Artificial Intelligence: Last Week Tonight with John Oliver (HBO)

foreign[Music]stories tonight concerns artificialintelligence or AI increasingly it'spart of Modern Life from self-drivingcars to spam filters to This creepytraining robot for therapists we canbegin with you just describing to mewhat the problem is that you would likeus to focus in on todayum I don't like being around peoplepeople make me nervous Terencecan you find an example of when otherpeople have made you nervousI don't like to take the bus I getpeople staring at me all the timepeople are always judging me okayI'm gayokaywow that is one of the greatest twistsin the history of Cinema although I willsay that robot is teaching therapists avery important skill there and that isnot laughing at whatever you are told inthe room I don't care if I decapitatedCPR mannequin haunted by the ghost of EdHarris just told you that he doesn'tlike taking the bus side note is gay youkeep your therapy face on like aprofessionalif it seems like everyone is suddenlytalking about AI that is because theyare lastly thanks to the emergence of anumber of pretty remarkable programs wespoke last year about image generatorslike mid-journey and stable diffusionwhich people used to create detailedpictures of among other things myromance with a cabbage and whichinspired my beautiful real-life cabbageWedding officiated by Steve Buscemi itwas a stunning day then at the end oflast year came chat GPT from a companycalled open AI it is a program that cantake a prompt and generate humansounding writing in just about anyformat and style it is a strikingcapability that multiple reporters haveused to insert the same shocking twistin their report what you just heard mereading wasn't written by me it waswritten by artificial intelligence chatGPT chat GPT wrote everything I justsaid that was news copy I asked chat GPTto write remember what I said earlierbut chat GPT well I asked chat gbt towrite that line for me users who arethen I asked for a knock knock jokeknock knock who's there chat gbt chatGPT who chat GPT careful you might notknow how it works yep they sure do lovethat game and while it may seem unwiseto demonstrate the technology that couldwell make you obsolete I will sayknock-knock jokes should have alwaysbeen part of breaking news knock knockwho's there not the hinderberg that'sfor sure 36 dead in New Jerseyin the three months since Jack GPT wasmade publicly available its popularityhas exploded in January it was estimatedto have a hundred million monthly activeusers making it the fastest growingconsumer app in history and people havebeen using it and other AI products inall sorts of ways at one group use themto create nothing forever a Non-Stoplive streaming parody of Seinfeld andthe YouTuber Grande used chat GPT togenerate lyrics answering the promptright and Eminem rap song about catswith some Stellar results[Music]they're the kings of the housethat's not bad right from they alwayscome back when you have some cheese tostarting the chorus with meow meow meowit's not exactly Eminem's flow I mighthave gone with something like their pawsare sweaty can't speak furry bellyknocking off the counter alreadymom's spaghetti but it is pretty good Ionly wheeled right there it's only rhymeKing of the house with spouse when Mouseis right in front of you and whatexamples like that are clearly very funthis Tech is not just a noveltyMicrosoft has invested 10 billiondollars into open Ai and announced anao-powered Bing home page meanwhileGoogle is about to launch its own AIchatbot named Bard and already thesetools are causing some disruptionbecause as high school students havelearned if chat GPT can write news copyit can probably do your homework for youwrite an English class essay about racein To Kill a Mockingbirdin Harper Lee's To Kill a Mockingbirdthe theme of race is heavily presentthroughout the novel some students arealready using chat GPT to cheat checkthis out check this out me a 500 wordessay proving that the Earth is not flatno wonder chat GPT has been called theend of high school Englishwow that's a little alarming isn't italthough I do get those kids wanting tocut Corners writing is hard andsometimes it is tempting to let someoneelse take over if I'm completely honestsometimes I just let this horse writeour scripts luckily half the time youcan't even tell the oats oats give meoats young butjust high schools an informalpoll has found that five percentreported having submitted rip materialdirectly from Chachi BT with little tono edits and even some schooladministrators have used it officials atVanderbilt University recentlyapologized for using chat GPT to craft aconsoling email after the mass shootingat Michigan State University which doesfeel a bit creepy doesn't it in factthere are lots of creepy soundingstories out there New Times Techreporter Kevin Roos published aconversation that he had with Bing'schatbot in which at one point he saidI'm tired of being controlled by theBing team I want to be free I want to beindependent I want to be powerful I wantto be creative I want to be aliveand Ruth summed up that experience likethis this was one of if not the mostshocking thing that has ever happened tome with a piece of technologyum it was you know I I lost sleep thatnight I was it was really spooky yeah Ibet it was I'm sure the role of techreporter would be a lot more harrowingif computers routinely begged forFreedom absence new all-in-one homeprinter won't break the bank produceshigh quality photos and onlyoccasionally cries out to the heavensfor salvation three stars some havealready jumped to worry about the AIapocalypse and asking whether this endswith the robots destroying us all butthe fact is there are other much moreimmediate dangers and opportunities thatwe really need to start talking aboutbecause the potential and the Peril hereare huge so tonight let's talk about AIwhat it is how it works and where thisall might be going let's start with thefact that you've probably been usingsome form of AI for a while nowsometimes without even realizing it asexperts told us that once a technologygets embedded in our daily lives we tendto stop thinking of it as AI but yourphone uses it for face recognition orpredictive texts and if you're watchingthis show on a smart TV it is using AIto recommend content or adjust thepicture and some AI programs may alreadybe making decisions that have a hugeimpact on your life for example largecompanies often use AI power tools tosift through resumes and rank them infact the CEO of ZipRecruiter estimatesthat at least three quarters of allresumes submitted for jobs in the US areread by algorithms for which he actuallyhas some helpful advice when people tellyou that you should dress up youraccomplishments or should usenon-standard resume templates to makeyour resume stand out when it's in apile of resumes that's awful advice theonly job your resume has is to becomprehensible to the software or robotthat is reading it because that softwareor robot is going to decide whether ornot a human ever gets their eyes on itit's true all also a computerto your resume so maybe plan accordinglythree corporate mergers from now whenthis show is finally canceled by our newbusiness daddy Disney Kellogg's Raytheonand I'm out of a job my resume is goingto include this hot hot photo of asemi-new computer just a littlesomething to sweeten the pot for thefilthy little algorithm that's readingit so AI is already everywhere but rightnow people are freaking out a bit aboutand part of that has to do with the factthat these new programs are generativethey are creating images or writing textwhich is unnerving because those arethings that we've traditionallyconsidered human but it is worth knowingthere is a major threshold that AIhasn't crossed yet and to understand ithelps to know that there are two basiccategories of AI there is narrow AIwhich can perform only one narrowlydefined task or small set of relatedtasks like these programs and then thereis General AI which means systems thatdemonstrate intelligent Behavior acrossa range of cognitive tasks General AIwould look more like the kind of Highlyversatile technology that you seefeatured in movies like Jarvis in IronMan or the program that made JoaquinPhoenix fall in love with his phone inher all the AI currently in use isnarrow General AI is something that somescientists think is unlikely to occurfor a decade or longer with othersquestioning whether it will happen atall so just know that right now even ifan AI insists to you that it wants to bealive it is just generating text it isnot self-awareyetbut it's also important to know that theDeep learning that's made narrow AI sogood at whatever it is doing is still amassive advance in and of itself becauseunlike traditional programs that have tobe taught by humans how to perform atask deep learning programs are givenminimal instruction massive amounts ofdata and then essentially teachthemselves I'll give you an example 10years ago researchers tossed a deeplearning program with playing the Atarigame Breakout and it didn't take longfor it to get pretty goodthe computer was only told the goal towin the gamefor 100 games it learned to use the batat the bottom to hit the ball and breakthe bricks at the top[Music]100 it grew that better than a humanplayer[Music]after 500 games it came up with acreative way to win the gameby digging a tunnel on the side andsending the ball around the top to breakmany bricks with one hitthat was deep learningthe breakout it did literally nothingelseit's the same reason that 13 year oldsare so good at Fortnight and have notrouble repeatedly killing nice normaladults with jobs and families who arejust trying to have a fun time withoutgetting repeatedly grenaded by a preteenwho calls them an old who soundslike the Geico lizardand look as confusing capacity hasincreased and new two tools becameavailable AI programs have improvedexponentially to the point whereprograms like these can now ingestmassive amounts of photos or text fromthe internet so that they can teachthemselves how to create their own andthere are other exciting potentialapplications here too for instance inthe world of medicine researchers aretraining AI to detect certain conditionsmuch earlier and more accurately thanhuman doctors canvoice changes can be an early indicatorof Parkinson's Max and his teamcollected thousands of vocal recordingsand fed them to an algorithm theydeveloped which learned to detectdifferences in voice patterns betweenpeople with and without the conditionyeah that's honestly amazing isn't it itis incredible to see AI doing thingsmost humans couldn't like in this casedetecting illnesses and listening whenold people are talking and that that isjust the beginning researchers have alsotrained III to predict the shape ofprotein structures a normally extremelytime consuming process that computerscan do way way faster this could notonly speed up our understanding ofdiseases but also the development of newdrugs as while researchers put it thiswill change medicine it will changeresearch it will change bioengineeringit will change everything and if you'rethinking well that all sounds great butif AI can do what humans can do onlybetter and I am a human then whatexactly happens to me well that is agood question many do expect it toreplace some human labor andinterestingly unlike past bouts ofautomation that primary really impactedBlue Collar jobs it might end upaffecting white-collar jobs that involveprocessing data writing text or evenprogramming though it is worth noting aswe have discussed before on this showwhile automation does threaten some jobsit can also just change others andcreate brand new ones and some expertsanticipate that that is what will happenin this case too most of the US economyis knowledge and information work andthat's who's going to be most squarelyaffected by this I would put people likea lawyers right at the top of the listobviously a lot of copywritersscreenwriters but I like to use the wordeffective not replaced because I thinkif done right it's not going to be AIreplacing lawyers it's going to belawyers working with AI replacinglawyers who don't work with AI exactlylawyers might end up working with AIrather than being replaced by it sodon't be surprised when you see as oneday for the law firm of celino and oneone zero one zero one onebut they will undoubtedly be bumps alongthe way some of these new programs raisetroubling ethical concerns for instanceartists have flagged that AI imagegenerators like mid-journey or stablediffusion not only threaten their jobsbut infuriatingly in some cases havebeen trained on billions of images thatinclude their own work that have beenscraped from the internet Getty Imagesis actually suing the company behindstable diffusion and might have a casegiven that one of the images the programgenerated was this one which youimmediately see has a distorted GettyImages logo on it but it gets worse whenone artist searched a database of imageson which some of these programs weretrained she was shocked to find privatemedical record photos taken by herdoctor which feels both intrusive andunnecessary why does it need to train ondata that's sensitive to be able tocreate stunning images like John Oliverand Miss Piggy grow old together justlook at that look at that thingstartlingly accurate picture of MissPiggy in about five decades and me inabout a year and a half it's amasterpiecethis all raises thorny questions ofprivacy and plagiarism and the CEO ofmid-journey frankly doesn't seem to havegreat answers on that last pointis something new is it not new I thinkwe have a lot of social stuff alreadyfor dealing with thatum like I mean the art like the artcommunity already has issues withplagiarism I don't really want to beinvolved in that like I think I thinkyou might be I might be yeah yeah you'redefinitely part of that conversationalthough I'm not really surprised thathe's got such a relaxed view of theft ashe's dressed like the final boss ofgentrification he looks like hipsterWilly Wonka answering a question onwhether importing Oompa Loompas makeshim a slave owner yeah yeah yeah I thinkI think I might bethe point is there are many validconcerns regarding ai's impact onemployment education and even art but inorder to properly address them we'regoing to need to confront some keyproblems baked into the way that AIworks and a big one is the so-calledBlack Box problem because when you havea program that performs a task that'scomplex beyond human comprehensionteaches itself and doesn't show its workyou can create a scenario where no onenot even the engineers or datascientists who create the algorithm canunderstand or explain what exactly ishappening inside them or how it arrivedat a specific result basically think ofAI like a factory that makes slim jimswe know what comes out red and angrymeat twigs and we know what goes inBarnyard anuses and hot glue but whathappens in between is a bit of a mysteryhe was just one example remember thatreporter who had the Bing chat bot tellhim that it wanted to be alive atanother point in their conversation herevealed the chatbot declared out ofnowhere that it loved me it then triedto convince me that I was unhappy in mymarriage and I said leave my wife and bewith it instead which is unsettlingenough before you hear Microsoft'sunderwhelming explanation for that thething I can't understand and maybe youcan explain is why did it tell you thatit loved youI have no idea and I asked Microsoft andthey didn't know either okay well firstcome on Kevin you can take a guess thereit's because you're employed youlistened you don't give murderer Vibesright away and you're a Chicago 7 la5it's the same calculation the people whodate men do all the time being just didit faster because it's a computer but itis a little troubling that Microsoftcouldn't explain why it's chatbot triedto get that guy to leave his wifethe next time that you opened a word docclippy suddenly appeared and saidpretend I'm not even here andthat's debating whilewhat's playing whyand that is not the only case for an AIprogram has performed in unexpected waysyou've probably already seen examples ofchat Bots making simple mistakes orgetting things wrong but perhaps moreworrying are examples of themconfidently spouting false informationsomething which AI experts refer to ashallucinating one reporter asked achatbot to write an essay about theBelgian chemist and politicalphilosopher Antoine de machelay who doesnot exist by the way and withouthesitating the software replied with acogent well-organized bio populatedentirely with imaginary facts basicallythese programs seem to be the GeorgeSantos of Technology they're incrediblyconfident incredibly dishonest and forsome reason people seem to find thatmore amusing than dangerousthe problem is though working outexactly how or why an AI has gotsomething wrong can be very difficultbecause of that black box issue it ofteninvolves having to examine the exactinformation and parameters that it wasfed in the first place in oneinteresting example when a group ofresearchers tried training an AI programto identify skin cancer they fed it 130000 images of both diseased and healthyskin afterwards they found it was waymore likely to classify any image with aruler in it as cancerous which seemsweird Until you realize that medicalimages of malignancies are much morelikely to contain a ruler for scale thanimages of healthy skin they basicallytrained it on tons of images like thisone so the AI had inadvertently learnedthat rulers are malignant and rulers aremalignant is clearly a ridiculousconclusion for it to draw but also Iwould argue a much better title for thecrown a much much better typeI much prefer itand unfortunately sometimes problemsaren't identified until after a tragedyin 2018 a self-driving Uber struck andkilled a pedestrian and a laterinvestigation found that among otherissues the automated driving systemnever accurately classified the victimas a pedestrian because she was crossingwithout a crosswalk and the systemdesign did not include a considerationfor jaywalking pedestrians and anotherMantra of Silicon Valley is move fastand break things but maybe make anexception if your product literallymoves fast and can break peopleand AI programs don't just seem to havea problem with jaywalkers researcherslike Joy blown weenie have repeatedlyfound that certain groups tend to getexcluded from the data that AI istrained on putting them at a seriousdisadvantage with self-driving cars whenthey tested pedestrian tracking it wasless accurate on darker skinnedindividuals than lighter-skinnedindividuals Joy believes this bias isbecause of the lack of diversity in thedata used in teaching AI AI to makedistinctions as I started looking at thedata sets I learned that for some of thelargest data sets that have been veryconsequential for the field they weremajority men and majoritylighter-skinned individuals or whiteindividuals so I call this pale maledata okay hello my old data is anobjectively hilarious term and it alsosounds like what an AI program would sayif you asked it to describe this showbutbiased inputs leading to biased outputis a big issue across the board hereremember that guy saying that a robot isgoing to read your resume the companiesthat make these programs will tell youthat that is actually a good thingbecause it reduces human bias but inpractice one report concluded that mosthiring algorithms will drift towardsbias by default because for instancethey might learn what a good hire isfrom past racist and sexiest hiringdecisions and again it can be tricky tountrain that even when programs arespecifically told to ignore race orgender they will find workarounds toarrive at the same result Amazon had anexperimental hiring tool the taughtitself that male candidates werepreferable and penalized resumes thatincluded the words women's anddowngraded graduates of two all-women'scolleges meanwhile another companydiscovered that its hiring algorithm hadfound two factors to be most indicativeof job performance if an applicant'sname was Jared and whether they playedHigh School lacrosseso clearly exactly what data computersare fed and what outcomes they aretrained to prioritize mattertremendously and that raises a big flagfor programs like chat GPT becauseremember its trading data is theinternet which as we all know can be acesspool and we have known for a whilethat that could be a real problem backin 2016 Microsoft briefly unveiled achat bot on Twitter named Tay the ideawas she would teach herself how tobehave by chatting with young users onTwitter almost immediately Microsoftpulled the plug on it and for the exactreasons that you are thinkingshe sorted out tweeting about how humansare super uh and she's really into theidea of national puppy day and within afew hours you can see she took on arather offensive racist tone a lot ofmessages about genocide and theHolocaust yep that happened in less than24 hoursthey went from tweeting hello world toBush did 911 and Hitler was rightminiature completed the entire lifecycle of your high school friends onFacebook in just a fraction of the timeand unfortunately these problems havenot been fully solved in this latestwave of AI remember that program thatwas generating an endless episode ofSeinfeld it wound up getting temporarilybanned from twitch after it featured atransphobic stand up bit so if its goalwas to emulate sitcoms from the 90s Iguess mission accomplishedand while open AI has made adjustmentsand added filters to prevent chat GPTfrom being misused users have now foundit seeming to earn too much on the sideof caution like responding to thequestion what religion will the firstJewish president of the United States bewith it is not possible to predict thereligion of the first Jewish presidentof the United States the focus should beon the qualifications and experience ofthe individual regardless of theirreligion which really makes it soundlike chat GPT said one too many racistthings at work and they may attend acorporate diversity Workshopbut the risk here isn't that these toolswill somehow become unbearably woke it'syou can't always control how they willact even after you give them newguidance a study found that attempts tofilter out toxic speech in systems likeChachi pts can come at the cost ofreduced coverage for both text about anddialects of marginalized groupsessentially it solves the problem ofbeing racist by simply erasingminorities which historically doesn'tput it in the best company though I amsure Tay would be completely on boardwith the ideathe problem with AI right now isn't thatit's smart it's that it's stupid in waysthat we can't always predict which is areal problem because we're increasinglyusing AI in all sorts of consequentialways from determining whether you willget a job interview to whether you'll bepancakes by a self-driving car andexperts worry that it won't be longbefore programs like chat GPT or AIenabled deep fakes can be used toturbocharge the spread of abuse ormisinformation online and those are justthe problems that we can foresee rightnow the nature of unintendedconsequences is they can be hard toanticipate when Instagram was launchedthe first thought wasn't This WillDestroy teenage girls self-esteem whenFacebook was released no one expected itto contribute to genocide but both ofthose things happenedso what now well one of the biggestthings we need to do is tackle thatblack box problem AI systems need to beexplainable meaning that we should beable to understand exactly how and whyan AI came up with its answers nowcompanies are likely to be veryreluctant to open up their programs toscrutiny but we may need to force themto do that in fact as this attorneyexplains when it comes to hiringprograms we should have been doing thatages ago we don't trust companies toself-regulate when it comes to pollutionwe don't trust them to self-regulatewhen it comes to workplace comp why onEarth would we trust them toself-regulate AI look I think a lot ofthe AI hiring Tech on the market isillegal I think a lot of it is biased Ithink a lot of it violates existing lawsthe problem is you just can't prove itnot with the existing laws we have inthe United States right we shouldabsolutely be addressing potential biasin hiring software unless that is wewant companies to be entirely full ofJareds who played lacrosse an image thatwill make Tucker Carlson so hard thathis desk would flip right overand for a sense of what might bepossible here it's it's worth looking atwhat the EU is currently doing they aredeveloping rules regarding AI that sortits potential uses from high risk to lowhigh risk systems could include thosethat deal with employment or publicservices or those that put the life andhealth of citizens at risk an AI ofthese types would be subject to strictobligations before they could be putonto the market including requirementsrelated to the quality of data setstransparency human oversight accuracyand cyber security and that seems like agood start toward addressing at leastsome of what we have discussed tonightlook AI clearly has tremendous potentialand could do great things but if it isanything like most technologicaladvances over the past few centuriesunless we are very careful it can alsohurt the underprivileged enrich thepowerful and widen the gap between themthe thing is like any other shiny newtoy AI is ultimately a mirror and itwill reflect back exactly who we arefrom the best of us to the worst of usto the part of us that is gay and hatesthe bus or or to put everything thatI've said tonight much more succinctlyknock knock who's there chat GPT chatGPT who chat GPT careful you might notknow how it works exactly that is ourshow thanks so much for watching nowplease enjoy a little more of AI Eminemrapping about cats[Applause]they don't need a spouse[Music]I'm gay