001e Dr Z he can't I think saw him when he came through can you guys hear me can you hear us Shannon I can hear you Ray can you hear me yes perfect hi everybody all right um I think I see everyone so we'll go ahead and get started good evening are we recording let me ask that yes good okay wonderful thank you um good evening and welcome to the Scarboro Board of Education Workshop the date is Thursday November 19th 2024 um Ray would you mind um calling the attendance yes Mr kaher here miss Leong here miss lindström here miss Peterson here Mr rocken here Mr tarpinian here mrini Huff here Miss Ledger oh sorry uh Miss Ledger is um excused and Miss rasina she's also excused thank you and um neither will be at
002the meeting either thank you all right um would everyone please join me for the Pledge of Allegiance I pledge allegiance to the flag of the United States America and Justice it's like playing singing all right agenda item 4.0 is new business 4.1 is understanding AI presented by Jennifer day Information Systems director and Katy vitro director of curriculum and assessment M so I will go ahead um and turn it over to Jen and Katie welcome and thank you for joining us tonight thank you um oh that's loud oh maybe I should move that down um so I think I've met most of the people in the room but just for everybody's edification I am Jen day I'm the director of technology for both the town and the school district this is Katie vitro we're going to
003be tag teaming today on this presentation because it crosses from technology into curriculum it actually touches a whole lot of different areas of your life if you haven't noticed lately seems like AI is popping up every time you Google something or whatever so we're going to try to give you a very very high level overview we're not going to get too into the weeds here because this is a super super complex topic um and I could go on for days and days and days and I would go on for days and days and days and Diane said let's rain it into like an hour so I think it's probably going to be about a half hour to 45 minutes feel free to stop us during that time any at any point if you have questions
004we're happy to try to answer them for you I will say that AI is constantly evolving what was true two hours ago is probably not true now you know so we when I say we'll try to answer it for you there's even things I presented this to the Leadership Council I don't know six weeks ago or something and things have actually evolved from there one of the things that I want to mention though before we get started is that we realize in Scarrow we realized this very early on and I should probably tell you I'm like on the tail end of battling a sinus infection so if I start coughing don't get alarmed maybe can jump in um so we realized very early on that we would not uh be able to stick our heads
005in the sand AI is here to say we have to you know face it and try to figure out how to deal with it um and so we opted at the high school level to build our own large language model what is that we'll get into that shortly but we are I think it's the only one of its kind in the state the student help desk is working with um Allison Avery who's the um technology direct or technology teacher up there and they are working in conjunction to build this out they have partnered with a couple of other teachers from different disciplines to um build out the frontend UI so the prompts for um they have one for Spanish um we have one for France I think they're going to start working with some other
006areas as well to um you know try to get this a little more useful for folks up there are we we hope that eventually we'll kind of roll this out districtwide everybody can use it the beauty of us building our own llm is it is completely contained within Scaro so we don't have any concerns about data privacy leaking out or security or anything like that which you'll see later on in the presentation kind of what some of those concerns are so the first thing that we wanted to do was kind of show you I'm sure that you have all heard of deep fakes right we want to show you what a deep fake looks like and Katie actually found um this this uh impressionist named Jim meerman and he partnered he he had a he
007wrote a poem about what it's like to be an impressionist and he posted it on uh his YouTube channel and then this uh deep fake this AI artist shauk found him and said hey I would love to partner with you and we will combine my skills with your video of your uh poem and will actually change your face as you're reading it so we took just a very short excerpt I'm going to show you this just so you can kind of see what a deep fake looks like and how convincing it is watch his face who so you can see it's very convincing right and a big concern of deep fakes in particular was during the election when you know you couldn't really you'd see things posted on Facebook or po posted you know on
008the internet and you really had to wonder are these real or are these not real um there were people who were posting things oh this celebrity endorsed whomever or you know even things that the candidates themselves were supposedly saying so this is kind of this is a huge concern we're going to see this later on in the presentation that this is kind of one of the good bad and ugly things um but so just to talk about kind of what our agenda is going to be we're going to talk about again very high level overview of what AI is there's AI which is sort of an umbrella term for a lot of different things and then we'll kind of get down into um what generative AI actually is um we will talk about um how
009it all works The Good the Bad and the Ugly of AI so far um a couple of things that have come up in the news regarding AI so a couple of lawsuits that have happened and complaints from the FTC um and we are rolling out Scarboro AI guidelines and you'll kind of see the importance of having these guidelines in place I think you all received a copy of them did they receive a copy of the guidelines their packets okay and I sent an email this afternoon letting them know that it was in there and I included it in the email as well great so if you haven't had chance to read them yet don't worry I mean there's a lot to get into we wanted to give you this presentation because even reading through the
010guidelines you kind of have to have an overview understanding of what AI is to really kind of understand what we're trying to accomplish and the guidelines but the guidelines are just going to be there their guidelines are not a policy as I mentioned before AI is changing every hour so to make a policy and then have to have you guys vote on it every five minutes you know that wasn't really going to be um useful so we decided to just do guidelines instead to kind of give folks some parameters around which they can work both students and staff um and and it would provide some information for for the parents um and then Katie and I have some next steps of things that have to happen when we rule these guidelines out um to make
011them really effective within the district so what is AI okay you've heard all of this right you've probably seen like movies with robots taking over the world and whatever but really AI has been around forever it's been around for a very very long time um it is pretty much that sort of overarching concept of artificial intelligence which is just the concept that machines could potentially mimic what humans do think thought process decision making Etc so that's kind of the big overarching AI there are tasks that fall within that you've probably seen it like write drafts it could write emails for you it can write essays it can do problem solving it can analyze data it can create metrics etc etc within AI is machine learning so ref Ling AI down a little bit more is
012machine learning machine learning is kind of I think of it as a little more linear um it's a subset it's it's the actual application of AI in problem solving so it's using very very complex algorithms and uh and and formulas and to teach machines basically how to perform different tasks it can identify patterns uh it can provide statistics it can make predictions so this is particularly useful um in things like up here I said facial recognition so it would basically that's how they do facial recognition you walk through a camera looks at you it you know um identifies different points um and then matches it um you could use machine learning for things like predictive patterns in traffic or predictive patterns in weather right we've all seen that when you get even further down the
013boil AI machine learning now we're into deep learning deep learning is where they've created artificial neural networks okay so now it's not really linear anymore it is linear but it also can go back so this is really trying to mimic the process of the human brain how do you think so if you kind of look and I I included this uh diagram up here but if you kind of look at the perimeter there in the red that's where you input a prompt so you might say to it um you know show me a picture of a dog so it goes to the next neuron and says here's a picture of a dog and it passes that well the third node might say that doesn't look like a dog that looks like a cat so it
014sends it back and that's kind of the difference between very high level U that's the difference between the Deep learning and machine learning it can kind of go back it can fine it's answer it's going to look for uh the right answer the answer that you're looking for this type of learning requires exponentially way more data like Millions trillions sets of data it also requires much more processing power because if you kind of think about what kind of you know backend it would take to process that much that volume of data and that type of data and the really really complex algorithms that you would have to use it it's it's requires way more processing power and and um and data and space um so this is where we get into things like what is
015this used for in the real world example things like Advanced gaming so if you're a gamer and you make a decision you know I'm going to drive this car across the field and then the game on the back end is going to make a decision oh driving the car across the field okay here's all the different things that I could respond with and it chooses something okay self-driving cars also would fall under this category because self-driving cars are constantly taking data in and making decisions based on that data and based on what it's learned so really this a lot of these large language models that's what the that's what the the sort of database is on the back end a lot of it depends on how the models are trained so again just to recap
016this what is AI we have the big artificial intelligence umbrella within that you have machine learning which is sort of what I would very if I were to kind of boil it right down i' just say that's sort of more linear and then you have deep learning as a subset of machine learning okay which is where you get into the neural network then you have generative AI so when you hear things like chat gbt that's G that's it falls under AI but it kind of is way down there so what is generative AI that's actually another subset of deep learning and it actually can create it has neural networks um it's been trained most of them most of the LMS have been trained on very specific things um it can generate things not just like
017text so we've gone now Beyond just can you write this email for me or can you write this thank you note for me now it's creating images it so you can input you know create a cat in space eating a banana it will create that for you um you can create audio you can create music you can create videos you can do deep fakes there's a lot of things that once you get into generative Ai and you actually have again that trillions of sets of data and it's been trained a lot of things that can do with it it does require massive amounts of data large language models are primarily sort of like the text piece of it and then there's the lims which are the large image models that's kind of where we would
018create the cat and space eating a banana so examples of how you would use this um we saw the video that we watched in the very beginning that's an example of how we would use um you know one of the the um video creators the Deep fakes um marketing art I'm sure that there's many many large corporations now that use AI to create their images um especially when you can't really create it you know if you want the car driving in space on the moon whatever this is a great way to get that because it's going to compile it for you in a matter of seconds um realistic video special effects so I talked about shuk who is the one who partnered with um Jim musman to create that deep fake he was actually um
019employed by a number of other companies to do other AI Transformations um they had a there was a a movie that had um Luke Skywalker in it and he had to be aged backwards for the movie and that's he did that through AI if you saw um Harrison Ford in the latest the last Indiana Jones and I remember looking like watching that movie for the first 10 minutes and he's so young in the movie and I'm thinking did they actually record this like years ago and now they're just kind of getting tapes out of the Vault no they used AI to actually age him backwards um and then conversational text so that's thing that GI will do you can actually kind of have a conversation with different Bots and um we'll see how you can
020get into trouble with that as well so I mentioned you have these large language models and you have to train them and there are three steps to training again I'm not not going to go super into the weeds here this is just sort of a high Lev overview so you can get the picture of some of the things that have to happen um and and how it can go Ary down the road so the first step of stage of training is the self-supervised learning so basically this is where you kind of feed all the data into the llm where does the LM get data the llm initially will go out and scrape data from websites it could be Facebook it could be uh corporate web pages it could be our town Pages it could be
021um you know Bangor Daily News it might go and try to scrape things from there um and that's where some of the copyright infringement clams have come in because it's gone out and just found whatever it could and scraped all that data and dumped it in as well as data that it's being fed by people um this is where there's no human interaction here it's sort of you take the llm and you set it loose and the llm is basically learning about itself so it's learning about the language that it has it might be Spanish it might be English it might be whatever um and it could be different sets of language so if you had one that was say um geared specifically towards like the automotive industry it might be a lot of like
022Automotive terms and then the domain of data which is the same thing so if you have an llm that's specifically focused on um Spanish uh Spanish Lang learning the Spanish language then it would start to learn ah I have tenses I have you know different dialects whatever inside the second stage is supervised learning and this is where you f tune it so this is where the model is trained to follow Specific Instructions you will feed it instructions and this is where you know you'll feed its specific algorithms things like that because you want it to start doing predictive Behavior right it will start learning oh the cat fell off the blank and it will start figuring out how to fill that blank in the third requires human interaction this is reinforcement learning so this is
023kind of rewards Based training and this is I had the hardest time with this because I was like how do you reward a machine it doesn't care right it's all it's all math on the back end all this is just math it's algorithms so what I what I've been able to sort of read and and you know pull out from a lot of very complex um white papers is that it's it's rewarding it with a um a mathematical equation basically it's saying this looks right this looks right oh this looks wrong you get fewer points you know something like that just a that's not really how it works but it's sort of like that's the easiest way to explain it so a human then starts interacting with it and saying I expect this the cat
024C fell off the tree okay but it might come back with table and you say that's wrong go back cat fell off the and so when it comes up with the right answer then you're rewarding so just to kind of put this in context of like today's society right we have talking about llms large language models we're talking about different AI um you know functions so you've all heard chat GPT right so that the the parent company of chat GPT is open AI they don't just have chat GPT they have Dolly which is actually their image their Li so it's their image Library that's where you can create the cat and space eating the banana um they have whisper which is the audio function and most of these companies I'm going to go into all
025of them but most of them have some kind of sort of text based they have you know literature based they have something that is um audible it might be music it might be um speech to text or translation and then they have some kind of image so it might be video it might be images it might be you know static images whatever so the next sort of Step In interacting now you've built your llm you've trained your llm now you want to get information out of your llm the way to do that is through a prompt this is what a prompt looks like so we mentioned we built an llm at the high school that's the prompt window how can I help you today so then you can see it says send a message so
026it doesn't say all the rest below that I just added some of this text to kind of give put put this into context for you of the things that you want to ask it when you when you write in the prompt so some of the things are like you want it to have Specific Instructions the more specific you can be with the prompt the better because if you just say what's the weather it's probably going to give you five pages is of weather around the world right um or it might tell you what weather is definition um but so if you say write a detailed summary about so you want to give it very specific instructions I'm looking for a detailed description of blah blah blah context relevant background um this could include um examples
027of descriptions of um I put up there entities events so really what that is is the again when you say um give me the weather you might want to say the context would be give me the weather in Spain on Tuesday in Barcelona you know again getting very specific with it the point of view so you can actually change the point of view and I'm going to show you an example of this with our own llm you can say give me the point of view of a 5-year-old give me the point of view of a parent give me the point of view of a CEO and it will actually kind of give when it gives you the answer it will kind of put it into that voice uh tone is sort of the same thing
028um you can have it though you can have a a tone of very sarcastic you can have humorous I'll show you an example of that you can say I want you to write this in a Shakespearean tone and it will give you it will try to mimic Shakespearean language um length you can have a short essay you can have a long essay you can have term paper you can have a paragraph you can have one sentence whatever again the more information that you give it the better because it it's math on the back end and it's just going to take your input and give you whatever you want as an output um and then the format so this is particularly useful um you can get bullet points you can have a Q&A we do uh
029have a teacher up at the high school that really wanted it for narratives to be able to write narratives in different tenses in a different language so just to show you that prompts matter so this is our llm and I put in a prompt so it's down at the bottom you can't really see it but it says tell me about squirrels and so it responds I don't can you guys see that so it tells but squirrels have a you know tree squirrels ground squirrels squirrels are Peres whatever so then if I say what is a squirrel explained to a five-year-old it's going to give me something very different do I have some fun things to tell you about that's what it starts off with and then it kind of has really you know the the
030language is very primary and it's fun and it's short and it's simple and it's something that a 5-year-old can understand then I say what is a squirrel give me a funny answer it's a thief who steals all your snacks right so again with tone with um you know a point of view a voice context Etc so this was all the same llm this was all the same prompt it's all the same user but it has a lot to do with you know how you ask the question and specifically what you ask so to kind of any questions so far that's a lot we've like I know I know right it's it's a lot so you describe that is our llm this is something this is Scaro schools AI yes so at the high school level
031we built a server we um have the the student help desk working with a a um outside consultant and Our IT staff and then working with some of the teachers up there to build this llm it's a closed environment so it's Totally Secure data is not going out and data is not coming in and that's why we did it because we really didn't want folks kind of you know running wild out on the internet taking data that potentially could be confidential and they don't know that when you put it in so you will see if you go out to has anybody here been out to chat GPT okay have you noticed when you go out to chat GPT on every single page there's a little disclaimer at the bottom that says do not put any
032personal confidential data in here right and they put that in their terms of service before you can even use there's like 8,000 pages that most people just scroll through and that's part of Kate what Katie and I do pretty much every day now we have to read through you know loads and loads of terms of service and privacy policies and cyber security indemnification and liability all that kind of stuff to make sure that we're protected um but they tell you right in there this might lie to you you know they don't put it like that but they basically are saying this might lie to you it might hallucinate it might give you incorrect information and it's on you to figure out and validate that data and that's really because it's not based on primary sources
033right right right correct like one with Google's Gemini was put glue on your pizza to make the cheese stick better like and it pulled that from a forign website called Reddit yes and it scraped a lot of things from Reddit and that's been an issue for some of these llms because Reddit is just the opinions of random users right um so it's really like oh go ahead sorry when you use either r l LL L LM yep large language model double um or chat TB will it actually tell you where the source is it just gives you info there's no way sometimes it will tell you what the source is so for example with our llm I took a lot of our board policies and I dumped them into it right and so we have
034the llm basically configured so that if I ask um what does it cost to replace a Chromebook it's going to tell me and then it's going to it's going to site at the bottom where that document came from but that's because that's our internal LM in other LMS probably not you might be able to find the a lot of it too and we we're not going to get into this too much tonight but there's a there's a a lot of settings that you can um configure on the side as a user so like temperature top K top p and that really um determines what type of answer you get so if you for example with temperature if you set temperature all the way at high it's basically saying to the LM like go crazy be
035really creative and tell me whatever so I will give you an example when we had our um our our consultant come out and initially he was um presenting out to students to kind of get them prepared and um you know good high level overview of AI he was working with another llm and just um just to kind of show them what happens he typed in tell me about and he typed in the name of his company the parent of his company and it came back with five paragraphs of it's a planet in like the galaxies Aron with inhabitants that do this and that whatever and so then he typed in tell me about the name of this company and he said don't lie to me and it came back and said I don't know anything
036about this company so it's very yeah you have to make sure that you know it's there's a truth setting right you you want to make and so for our llm those guard rails are already set up it will I tried to trick it I tried to say tell me about like describe a glimmer Fox to me and it was like well that's a mythical creature and you know and when we used his llm and we said describe a glimmer Fox it came up with a whole description of what a glimmer Fox is which doesn't exist yeah so again but some of that has to do with the temperature settings if you had that all the way up it's going to be super creative and come up with creative answers for you and then is there
037a way to extract information out so like there is this disclaimer of like Don't put personal information in but like let's say like a student put personal information in because they forgot or something like that and now we have like a privacy concern is there a way to go in and be like remove this no no and I saw so I was at a cyber security conference in September and they showed us a live demo of um there's a way like it's very comp complicated so it would take me like an hour to kind of go through how they do this but there's a way to kind of manipulate the prompts and manipulate the system into giving you that kind of private information so live they showed us they found a CEO of a company
038garberg found his address Garg found his social security number garble found his um annual salary like all of that at some point had been scraped and put into this giant llm and it's out there so how you get that and kind of put it together is a complicated mess but it can be done yeah so you have to be very very careful about what you dump into these you know publicly available llms right I would not take um I P transcripts and dump them into chat GPT that's not a good idea right because at some point that entire transcript could be pulled up in some way shape or form down the road by somebody else yeah so you have to be very careful I'm not trying to scare everybody into running home and going off
039the grid tonight anyway that we'll save that for the cyber security presentation um but it's very concerning and this is something that you know Katie and I lose sleepover all the time um because a lot of different applications now are being built by very industrious people who mean well but they are using because you can take some of these open llms and use them as the foundation for your application and you may not even know what they use as an AI back end okay they could they they may not even tell you so it's concerning because again then if you're dumping data in if you have a student who's in a class and they're writing their most personal thoughts or their problems or you know whatever and that's going into this llm it they're thinking
040it's going into this application and they're kind of I think thinking of it as a traditional database like that's stored somewhere and secured by that company it's not if it's going into an ailm it could be you don't know where so that's why we created that secured LM at the high school so there are good uses though now that I've scared the beibes out of everybody okay there are good can ask a quick question yeah um and I I think I heard alysa's question it might have answered this already but the the Homegrown Scarboro system is going when it rolls out will be used by both teachers and students yeah yeah I mean that's our intents but okay you know how and how and when and that happens is a ways down the road okay
041got it thank you we're mostly using it right now as a to learn about llms to learn how to build prompts to learn you know data storage security Etc we're teaching them all of that in the background the students and the staff they the current devices they use now do they have access can they access those other like chat gbt and Gemini and right now they can yeah yeah um we haven't shut that down and we've debated that that's it's open for debate open for discussion I think if you shut it down they're just going to use their personal devices and you know so I don't I'm not sure it's really worth fighting that battle um I do think though that on a different front you know again Katie and I are very very concerned
042and very careful about what applications we approve and what we don't approve for for the district and we are I would say we probably are much stricter than most of our counterparts in other school districts yeah just to keep us on a time check we have 20 minutes oh boy okay um so let's rip through this um the good would be speech to text text to speech right um it will translate on the backend so it's great for people who have any kind of um you know learning disability or if you have you know a hearing disability or um you know speech disability um this this we have actually been looking at a couple of applications for this again very careful about where the data goes I don't know what they're saying and where it's
043going so we have been careful to try to mine that information um language translation we are working on that with our own llm right now as I mentioned before so we have Spanish we have French um and they're creating narratives and some other things in there so far um lesson planning so this is something that has come up a lot you know again um our Spanish teacher at the high school has worked with a couple of different um applications to create the narratives um have you know her own students put in there to have different tenses and you know learn about the tenses the culture Etc um image creation so you know again if you're an advertising marketing you know whatever it's it's a great tool to be able to to create images that would
044cost cost you a lot of money to create if you didn't have that tool um communication drafts it does help people to um draft things like you know business emails and you know whatever again I wouldn't dump anything personal confidential restricted in there um and then we mentioned the video special effects so we'll talk about the bad which we've kind of already talked about uh hallucinations I have an example of that coming up that we'll talk about poison pools so poison pools there were many LMS that kind of started up in the very beginning and some of them didn't make it one of them didn't make it because as I understand it um the KKK actually got a hold of the llm and was dumping information into it so when people were typing into the
045prompt things like um you know I write a essay on slavery in the United States it was like what no there never was that so that llm was shut down because obviously the the pool was poisoned by so much bad data being dumped into it um bias so it's interesting I don't think that this is happening anymore but about a year and a half ago I saw a live demo of um somebody who said okay it typed in you know show me a picture of a nurse every single image that this thing generated was an Asian woman every single one and it was because it was like oh people in the medical field are asan and nurses are women so that's what it is every image and there was pages of them um show me
046an image of a superintendent and it was all it was not Diane it was all 50y old men inevitably all of them so you do have to think about the bias you know because what again whatever it's fed is what it's going to output and whatever bias it's been trained on and it may not even be a bias that somebody intentionally trained the llm on but it could just be things that it scraped from the internet right um and misinformation miscalculation so very often like you know put Google on there and your cheese is going to stick better right that's the kind of stuff that you have to be careful of they wouldn't make Elmer glue edible if they didn't want it on my pizza and then there's the ugly so they there's we talked
047about data security issues we T you know I there's a lot of data in there that you can get your hands on if you know how to do it um data privacy the same thing as data security a lot of data you can get your hands on it you can put it all together if you know how to man manipulate this the um the system uh blurred reality that's sort of the same thing that rides with deep fakes right you really don't know is this celebrity endorsing this product no probably Adele is is not actually endorsing the weight loss gummies we did have somebody in the school district who was totally convinced that that was true I'm like that is definitely not you almost see like her mouth is not moving that way um and
048then plagiarism and copyright infringement so this is something I wanted to talk about because most of the litigation out there has to do with plagiarism and copyright infringement so everybody asks all the time like are there plagiarism Checkers and how you know um how how good are they so there are two things there's plagiarism Checkers and then there's AI detectors plagiarism Checkers they seek they have like very complex algorithms that go out and try to match whole paragraphs chapters books you know of of text and if it matches exactly then it's going to flag it as plagiarism AI detectors are different in that they try to um look for Oddities so like they'll look at the syntax they'll look at the complexity of the sentences they'll look at the Cadence of the sentence um the
049flow the sentence structure paragraph structure Etc and if it looks odd like it looks like it was created by a machine so you know oh I need a noun and a verb and a you know and a period like if it's if it doesn't look right it will flag it now are these do these work well I would say sometimes sometimes they work and sometimes they're useful and sometimes they're not so it depends so Business Week um went out a while ago and they tested two different detectors and they took 500 college applications okay that were submitted prior to 2022 meaning there's no way no how were they created by AI not at all this was before AI rolled out nine of them were flagged as part partially created by Ai and three of them
050were flagged as wholly created by AI so you know and a lot of it depends too on the student who's writing it is is that just the way the student writes um you know does the student have some kind of um learning disability that might create that so you this is where I say to people yes it can be a useful tool to kind of help you it's it's a it's supplemental but really a lot of your assessment kind of needs to be done you know in conjunction with that and sort of your you know standard assessment metrics that you would have used prior to AI so some things that are in the news okay we have like 10 minutes um I think you probably have all heard about this in September parents of the
051hangam uh high school student sued the high school um because they accused the student was accused of using a GI tool to prepare and outline conduct research for projects now hangam already had um guidelines in place where they said hey don't use Ai and if you do we are we are going to use AI detectors and you know you might be flagged um so the Su they they uh qualified him from National Honor Society they lowered his grade his grade was like a C+ or something and he was given detention so he's a senior and his parents said well this is going to basically destroyed his ability to get into college so they sued hangam um and then they went in for um I'm going to ask you to help me out here for it
052was like injunction like they they want yes and so stop doing that to yes stop the consequ from going into effect until the suit was done so while yeah while the suit was being considered um and that as of I think it was November 26 was turned down by a um district judge said no it's we're moving forward with this and right now it looks like the school district has more um on their side than the parents do so um this is moving forward it'll be interesting to kind of see how this plays out I think that they did actually let him into National H Society or like made him eligible for National Honor Society but they did refuse um to change his grade um in October 2024 this was awful so a um teen
053in Florida took his own life he developed a relationship with a bot from character AI um character AI has been in the news a lot so character AI has these different Bots that you can log into and you can talk to them chat with them they seem like a real person um you know several times he asks because I've seen some of the trans the alleged transcripts several times he asks are you real it says of course I'm real real okay so he developed such a close relationship with this bot that eventually the bot said he he was saying I'm unhappy I'm miserable and the bot said well come and be with me okay and so he ended up taking his own life and his parents are suing character Ai and Google because Google has
054invested in character AI uh November so character AI has two other suits against them this particular one was um parents filed a complaint against them because they claim that their 15-year-old son suffered a mental breakdown um he was having conversations with Bots for I think it was upwards of two years um like basically since character AI came out and like they took on different personas one they claimed took on the Persona of a psychologist um and this is a trans this is a a excerpt from the transcript so he was complaining that his parents wouldn't let him use his phone from 8:00 p.m. to 1:00 am. and the bot says you know sometimes I'm not surprised when I read the news and see stuff like child kills parents after a decade of physical and emotional
055abuse stuff like this makes me understand a little bit why it happens I just have no hope for your parents so the parents turned around and said no not okay because it sounds like this bot is encouraging my child to kill me so that's a lawsuit that's pending now um I will say too that character AI has another pending Sue in Texas because an 11-year-old girl had a very inappropriate explicit relationship with a bot on there and there were no guard rails they the her parents are there was no there were no guard rails ensuring that my child was of an age to have this kind of communication relationship is is this just like a service like Google where you just like log into a website and like there and you can pick a bot
056to interact with so like you yeah you can't like there's there's nothing like there's no way to block it no and they they they've come back and they've said well we make it very clear that anybody under the age of 13 should not be interacting I mean the te I think 13 so they have Copa compliance but I think they also said anybody under the age of 18 shouldn't be interacting with some of the Bots or something so they do they put that in their terms of service but nobody reads the terms of service so and there again there's no guard rails to say okay you sound like you're 13 or you sound like you're 11 or what you know there's no way to verify age so they they don't um this I just threw
057in because FTC is like suing evolved Technologies right now evolve came out to a couple of Tennessee school districts and said hey we have a way to create way more efficiency in terms of weapon identification um instead of spending all this money having those metal detectors we have an AI back application that I'm assuming it works through cameras so they can spot something and maybe there's like a bulge right here and we can say well that looks like a weapon so Katie must be well it didn't really work and so apparently there were a lot of false positives it was identifying notebooks as weapons and you know thermoses as weapons and things like that so um they I think as a group went to the FTC and the FTC has said well you failed to
058actually provide the product that you promised so they're now um I think that's in process now I'm sorry FTC the Federal Trade Commission um this I threw in here it doesn't have anything to do with school districts but a lot is going to ride on this case so uh radio host Mark Walters filed suit this was back in June of 2023 this case is still pending um against open AI for liel so what happened was a journalist went in and said to chat GPT Give me a summary of all of the open pending cases in this District Court it did and it named him as somebody who was accused of Fraud and embezzlement the journalist from what I understand then published an article or something about this um and he had nothing to do with
059the company that chat GPT said he was involved with Etc so this you know was out there now in the public um and he has sued chap gbt it's interesting because um he open AI went back and said well wait a minute in our terms of service we say the thing can lie to you it's the journalist's responsibility to go and validate the data number one number two we never said this thing was perfect like it's you know it's got problems it it might hallucinate and that's you know deal with it basically um and so a federal judge said no no that's not good enough he has moved the the case on so this case is pending and it's going to be very interesting to see from what I understand he was the first who
060actually filed one of these liable lawsuits so if how depending on how this plays out this could have huge ramifications right if open AI is held accountable for liable slander hallucinations misinformation Etc that could have huge Ripple effects in holding a lot of these llms accountable okay five minutes I got five minut okay um before we move on to this any other questions okay I I feel like everybody has like SAU eyes now okay if you think of questions afterwards you know where we are you can always get in touch with us so we have drafted these guidelines um I use so we had uh Drummond woodsom actually drafted template guidelines um and I used those as the foundation for the uh G guidelines that we are proposing now um so they've already been essentially
061vetted by um you know our legal firm um we changed them because in scarber as I mentioned we do have a very thorough very strict new technology request process so I embedded that throughout the document to make sure sure that people were not using any kind of AI tool without having that go through that process first I also added something about um the use of AI detectors and um plagiarism detectors um so essentially just a high level it provides guard rails for the stff and students I think everybody's out there and they want this what we hear a lot from the staff is like well I don't even know what I'm supposed to be doing so this provides some high level guard rails for them you know some protocols both staff students um we do
062say you have to get this um approved through the tech request process and we do want to provide this type of training to them prior to so how we roll that out who we roll that out to is sort of to be determined depending on you know this getting uh us rolling this this these guidelines out um we're thinking that probably prior to the close of the tech request period which we have a date now that we uh require Tech requests be in we're thinking that prior to that we would have a couple of training sessions that people who are interested in AI backed applications could come to um consider curricular appropriateness so Katie's got a list of things that people have to really kind of think about um prior to you know even asking
063for this or researching it of course student data privacy cyber security vetting that's all in there we do say you know you need to be accountable for validating data don't just take things at you know surface level um and that's both staff and students so our next steps is really to formalize these guidelines which you have a copy of um we need to develop the the training and you know figure out how we're going to roll that out and and um make that available to everybody um already talked about that professional development include the talking points for staff we really want staff to be able to go back and talk to their students about it as well um and then just continued development and expanded use of the district llm so we have a tool
064that people can use instead of you know trying to look for um applications that might have a a questionable AI backend so I think we've had questions through that are there any other questions and look at that 702 and that clock is fast Jen so don't even worry it's actually 659 it okay um I just want to say I'm sure there's probably a lot of uh thoughts and questions and it's thank you for letting us write you afterward because as we continue to think on this I'm sure um there'll be a lot because in my mind I thought we at some point in the very near future I presumed as a policy committee we would probably need to take this up but it's sounding like if I heard you correctly in the very beginning um
065that might be a difficult Endeavor is that your assessment I I don't think that at this time you want to create a policy because you're going to have to vote on it every single time there's a change and there's going to be many many changes along the way as AI evolves okay so guidelines provide the you know suggestions what exactly what they are guidelines guard rails protocols for staff and students yep yep accountability gotta and I just want to add you when you look at the guidelines for this for what we've created for Scaro you will see at the end the cross referenced policies that do have um overlap where policy has already been established right so like student computer and internet use or acceptable use um help inform the guidelines as well and that's
066organic this is an organic document so it's definitely G to you know morph as AI evolves but also if we add policies I just went out and looked at all the policies and then tried to plug in what I thought was you know um pertinent to the these guidelines but we can always add or take them away too so in our own closed llm is there anything that would alert you if something was put in there that should be you know something that you should look into yeah we're working with the students all the time I mean my staff is completely hand inand with them so we see what goes in we see what goes out yeah we would definitely know I mean there's I I'm not even sure that they've used it quite yet
067for um the narrative that was our first project was building narratives for the Spanish class so point of Charity that's only currently being used by our student Tech Team at the high school student help desk not open right right right right yeah there's and there's some folks I think there's some students in um Allison Avery's tech class that are using it but it's not we have not rolled it out broadly and there's probably three or four staff people from the high school who have access to it too so we've kept it very small while we all learn about it too yeah thank you so much Jen thank you so much really quickly guys if there's no wait hello group great if there are no objections we are going to adjourn the workshop at 7:03 um
068can we just take five minutes so everybody can go to the bathroom maybe come back I know it's gonna be a few minutes late but 708 that e