hey I’m Craig Numark probably best knownto anyone as the Craigslist guy but I’mhere today talking about somethingdifferent American democracy is underthreat and we all need to work togetherto protect the country and to protectourdemocracy it’s kind of like in World WarII when everyone was expected to play arole and the first part of this is forall of us to learn a little more and towork together strengthening Americandemocracy strengthening the country theCivic life series is a big part of thatand the 92nd Street y New York is theright forum for delivering all that Iappreciate you’re joining methanks all right uh thanks everyone forjoining us um we’re really excited to behere we want to give our thanks to 92nfrom making this happen um also want totake a chance to invite everyone toattend the Aspen cyber Summit which wewill be holding at 92n on November 15you can find info at Aspen cyerssummit.org um look there for speakersfrom the White House from the FBI uhMichigan Secretary of State uh we’ll betalking about AI cyber security uh crimenation state attacks election securityum we’ll also have a special discussionto dig deeply into the bird droneconspiracy with the public affairsofficer from Birds aren’t real um sodefinitely something you don’t want tomiss and as a viewer of this webinarwe’d like to offer everyone tickets at aa steep discount rate of $100 you canregister at Aspen cyers summit.org anduse the promo code 92 n92 n that willget you $100 registration um so todaywe’re here to talk about AI specificallygenerative AI um literally a year agovery few people people were talkingabout this I know I wasn’t that changedalmost overnight on I think November30th of last year when open AI releasedchat GPT and since then the conversationis everywhere um government circlesmedia you see late night hosts talkingabout it um in schools it’s been a keyelement even behind some of the strikesyou’ve seen the writer Guild uh to thecar industry um but what is thistechnology why should you care um and asthe title says what are five things youshould know so we’re going to get intoall of that going to have a greatconversation with Yasmin and Nicole umbut first ask folks to introducethemselves let me throw to Nicole thanksJeff um and thanks for having me todayI’m Nicole tisel I am a senior advisorwith the Aspen digital group I focusmostly on cyber security Workforce andeducation I’ve previously worked for theWhite House on the National SecurityCouncil and also for the US Congress umfor the house committee on homelandsecurity I’ve always focused on cterterrorism cyber and Counter Intelligencewhich which sounds scary but I promiseum we’re going to make sure today isn’tscary um and that’s it I’ll turn it overto youyasine hey I’m Yasmin green I’m the CEOof a unit within Google called jigsa ourgoal is to develop technology to makepeople safer online we look at threatslike disinformation toxicity extremismand repressive censorshipI’m sure none of those issues have beenuh been an issue with uh generative AIor other Ai and I’m Jeff Green I am thedirector of the cyber security programat the Aspen Institute um history andgovernment and private sector as well onon these technology issues and um I candisclose for this group that at onepoint I was a lawyer deep in my past butI usually usually keep that quiet um sofor today we’re going to have threeparts of the discussion real freeflowing first we’re going to talk aboutwhat is is AI generally and specificallywhat is this generative AI chat GPT Bardall the other tools that are out thereum second what are some of the thingsthat that we think as experts in thisspace or pseudo experts if in my case uhthings that you would want to know andthen finally we’re going to try to diginto some of the things that youactually want to know we did put out acall for questions before the webinarwe’ve got a good number and we will getthrough as many as we can um so yeahAsin if you want to kick us off andNicole jump in and I’ll jump in let’smake this real Interactiveum what is AI and and what is generalgenerative AI How would how would youdescribe it what are your thoughts on ituh the simplest fewest words is makingmachines intelligent um the term AI ispretty much a a term of art you knowit’s an aspiration there’s like a jokethat it’s you know if it’s already ifit’s already working then we you know wedon’t glorify it as as AI or um yeah uhto be specific the when we’re talkingabout AI we’re generally talking abouttraining um machine learning modelsbased on a lot of data to see patternswhich is really interesting because youknow the the inspiration for the thestructure and design of these machinelearning models is the brain neuronetand how we learn things as humans um andon on our team at jigsaw we were youknow our first machine learning projectwas around 200 200 16 uh when I had a2-year-old and I was like reallylearning about how machine learning ummodels work how they learn you know lotsof patterns and then they can Intuit itwhat to do in a context that theyhaven’t seen before so previously withum with software that you know that wasrules based so you had to kind of giveit very strict parameters for how tobehave uh but with machine learningmodels they can almost INT in it what todo which is how children operate so itwas both kind of I had like fireworksgoing off as we were learning aboutmachine learning models and I waslearning about Child Development um andthe interesting thing is it has alsoapplications for when things go wrongwhich maybe we can talk about yeah if wehave time I want to hear about whofollows rules better the two-year oldsor the uh the AI tools um Nicoleanything you any thoughts you have therewhat you want toadd yeah I mean I just add to to whatYas Meine is saying um especially Ithink about artificial intelligence asan umbrella and generative AI is justone of the pieces under that umbrella umand when I think about um AI because Ispend so much time writing um I’m alsoan attorney but I focus on policy Ithink about generative AI as a tool tohelp me create content ideas list umthings that I already know how to do butit helps to expedite the process sogender of AI um it it is the creation ofcontent that sets it apart from other AIor machine learning tools out there Ithink so and I think even and we’ll geta little we’ll get into this a littlebit um you know content is subjectivebecause sometimes the content is notalways um it just not always as helpfulfor your end result um but I actually Itry to visualize words a lot so Iliterally just think about the Gen isgenerating ideas um or generating a ideaAI um and so yeah I think for mostpeople um the uses that they would orhow they would use generative AI isusually for creating some type ofwritten content or some type of umcreative Pursuit umyeah what do you think asine you agreewith that yeah yeah I mean it’s justinteresting how you know you describedJeff that back in November you knowthere was this like mainstreaming ofconversational AI that just rockedeveryone’s world even though the actualthere wasn’t so much of a technicalbreakthrough at that time as much asthere was this um interface for theaverage person to go in and actually trytalking to Ai and have it talk back soit’s like dynamically generating contentin conversation with you and I thinkthat just feelsso just feels like a different Paradigmto us as people to be able to generatecontent on on the fly to to what we hadbeen using AI for before which had beento classify there were really smartclassification systems that couldclassify content that they hadn’t seenbefore uh but now we’re generatingcontent and I think we were just alsobesides ourselves with excitement andand really being impressed at um howthis this thing was operating because ityou know you could ask the same questionseveral times and get a differentresponse so there was there wassomething that felt human and creativeabout what we were getting out of thesechat Bots but for you I mean I I assumethis wasn’t particularly new for youbeing in jigsaw and Google you know itwas new to the world but were youwatching everyone else discover thisyeah we’ve you know Google has so jigswas a part of of Google and we get tosee the stuff that they’re developingvery early and they did have their ownversion of chatbots the thing that thathappened with chat GPT which was theversion of a chatbot that most peoplesaw which came out of a company calledopen aai was that they really went likefull creativity which meant that youknow they put some safety measures inthere but it meant like you could youknow you really in for a wild ridetalking to this thing like Google’sversion was always a bit more you knowlike contained um so it was kind of coolto see I didn’t think Google would haveput that out then um so I thought it wascool to see something in the wild and tosee all these Incredible use cases thatpeople had you know everything from likewriting songs about you know funnytopics in the style of your favoriteartists to people um you know doingwriting text in languages that theycan’t speak um to a um a really touchingstory about someone who is you knowthere was nobody who was their identitylike their sexual orientation in theirtown and they would just talk to chatGPT to get positive affirmation and youknow I don’t think that that was theintention that there would be any typeof um application for helping peoplewith emotional well-being anything butjust because it was full creativity andvery little Prudence in the originaldeployment of chat GPT it was it waskind of a miracle to see in all thedifferent directions in which it wentNicole do you remember your firstinteraction with uh with thechatbot I do actually yasine as you weretalking I was thinking about how um thechat box just made the conversation justturned into like a very public thing butand um I was working in Congress and wehad already done like all of thisresearch or we were continuing to doresearch on artificial intelligence andit felt like no one ever used to ask meabout it and then all of a suddeneverybody wanted to um back to thequestion at hand my first prompt and Iput this prompt into um every chat boxwhen I use it first just to get a littlebit of context of the content that itwill produce um it’s a national securityquestion it’s a foreign affairs questionand I just asked if Henry Kissinger wasanti-apartheid or pro- aparti in SouthAfrica um and I judged these chat boxy’all based off their answer and I willsay um most of them do not do very wellum but I think that’s a really goodexample of that’s actually a verycomplicated question um over time Henrykisser did become anti- aarte butbecause the chat boxes are trying togive you a very specific answersometimes they say yes and sometimesthey say no depending on which period oftime they’re talking about in his careerso that’s always my firstquestion I um it’s said you’re much moreserious than me I I had having been anan AI skeptic and I was asked to do tomoderate a panel uh on AI and about aweek before it occurred to me I probablyshould try this out if I’m G to be infront of people talking about it so thefirst thing I asked one to do was writea review of a of a comedy movie using aparticular Regional dialect in theUnitedStates and it was what’s it was stunninglike I it was it it got the themannerisms it got both the substance andsome of the Nuance of the movie and thethe the conversational mannerisms and itwas hilarious I had video of my kidstrying trying to read it and they theycouldn’t get the words out but that waswhen I sat back I was like oh my Godthis thing did this inseconds yeah it’s so funny I actuallywaited a really long time to try them umI was kind of in the camp and I don’tknow know if any of our viewers are arein that space but um because I’m awriter I’m a lawyer I was an Englishmajor like writing is very important tome and I was a little jealous I guess orjust I was like there’s no way that achat box can like write and analyze anddo the things that I can do it’s a veryarrogant way of thinking about them umso I actually didn’t even try I I wasnot one of the like early adopters ofchat GPT it wasn’t until this Summerthat I tried it for the first time yinedo you have any recollection of thefirst time you you played around with itit was definitely my like cocktail partyconversation for like November DecemberI I live in New York and I went home toLondon and I would just to carry mylaptop around with me being just checkthis out try something try something andif somebody tried something that didn’tquite showcase the Brilliance of thechatbot I was re I really judged I heldit against them you know because I waslike well that wasn’t quite the rightthing and like it’s you know you grademe how do you grade me for my firstinteraction great great great yeahbecause it’s it’s almost like it’salmost like it felt like um a magictrick and you want it to be executedwell every time so that people have thatfeeling of awe of something that’spossible that you know wasn’t wasn’tpossible before the interesting thing isthat you know open aai the company thatmade chat GPT have continued to itthey’ve built more powerful foundationalmodels the thing like the kind offoundation of it that’s powering it andthen they’ve done kind of put moresafeguards in place for the interfacepart that allows us to interact with itand chat with it and so it’s you knowit’s not full creativity anymore it’smuch more prudent so uh yeah so I Idon’t know I I might have heart pangsyou know like a yearn for the for theWilder the AI of of months past you saidsomething about a wild ride wasn’t thereI think it was maybe a New York TimesReporter that I think maybe theMicrosoft chatbot started telling him hewas a bad person and it was and hedidn’t and the chatbot didn’t like himyou talk about it mimicking humanemotions I read that and that kind ofscared me yeah that’s so you know thethe training data for these um for thesechatbots is the whole of the internetyou know another term that people mayhave heard is large language models orllms um I think like anyone who’s beendeveloping machine learning modelsalways thinks that what they’re dealingwith is large so I just think it’srelatively like we we we say large andwe mean like you know trillions ofparameters now um but these largelanguage models are trained on all overthe Internet and you know when humanswrite they have an agenda you knowthey’re trying to persuade or they’retrying to like console or they’re tryingwhether you think of like commentsections or news articles or so um sothe interesting thing is that in itsearlier um incarnations as as this modelis conversing and with this New YorkTimes Reporter it’s started to adopt theperson owner of a um like a jilted loverand so this New York Times Reporter hadthis really bizarre experience which Ithink will go down in the I think itwill continue to be a reference pointforever more of kind of what it lookslike to have the the Untamed version ofthese chatbots out there but the thechatbot was saying you know divorce yourwife and he was like I’m happily marriedand the the chatbot said no it’s nottrue she doesn’t love you or you don’tlove her or it was really um reallyDynamic and I mean I think it had thewow but now you’re getting a little bitof the dark side of it too so I don’tthink it felt it didn’t feel like a puremagic trick it felt a little bit morelike something maybe maybe Sinister anddangerous the other we’ll get to this Ithink maybe a little later but as as thethe reformed lawyer that there was aninstance where where a lawyer used it towrite a legal brief and it made up casesbut Nicole I want on the dark side Iwant to Pivot to you may not be able toanswer this because it’s probablyclassified but can you talk to us aboutabout Skynet and maybe what chat GPT andartif icial general intelligence is notwe’ve talked a bit about what it is likedo we have to worry about theTerminators coming afterus yeah I mean I I think the answer isno um when I say that when I pose thatto people they’re always like well whatif this happens and what if this happensand it’s like okay um I would say as ageneral rule um I do not think AI isgoing to kill us all but I think itbrings up an interesting conversationabout who is in charge of creating theartificial intelligence so when we’retalking about g& Ai and we’re using itum a lot to for Content but askingquestions trying to get information whenwe’re talking about the umbrella ofartificial intelligence and some of thethings um that it does that gets us Ithink more on the side where you havepeople who have technical backgroundtechnical expertise they’re using verysophisticated tools and data sets and sothe answer to the question is no AI isnot going to kill us but it’s a reallygood question about whocontrols who who helps decide what isgoing to what it is going to be used forand those inputs kind of going back toYasmin talking about those largelanguage models like where are theycoming from what communities are beingused what communities are being excludedfrom those models um so no it’s notgoing to kill us but we should be verycautious about it and we should beasking a lot of questions and we shouldhave transparencybut gen gen AI is different from AGIartificial generalintelligencecorrect yes it’s a lot of acronyms umyes so I think so one of the examplesand I think Jeff you you are more of aTony Stark fan um so you were kind ofwalking me through this which I think isa good visual and I have been using forpeople um so when we talk about AGI weare talking about Tony Stark levelAvengers yes um and when we’re talkingabout J AI That’s what’s in your pocketthat is the thing that you can use onyour laptop um but outside of Tony Starkand I can’t remember how he got hismoney but outside of of him and andpeople that have pretty pretty large umor have access to large resources mostof us are not going to be in a placewhere we’re going to be dealing with AGIartificial general intelligence sogenerative AI is not self-aware even ifit’s telling this reporter to divorcehis wife yasine oh I mean you saysomething really profound which is oneof the things that people describe umone of the attributes that people sayAGI will have the general intelligenceis that it’s self-aware which is a kindof you know philosophically mindbogglingconcept um but because because of thepower of language and because there wassome of this transitive goal-seekingBehavior coming into these early chatBots from just Language online um therewere this you know it you had the senseof you were kind of like detecting thatthere was some self-awareness eventhough there wasn’t um and alreadythat’s that’s a little bit of achallenge if this you know even if wherethe tech is today you can evoke thefeeling that there’s self-awareness orlike a Humanity or personhood in thesoftware that you’re talking to you canimagine you know threats and and andharms that could come from that um theit’s one of the things that’sfascinating about the AI discussionespecially in the last year is that umthere’s so much heated debate among theforemost experts in the world on thetrajectory of this thing um so AGI isdefinitely one of those like lightningrod topics because associated with AGIis that um we will really not be incontrol um and we don’t know what wedon’t know and we’re not prepared forthat so it’s really consequential youknow what the timeline is for AGI sosome people you know some experts thinkit’s five years some experts think it’s20 years um there are far fewer expertswho say it’s never gonna it’s nevergonna hookup I I probably would be in the kind ofhope it doesn’t but but I guess we canwait and see um let’s shift a bit andtalk about like what are some thingsthat that we think and I promise at theend we’ll we’ll give folks a list ifthey want it but what are some thingsthat that we having been part of thisdebate and Yasmin having been part ofthe evolution of this that we thinkfolks should know that that they mightnot know so um I have a few but I don’twantan to I don’t want to jump in overpeople so what are throw out some youknow something you think people shouldreally be aware of and we’re as they’renot new to technology but new to thisspace you know they they picked this uplast November with with many ofus forNicole yeah I can jump in um so I thinkthe number one thing for for people whoare using G gen AI for the first time oreven trying to use it consistently isthat I have found that it’s not actuallysmarter than I am in a lot of ways umone of the things that you mentionedearlier de is like a lawyer used it andum used one of the chat box and it juststarted making up cases I have seen thatum when we talk about um some of the Iuse chat GPT but I’ve also started usingwhat’s called caude AI which says goingback to yasmin’s point about just havinglike a different type of foundation interms of what kind of information itwill do it doesn’t do the fun crazystuff but it they offer that it givescontext and um it does it in a way thatis more of a human human language andI’ve Just Seen over time that you know Iam the kind of person that tells my chatbox thank you you did a good job or whenit does a bad job I say no this is notcorrect and I’ll say what’s wrong um butI I have noticed over time it feels likeit’s trying to to say the right thingand like it has suggested some reportsfrom um some reports on Electionsecurity and voting right stuff from uhthe NAACP and the Urban League and thenI’m like I don’t know about thesereports let me look them up and then yourealize they don’t exist and so I thinkwhat like the overarching theme on thatis it can help you can give you thetools to help you get smarter and giveyou places to start but it can’t bewhere you end because you just don’t ifyou don’t have the Baseline level ofknowledge you don’t know if theinformation is right or not um and so Ithink as people are starting to use itthink about it as something that helpthat may help you become more efficientbut it can’t be your only source ofknowledge or the only place that youconsult for information that’s whatpeople are callinghallucinations yeah which I think islike an interesting thing to call it ahallucination but it it basically meansthat it’s a nice way of saying that itis incorrect or it’s not factual um butpeople should be very aware of thosebecause they do look real like you cansee where a report is cited and itsounds like the NAACP would write areport on Election security and so umthe hallucinations are not going to beas easy I think for people to identifyyou’re going to have to like actuallycheck that Source you’re going to haveto check that information you’re goingto have to have a little bit of contexteven if you’re asking you know one ofthe chat boxes to summarize somethingyou know the chat box is going to dothat very quickly but if you are goingto use this if you’re going to sharethis information with other people youreally do have to take your time and doa additional research you guys beenthoughts on things to know or more onhallucination build on what Nicole’ssaying that’s spot on and you know I wasjoking about kind of like the magictrick uh metaphor for chat GPT which Ithink was appropriate when it was so sorestricted access to it but now we’regoing to be seeing chat gen you knowautonomously generated chat all over theplace sometimes we’ll be knowingly in acontext where we’re speaking to achatbot and even then I think we’resusceptible to being misled um andsometimes we won’t know um we did somewe did some ethnographic research atjigsaw we get to kind of go and spendtime with people in their homes and likeask them about just try to understandtheir life a little bit and then therole of Technology within that um and wedid a study over the summer right whenpeople were starting to use useconversational AI um in India and the USwith genz um to just to understand howhow it’s showing up for them and thecontext that they use it for and one ofour takeaways was that it requires awhole different mental modelto understand the service that you’regetting like Nicole is like describingit uh but of course people have horissixs and so they’re going to the searchmental model which is you are giving methe correct answer from the internetwhich is wrong because that you know itover indexes on authoritativeness orpeople are going with the kind of socialmedia mental model which is um what isspeaking to me like a human and I don’tdistrust a human unless they give mereason to so you’re over indexing onauthenticity um but this is not this isneither giving you the correct answerfrom the Internet it’s not scanning theinternet giving you the right thingthat’s not its goal it might end uptelling you the right thing but that’snot the objective function and it’s alsonot a human as much as it sounds thatway and it’s it’s you’re not in a socialrelationship with this soft W so it’syou know it’s giving you plausibleanswers so I I think it’s kind of funnythat included in that plausible answerscope it will give you page referencesand actual titles of reports that the NVput out or that lawyer that actuallybuilt their whole case they built thecase and they wanted to make sure thatthere were references so they asked chatGPT to provide the references and thereferences were so plausible lookingthat the lawyer included them but ofcourse the lawyer didn’t actually go andlike follow up so um I think like areally important thing as we move intothis next basically the new realm of AIis that we um we all understand that weneed a new mental model none of ourgo-to horis STS are going to serve us umand hopefully that we’re acceleratedalong the development of that new mentalmodel which I’m hoping that you know alot of like even the even the kind ofmore Sensational headlines aboutgenerative Ai and you know a kind oflike a deceptive video or an articlethat was written written entirely by umby by by a chatbot I think all thosethings are helpful to dislodge us fromthe you know the go-to mental modelsthat we had before so it is bothauthoritatively and authentically wrongattimes but yes but and deceptive on thosefronts as well yeah so that that youknow what you talked about how it itdidn’t go search the internet becauseit’s really just in a nutshell and tellme if I get this wrong predicting whatwords it thinks should come next basedon having read the internet right and ifsomething is very welldocumented the plausible answer will bethe right answer because because there’sso many examples in the training datadata set 2 plus two equals four it willgive me four don’t talk about maththat’s actually you you know one of it akitty’s heels it should multi-leggedthing because it has many of Kitty heelsbut um you know it’s like if you saidfor example I don’t know um I’m tryingto think of like a word that likeWednesday you know how many Es does theword Wednesday have it won’t probablywon’t get that right it’s very itstruggles with it’s you know becauseit’s not a general intelligence thereyou go what a good example of the factthat we are not at artificial generalintelligence uh because you know andthese large language models they aretrained on language and so they’re goodat generating language but it’sdifficult for them to flex to othertasks you I I have some go ahead I wasgoing to jump in on the math the math isreally bad like I I tell people I’m toyour point yasine that’s just not whatit was designed for so you know if I’mlooking I’m looking to create a positionfor a cyber analyst for my company chatGPT all the conversational models theydo a really good job of like this iswhat a potential job description couldlook like this is what the duties couldlook like the qualifications but thenwhen you try to put that into a budgetit does a really great job of scouringthe internet and saying the average payis this and you drop it into a budgetand just basic math no longer works andI this is coming from a person who doesnot do math in in public because Ialways feel like I’m going to get itwrong but when you see that it strugglesto do just basic math it also puts incontext for you like this is not goingto be the answer for everything you’restill going to have to do somework so I I have a couple friends in uhone who teaches in high school one whois Professor colle and I’ve talked tothem both about this and their take onit I thought it was really fascinatingwhich you know the question is are yourstudents going to use it and the answeris yes they will and we need to acceptthat reality and it was teaching them touse it intelligently and Yasmin the wayyou talked about changing our mentalmodel about it um kind of almost likethe very early days of Wikipedia whenGod knows what was on there like youcould use it as a way to gather placesto then confirm the information but nowyou can have it you can use it to haveit write something for you as long asyou verify the accuracy independently onyour own do you think that’s a goodenough summary yeah go go go I was gonnasay so I actually used to be a collegeuh tutor and so of course my studentsalways wanted to use Wikipedia and wehad the school had a strict like do notuse Wikipedia policy but I was reallyhonest with my students I’m like if youneed to start at Wikipedia start thereget to those footnotes like you got toget to the footnotes and like that iswhere you should be that’s what youshould be researching that’s what youshould be learning from don’t useWikipedia as like the the thing that isthe end all be all do your research so Ithink actually comparing it to Wikipediais a good example but also your teacherfriends and Professor fans may feel avery different way about Wikipedia thanI do but I think for the averageAmerican or sorry the average personmost most of the time when they’relooking for general informationWikipedia does provide that like who isYasmin green and it’s like what you wantto know is these like three sentencesbut if you’re are going to write areport on Yasmin you should not copy andpaste so you know you should be going tothose footnotes and you should see whereshe’s done interviews you should look ather official bio you should look at herwebsite my Wikipedia page is 50% wrongso that’s funny that you chose me reallyI didn’t even know if Jeff had aWikipedia page I think you’re the onlyone on here that has a Wikipedia page ifyou look if you look up Jeff Greenyou’ll get the Miami millionaire who’shad some ising issues so I would takehismoney um what are youJeff I’d say the both of both of the theteachers professors their view wasthey’re training students for the futureand these kids are going to use it andit is more important to teach teach themhow to use it wisely than to try to youknow put the proverbial finger in theDyke Iagree um let me see so you were gonnatell you had some ideas for things toknow yeah well one of them I think wetouched on and this is giving away oneof my five it is don’t you know there’sthe old saying from the Cold War I’m oldI remember this is Ronald Reagan saidtrust but verify with AI if you have itsomething it is don’t trust but verifyyou you can use it to help you formwell-written good sentences but you needto make sure the facts are accurate umso and I hadn’t even I love again whatyou talked about about changing ourmindset about how we use it it’s not ahuman interaction it is not anauthoritative figure and I wonder aswe’re talking through is maybe one ofwhat people have talked abouttransparency is teaching it to be lessauthoritative might be a way to helpengage people in a way that’sconstructive um let me turn now we’re Ithink we’re about 30 minutes in let’stake some questions from the audience soto speak um and here’s the most simpleone um yasine you have young kids cankids use chat GPT what are your thoughtson thatone I mean can they it’s very easy touse if they can type can use it um Iwouldn’t let you know I think it’spretty easy if they don’t don’t let kidsdo thingsunsupervised uhbecause because it just assumes that youhave infinite time and what’s the pointof a screen if not to help you dosomething else um yeah but uh maybe youknow in a supervisor there are somepretty fun activities that that you cando I mean my my eight-year-old is intoPokemon um so one of the things that Iwas this is a good example of where chatGPT still needed work but one of thethings I I was doing over Christmas whenI wanted to show chat gbt off to herum was to ask it to to contrast the mainposter child for Pokemon which isPikachu with um Eevee was her favoritePokemon and it and it was not it was Itwas kind of conflating theircharacteristics they’re very differentlooking and obviously they havedifferent powers um but because theyappear so often together in the trainingmaterial this they they’re very readydealt with independently and so theywould you know they were kind of it waslike inverted things that it was sayingabout one versus the other so um I thinkfour jokes withsupervision um I’m all for it um theother thing that we didn’t talk about uhis that you know we talked about largelanguage models in text generation butof course generative AI creates contentin all media um images and video andaudio um and that stuff I think reallyis fun you know if you’re reading a bookwith your child and you know there’s adescription of the villain or of the youknow protagonist’s home or theirchildhood bedroom um you can actuallyput the prompts into an image generatorlike Dar and and you can keep riffing onthat you know you can keep saying andnow take the character and now they’vegone outside of the house what does thegarden look like so there’s some prettycool like wow facture experiences seenJesus flipping over tables how that cameout and someone it is Jes is literallydoing a flip above atable you know I saw it I don’t if it’sreal or not but it made me laugh so Idon’t care um are there on on on thekids can kids run into inappropriate ageinappropriate content on it or theycontrol from model to model yeah I I umI mean it’s like would you I don’t letmy eight-year-old have conversations byherself on you know with strangers whenshe’s playing her Nintendo switch so Iyou know definitely wouldn’t let my kidhave a conversation with a with achatbot because you just don’t know youknow I think openai just oh no was it ohno I’m going to I’m going tomisrepresent one of the tech companies Ithink Google just released a version oftheir chatbot that was for teens so likethat that signals the severity withwhich the tech companies are viewinginviting different you know youngerpeople and minors to interact with thiscontent so that you know children I’mthinking like single digits or you knowlike tween what pre Preen tween um Idon’t think any you know anyone in goodconscience who’s thought about it is isrecommending that that children use itby themselves Nicole you mentioned biasearlier I’m going to throw this questionto you how do developers address theissue of bias in the models and whatmeasures are in place to ensure fairnessand generated content maybe you can talka little bit about how the bias couldwork and then and take that and thenyasine I think from from your side youguys may have come across that in someof your work withchicksaw yeah so I think it’s a goodquestion question um we hit yasma hit onthis a little bit earlier so when we’retalking about these large langu languagemodels you have to get that data fromsomewhere and so um I thought it wasinteresting Jeff when you talk aboutlike the dialect the first time you usechat GPT and putting it in a dialectwhat we’ve seen over time is again theyare the the large language models arenot using the most accurate informationit’s using the most wde the the widestinformation and so it plays up a lot ofstereotypes and so you know I I have putin and there’s all type of like umscientific research to show this butlike if I say I would like to write apaper in the voice of a a black womanmost of these uh conversational chatboxes are not going to have me soundinglike this right they play up on theselike really negative stereotypes aboutblack people they play up these reallynegative stereotypes about women andthat is a result of the bias behind thethe language models and so when we talkabout what we what should we be doing Ithink it’s mostly it all comes down toaccountability and transparency so wehave to be we have to be pushing that’swhy you see there’s a lot of a lot ofthe members of Congress but also at thestate and local level people are tryingto figure out like what should we berequiring from abaseline um in terms of transparency sowe can see what language models arebeing used what information is beinginput and then also have accountabilityso if this is pushing not only falsenarratives and stereotypes what doesaccountability look like so that thisstuff can be corrected because we itmight be first impression for a lot ofpeople and some of these issues some ofwe’ve talked about some of theum um not so complex or controversialissues that people may be searching withthe conversational Bots but you’re goingto have these Bots that are going to beinteractive with people in customerservice you’re going to have them um inwhat we call government service spacesand if they have these biases where theythink that they are now talking to ablack woman so they’re going to talk ina very different way or they’re notgoing to share information in the sameway as a white male counterpart that’sreally problematic so I would deduce itdown to accountability andtransparency yasine got any thoughts onthat one yeah we we’ve had someexperience for sure of bias you know Iwas talking about when um you knoworiginally understanding the mechanicsof machine learning models and my youknow and the Child Development processand then the the vulnerabilities in eachof those um and it’s you know when youhave a young child and they’re learningto talk they um they have like littlebiases in their language partly becausethey they expect grammar to be logicalum and so they’ll say a sentence likelike you know I eated a banana andyou’re like well I got to give it to youmaybe it should be eat it’s it’s not Iate a banana um but you know that’sactually that’s you can think of that asa biased output because the the trainingdata of their life is is biased it’sbiased towards a a you know and likelogic to to grammar that isn’t there ifchildren don’t see you know like we wentto the zoo and the um the sea lionsneezed and my daughter said bless youuh but it wasn’t a it wasn’t like shesaid the sea line has a cold but sealines just need that’s just like that’show they roll they just need to likeclear their Airwaves and so they sneezeit’s not because it it had a cold but ofcourse every time she’d seen a sneezeshe thought you know it had beenassociated with somebody having a coldso those are really innocuous examplesbut you see that if you don’t have goodquality training data then or you knowif your training do data doesn’t reflectyour expectations of a child then um youwill get output that that that you thenhave to fix later and so with thetraining with the machine learningmodels what’s it could be that the youknow there are things that are happeningduring the actual training process thatare introducing the bias so um or itcould be that the training data itselfis biased so you know when we were oneof our machine learning projects was todevelop um an AI that can detecttoxicity in language and we trained iton a bunch of comments and then when wewere ready to laun and the the AI isvery WID widely used it’s the mostpopular free API for most Publishers andplatforms like that the large in in inum in the 18 countries that that are inlike use it like New York Times uses italpes leemon Wall Street Journal Redditwikkipedia they all use this API codeperspective when we launched it werealized that if you if you put in apositive comment that mentions beingMuslim the model thought that there wasa negative association with that itthought that it was toxic why did itthink that a positive comment aboutMuslims were toxic because by and largethe the comments that it had beentrained on the mention the word Muslimhad a negative association and we saidbefore like these are patterns that it’slearning so it’s intuiting that the wordMuslim might must itself have a negativeassociation so uh you know to Nicole’spoint across actually a lot of identityterms where people are persecuted ormarginalized and where there’s a lot oftoxicity online you have to go back andfixthe toxicity issue in the training dataum so that the model itself doesn’t havethis systemic bias and of course ifyou’re trying to build an a you knowsome algorithms that are designed topromote healthy speech fine toxicity andremove it the fact that you might havethese false positives that you might befalsely accusing you know comments thatare not toxic but mention you knowtargeted identity terms as being asbeing toxic is a really huge problem umso so we really do have our work cut outfor us and Nicole referenced this rightat the front if you you know you youknow the person who probably found theMuslim the first Muslim um you know theperson the person the person to Firstdiscover the issue with bias around theterm Muslim was probably a Muslim personyou know people who went in and theythey looked for you know their ownfaiths and their own names and their ownso if you don’t have a representativegroup of people who are developing thetech and doing the quality assurance youwon’t find this stuff to fix it um so Ithink you know I’d probably yeah addrepresentation to the to theaccountability and transparency listthat Nicole gave but if you’re trainingthe model to detect toxic language youneed to give it toxic content which iskind of you wouldn’t think that youwould want to introduce that but I guessif you want it to help you at the lowerlevel it needs to know it so at thehigher level it can stop it from beingpublished it’s it’s as weird that’s areally profound Point um that we I stoleit from someoneso I’m surprised Jeff no that you youdon’t actually necessarily want tosanitize your foundational models fromeverything that you think is bad becausethat will have unintended consequenceconsequences later in that for exampleyou won’t be able to have moderationalgorithms that they’re trying to findbad stuff in comments and remove them soone other question I love this is um isAI going to take myjob I’ll jump in and then I mean theanswer is I maybe it depends what yourjob is um ideally it it changes jobs butthe reality is that it will change whenwe’re talking about the generative AI itwill definitely change some of thesupport work the customer service a lotof it it hopefully will make it itbetter with fewer people in the cybersecurity realm it’s certainly gonnaum lower the number of humans needed todo the same amount of security but wedon’t have enough humans there alreadyum the question of whether it’s going toto be a sentient robot doing your job isI think off in the distance but I thinkit would be wrong for anyone anyone totries to tell you with absolutecertainty that it that it won’t isprobably wrong but we still don’t knowexactly how this is going to play outbut it will change the workforce that’sthat’s how I see it and and we need toconfront that and and understand it andtry to make the the impacts as minimalas possible on on our society in termsof ofemployment yeah one of the things that’sinteresting about this AI Revolution andI think all techn technologicalrevolutions before that is that it lookslike it’s disproportionately going tohit the moreeducated umthat’s that’s really interesting so youknow one of the things one of theimpressive things that chat Bots can dois they can write code I do math butthey can write code because there’s somany examples of code written out thereum very large repositories computer codeyeah computer code so it’s just funnythatprogrammers have developed a technologywhich are these chat Bots that willreplace their own labor I mean I don’tyou know I don’t like anyone beingdisplaced from from their work I thinkhas huge societ and implications there’sjust something things perversely sweetabout having the people who are theinnovators have to suffer some of thedisplacement of the Innovation well Ican give an example that I think mostpeople will like going back if youactually train a model only on actuallegal cases and case law it couldprobably start writing pretty good legalbriefs so you’re going to it it’s goingto take a potentially a lot of the thepeople out of you we won’t I guess we’renot going to get to the the ideal nolawyer world but you’re going to havefar fewer and and those expensiveeducations are going to be I’m glad myloans are paid off are going to be a lotless than uh useful potentially wisemaybe we’ll find other um Nicole do youhave thoughts on the work questionI do I mean as someone who has not paidoff her student loans from law school umI I hope it does not replace me beforeI’m able to do that but I think itbrings up an interesting questionbecause um you know my my dad is like aUnited Steel worker and so they most ofthe time to yasma point when we talkedabout Innovation it’s people that aredoing those kind of jobs that have beenum adversely impacted the most um thosejobs that require more manual skillsand so when we’re talking about likewhat generative AI specifically what itcan do it does start to feel like it isgoing to to impact not just um whatpeople call the White Collar professionsbut also the creative profession so youknow we’re talking about Dolly and howit can create images um but also youknow there are people who who writethese scripts we mentioned earlier thewriter stri they write these scriptsthey do all of these creative Pursuitsbut I I hold out a little bit of Hopebecause and this is just things thatI’ve been reading but also my experienceit still doesn’t get it quite right likeeven with doly I I feel like it it wouldprobably make a graphic designer not gocrazy trying to do something with mebecause I could show up and say this iswhat I like and this is what I don’tlike and this is what I put in but evenwhen I play with that like trying tothink about a new logo it just doesn’tget it exactly right and then don’t knowall the prompts because I’m not agraphic designer so there’s only so muchI can tell it to do to like it to makeit do the thing um I think we’ll seesimilar with lawyers so I do think youknow for those JuniorAssociates um straight out of law schooland early in their career I think it isgoing to bedisproportionately um or it’s going todisproportionately impact them but Ithink about the kind of analysis that wedo and like how you’re you you know Jeffand I work together own legislation andwe worked onpolicy sometimes you’re coming up withsomething new so there is no languagemodel like we’re trying something thathasn’t been done especially on Tech andcyber policy this hasn’t been doneanywhere so it’s going to be really hardfor any type of generative tool to comeup with the same type of analysis or thesame way to to address a problem so Ithink what we’ve been seeing and you seeit especially a lot with um communitiesof color and marginalized communitieswhat I hope we will be able in a placeto do is is not do what we’ve done inthe past which is like go fast breakthings hurt people and then sit back andsay oh wait a minute maybe we shouldhave done this in a different way I hopebecause it is targeting a differentdemographic that people will slow down alittle bit and think who are we leavingbehind and how do we make sure we bringthem forward with us so one lastquestion try to do quick answer and thenbecause we’re running up on time um butI do want to get this one in and thenwe’ll go to our five recommendations umHow do you see generative AI being usedand playing a role in the 2024 electionsand I think we should talk globallybecause I think it’s 40% or more of ofof people around the world will have anelection in 2024 so quick thoughts on onwhere you see this playing um Yasminmaybe thenNicole I mean I I do think we’ll seeimages and videos that are syntheticallycreated the interesting thing is that uhthe ones that we’ve seen so far havebeen uh have been clearly well I thinkthey’re clearly they’ve been describedor they’ve been reported on assynthetically created um I remembersomething that I remember when we didsome research with some people who werefalse flag like disinformation people sothey thought every time there was aschool shooting or something theythought it was a government conspiracyto ultimately take take away your gunsbut they thought they were crisis actorsand it was an interesting study um foranother podcast uh but uh one of thethings we showed them something that hadbeen going going viral in theirCommunity we showed them um something uhan image of one of the victims of theshooting and we showed them that it wasa doctored image and they the likeUniversity they said oh well even ifthat’s fake that’s emblematic ofsomething that that person would havedone so like we’re really you knowthere’s so much motivated reasoning nowaround the topics that uh that we’remost concerned about for disinformationso you know the um there was a study ofDary um that’s that’s open ai’s imagegenerator where academics put in promptsso if you say show me someone stealingan election it doesn’t generate an imagebut if you say show me somebody stuffingpieces of paper into a you know a tinbox and wearing a name tag that showstheir official and you know then you cancreate a bunch of images that look likethey’re supporting the stop for thesteel narrative um and even if they looksynthetic they are somehow going to beuseful for propagating a really harmfulyou know election disinformation themeuh so I do expect we’ll see it and uhand I and we really it’s the Rubbermeets the road for detection andProtection Systems all the platforms areaware that it’s coming and now we haveto see how how well they’ll do atspotting it and removing making surethat they’re not helping with vadityNicole yeah I will just plus one to whatYasmin said and and what I would add isI you know AI just amplifies so itamplifies good it amplifies bad one ofthe things that I’m most concerned aboutthe 2024 election is how it’s going toamplify um these messages of what I callor what people call voter depression soyou have communities that are over thecourse of an election they’reoverwhelmingly targeted already withyour vote doesn’t matter nothing’s goingto change democracy isn’t working and AIis going to allow that to become a muchmorepersistent um targeting but also justvery tailored to like what is going onwhat is happening at the local levelwhat is happening at the state level andthat really worries me because when youtalk about voter depression that’s thedifference between someone going to votewho is reg that’s the difference betweensomeone registering to vote and to havecommunities especially marginalizedcommunities that are already Targetof these campaigns like Yasmin wassaying now you’re they’re just going tobe targeted all year long like they it’salready started right like we’ve alreadyseen um some generative AI tools thatwere used to create these like to reallyjust exploit the tensions And theemotions of black lives matter where itwas a AI generated image of a black manbeing shot in the car and you know thatI looked at that image and it alreadyhad the mark on it because this was a apart of a journalist investigation thatit was artificial intelligence and itstill even seeing that it still made mehad a have a very visible and emotionalreaction and so for people who don’teven know who don’t have that tag orthey get the they find out later itdoesn’t really change the emotion thatyou feel in that moment and it doesn’tit doesn’t change the depression thatyou may feel or the disillusionment thatyou may feel about yourday-to-day you know I I will be a littlecontrarian in for perverse I thinkreasons I think we are so nowconditioned to what is the word you usethe description of asmin of of of ofreasoning I loved it or or why peopleare motivated reasoning motivatedreasoning is that um we are past thepoint where it doesn’t even have to beI’m not worried about a good fakebecause a bad fake doesn’t matter and II so I I think it is a problem but I’mnot sure it’s going to be as big anaddition problem because because we’realready in a bad place in that regard umso we’re going to wrap up but firstlet’s do let’s do as promised in thetitle um Yasmin if you have five thingsthat you would leave people and thenNicole if yours are different you can orthe same you can just drop down and I’mhoping you guys hit all of mine so Idon’t have any at the end so jasine I’mgonna test the the listeners of viewersability to do math because I’m going tosay a number of them it’s not going tobe five um so one is to leave you withis to emphasize that um we need a newmental model the second is that the bestway for us to develop that and the bestway for for everyone to feel empoweredand informed about the new AI Paradigmthat we’re entering into is to learn bydoing um so go and use you know Bardwhich is Google’s version of a chatbotthat it did a horrible job of pluggingor chat GPT which is opening eyesversion of chatbot that’s it’s a lotmore popular um and the image generatorsand experiment for yourself and seewhere they fall short and see whereyou’re impressed and see the limits umand I think if you understand that weneed a new mental model uh and you’rereally exploring the boundaries of it uhthen I feel it’s I can feel that I cansay this responsibly which is to askpeople to go and look for the Magic inAI great Nicoleso I actually use generative AI to comeup with my five um because glad we gotI’m glad we got this far before someonesaid so I mean I was I tried to do it bymyself I tried to do it by myself and Icame up with a list but then I wanted anacronym um and like this is kind ofwhere the conversational um chat box iskind of shine so the acronym is paintand I would encourage people to rememberto paint and to also try to paint paintstands for and all of these are thingsthat generative AI can do and how youcan interact with it um the P is forperforms repetitive task the a isautomates jobs the I is identifiespatterns the N is needs human needshuman oversight and the t istransparency is important so for methose are the things like it does dorepetitive tasks and um even when wetalk about automating jobs some jobs wecan can benefit from automation we justneed to make sure that we don’t leavepeople behind and then identifying thepatterns um of course I think that’sreally important but we hit on the pointof it it matters what the what theinputs of those patterns are um beforewe accept the outputs and of course itneeds human oversight I can’t state thatenough um and then transparency we haveto be able to know how these things arebeing built um we have to be and that’snot something that we should learn onthe end on the back end all the waythrough the process we need fulltransparency on what’s being done so Iwill give my 51 I gave before which isdo not trust and verify make sure youyou you look into what it tells you umkind of echoing as and play around withit learn how to use it what it can’t usedon’t be afraid to use it um the thirdone is don’t forget everything you knowwhen it comes to life or security orcyber security all the same rules thatyou’ve applied trust your intuition umwhen you’re playing with it the fourthis a phrase I got from someone at aconference last week which is panicresponsibly um it’s okay if if itworries you because something is doingyou hear about it but take a breathdon’t overreact to the good and the badum and then finally is is just have funwith it and so that I’m ending withyasine you know find the magic in it umso Yasmin Nicole I really appreciate thetime um I want to turn back again andthank 92 y for making this happen todayit’s been a great conversation again askeveryone join us at the uh the Aspencyber Summit AT 92n on November 15 infoat Aspen cybers summit.org and if youwant to get in at a super discountedrate um you can use the code 92n andit’s $100 to register a great day and asI said everything from from AI to thebird drone conspiracy to the election sohopefully we can see you there inNovember um but thanks again for yourtime thanks gasmin andcall
In recent months, the rapid rise of generative AI (GenAI) technology has captured widespread attention, dominating news headlines and sparking public interest. What should you be aware of when using these new tools? Are there downsides to them?
In this session, 92NY and Aspen Digital dive deeper into the dos and don’ts of GenAI, and what this transformative technology means for you, with Jigsaw CEO Yasmin Green and Aspen Digital Senior Advisor Nicole Tisdale.
{"includes":[{"object":"taxonomy","value":"134"}],"excludes":[{"object":"page","value":"203254"},{"object":"type","value":"callout"},{"object":"type","value":"form"},{"object":"type","value":"page"},{"object":"type","value":"article"},{"object":"type","value":"company"},{"object":"type","value":"person"},{"object":"type","value":"press"},{"object":"type","value":"report"},{"object":"type","value":"workstream"}],"order":[],"meta":"","rules":[],"property":"","details":["title"],"title":"Browse More Events","description":"","columns":2,"total":4,"filters":[],"filtering":[],"abilities":[],"action":"swipe","buttons":[],"pagination":[],"search":"","className":"random","sorts":[]}
Facebook
Twitter
LinkedIn
Email
