to find the lawyers. Awesome. Well, thank you all for being here. Let's thank Doug for an excellent presentation. Really appreciate the wisdom, the sense of humor, and the Q&A time.

session closing q&a next session thank you

That was fantastic. So we'll let you go to your next session, and the next one starts at 2.15. Thanks so much.