Written Testimony of Oren Etzioni CEO Allen Institute for Artificial Intelligence Professor of Computer Science University of Washington Hearing on Game Changers Artificial Intelligence Part I before the United States the House Committee on Oversight and Government Reform Subcommittee on Information Technology February 14 2018 Good morning Chairman Hurd and Ranking Member Kelly and distinguished Members of the Committee Thank you for the opportunity to speak with you today about the nature of Arti cial Intelligence and role of the federal government My name is Oren Etzionithe Allen Institute for Arti cial Intelligence A12 Founded in 2014 A12 is a non-pro t research institute whose mission is to improve lives by conducting high-impact research and engineering in the eld of AI all for the common good The goal of my brief remarks today is to help demystify Al and to cut through a lot of the hype on the subject What is Al Let's not confuse science with science fiction Superintelligence is over-hyped we are building narrow systems Distinguish intelligence and autonomy Regarding regulation 1 2 3 We can t afford to stifle innovation due to global competition Regulating Al itself is challenging as it s fast moving and is difficult to distinguish from software We can and should regulate Al only to prevent a patchwork of state rules as is the case with self-driving cars Principles 1 2 Al should have an off switch Al should associate clear responsibility if not liability with a person it s a tool not a mask to hide Al did it is not an excuse Al should identify itself think of the case of bots of fake porn Al should respect think of Al barbie talking to our kids Al should not exacerbate bias in data Data is historical and yet Al is predictive Let me start by setting things up Elon Musk the tech entrepreneur has told us that with Al we're summoning the demon that Al represents an existential threat to the human race We see headlines in the newspapers all the time Here s one from Newsweek Artificial intelligence is coming and it could wipe us out Is At really poised to wipe us out The famous roboticist Rod Brooks said if you're worried about the terminator just close your door Perhaps these robots are not as threatening as they're made out to be Andrew Ng from Stanford founder of Coursera said that working to prevent Al from turning evil is like disrupting the space program to prevent overpopulation on Mars It ignores the technical difficulties It also ignores the potential benefits Stephen Hawking in recent remarks said Al is either the best thing that will happen to the human race or the worst thing It's nowhere in the middle It's one or the other Before I continue l d love to get a sense from you How many people are really worried about Al and its impact How many people see it as a major positive for our society Okay it s a techy crowd Maybe I should quit while I m ahead I don t know if you saw a lot of hands went up on the positive side Let's talk about this in more depth I think the first thing we need to do is really separate science from science fiction and the Hollywood hype from the realities of AI We need to ask ourselves What is Al tOday The answer to that question is certainly very different than what you see in Hollywood films what We've seen an AI you see in Westworld etc program defeat the What is Al First of all for many years Al was a lot of hype Champion in chess That's something that's changed We have seen a number seen AI of very real Al successes We've seen an Al program defeat the world champion in chess We've seen Al do speech very recognition very well That's the basis of Siri Alexa and That s the basis OT Siri systems like that We have systems that can now do facial Alexa and systems like recognition with very high accuracy That's used in security that We have systems that can now do facial recognition with very high That s used by Facebook It s used in a variety of ways We have very powerful machine translation for those of you that have used translategooglecom or other such apps Skype will simultaneously translate as you re talking to accuracy somebody These are all major Al achievements over the last few years Of course we had IBM's Watson defeat the world champion in Jeopardy should point out that that s IBM Watson the program not IBM Watson the brand that came later That's a whole other story There really is a lot of hype there Of course there is success in robots My point is there are a lot of impressive things going on What s really interesting is that all these successes come from a very simple paradigm that s been remarkably successful This is probably my most technical slide I want to show you how all this Al success is generated even if you're nota computer scientist It really comes from something called machine learning I should caution you our field artificial intelligence machine learning is full of grandiose labels Artificial intelligence is still not quite intelligence Machine learning isn't exactly learning Let's talk about what it does do Machine learning is you take a bunch of categories ore-specified by a person In my example we have some cats We've got some dogs We ve got some flying animals Those are the categories You take some examples of these categories These are typically represented as vectors of numbers and you associate a label with each vector The vector says this is a cat This is a dog When I say you I should be clear This is done manually The computer is not doing that The definition of the categories the creation of the labeling of the data is all done by people There's a machine learning algorithm that does what I like to call the last mile of learning After everything has been set up with copious manual labor by people then a statistical algorithm looks at the labeled data and creates a statistical model that is then used to make predictions The next vector i see is a cat The next image see is a dog This is the paradigm of machine learning What s remarkable is that you can apply this paradigm to speech signal you can apply it to email messages categorize them as spam or not You can apply it to road following cars staying on the road this is what s been driving all the Al successes that you've heard of This is great stuff If you ask me what's going to happen over the next five or ten years we re going to see more and more of this in lots of arenas in healthcare in enterprise in all kinds of places At the same time we have to remember that this is still a very limited technology These capabilities are very very narrow If we can do chess we can't do other things The program that plays Go can t do other things We'll talk about that I like to call these things Al savants because they re so so narrow in their capabilities The poster child of this on the one hand tremendous but at the same time narrow capabilities is of course AlphaGo the program that in March 2016 defeated the world champion in the ancient game of Go which is a fantastically difficutt game with many many options The thing is it's still a very very limited system I imagine to myself let's say go to some Al cocktail party the kind of place where geeks go Let's say I meet AlphaGo and we were to have a little dialogue I might say AlphaGo congratulations You defeated Lee Sedol but can you play poker AlphaGo would say No i can t I'd say Can you cross the street AlphaGo would say No I can't do that either I d say Can you at least tell me something about the game of Go the game that you won AlphaGo would say It can t explain itself The remarkable thing is AlphaGo doesn t even know it won All it is is a fancy calculator applied not to multiplication and division but to analyzing the different possibilities in the game of G0 which of course is a game that s black and white with moves It's very very artificial and stylized What we need to understand is programs like AlphaGo are intelligent in a sense a narrow sense mjt's a savant but what they aren t is autonomous This distinction is really really important because we re used to thinking of intelligence and autonomy as going hand in hand The intelligence that we're familiar with is intelligence of people and peeple are of course autonomous as well With machines It s very scary to have it s really really different To make this distinction clear l'm a weapon that can make setting up a two-by-two table so we can analyze this What llife or death decision trying to show you here is that intelligence IS not autonomy These are orthogonal notions a human the loop The intelligence is not the problem It's the of teenagers drinking on a Saturday night in contrast on autonomy Let s look at the upper left hand corner here What we have is high autonomy with low intelligence that s a bunch the bottom right we can have high intelligence with very low minimal essentially no autonomy That's a program like AlphaGo autonomy is restricted to choosing which move to make in the game it can't even decide on its own to play another game That's how limited it is This distinction is really important Let s look at various other examples and remember it Let's take for example computer viruses They re dangerous potentially very destructive but their danger comes from their autonomy from the fact that they can go from one machine to the other over the internet duplicate themselves potentially causing massive damage They re not intelligent I like to point out that my seven-year-old is really more autonomous than any Al system He's very intelligent too but that's not my point He can cross the street He can speak English He can hear English at least when he wants to He doesn t play Go but he s a highly autonomous being again the opposite of this savant type behavior we see in Al systems Let's talk about Al weapons That s a topic that has received some attention Often it gets people's heart racing You start thinking about Do I really want intelligent weapons imagine a weapon that can launch itself fly halfway across the world and kill somebody That's the stuff of nightmares What I want to point out is that the nightmarehas to do with the autonomy lt's very scary to have - a weapon that can make a life or death decision without a human in the loop The intelligence is not the problem It's the autonomy Intelligence in weapons could actually prevent mistakes like we've had where innocent civilians get killed 80 the key thing that we want to avoid again is not intelligent weapons but autonomous ones or at least we want to think about very carefully Again l m just giving these examples to set up this key distinction To those of you who are very optimistic about Al I do want to pose the question about what's Al's impact going to be on jobs That's a very serious question that merits a lot of discussions that I don t necessarily have the answer for Hal Varian the chief economist at Google has said that old jobs are going away but new jobs are going to come and replace them I frankly don't think it s that simple It s a question of at what rate Old jobs have gone away New jobs have come But it seems like this change is happening at an unprecedented rate l'm not a Luddite l'm not suggesting that we just stop Al These are pictures of the Luddites throwing shoes into the looms because they were concerned about their jobs in the textile industry Obviously the progression of technology has a lot of benefits Think about washing machines antibiotics textiles it's not something that we can just rule one way or another Still peeple ask the question If it's not clear what impact Al will have on our society if there are pluses and minuses why don t we at least declare a moratorium on Al Why don't we slow down and give ourselves a chance to think about it That can be quite if we slowdown Al appealing progress in this country we very much d0 80 at BUT you say things like Bill Gates himself This is almost like peril Right HOW WE have a deity Bill Gates is a very smart guy He said Why don t something If an 3ng we put a tax on robots Which of course has the effect Bill Gates himself said when you come from Seattle of slowing things down at least the proliferation of robots The problem I have with that idea is that Al is very much a global phenomenon China has declared explicitly that they want to be the world leader in Al by 2030 It's right around the corner Putin has said that the leader in artificial intelligence will rule the world So if we slow down Al progress in this country we very much do so at our peril Right now we have something of an edge l'm not sure that we want to give that up Even in the case of weapons which is a very again tricky issue I don't mean to simplify it But when people ask me about Al weapons I say the one thing that fear more than highly powerful Al weapons in the hands of our military is highly powerful Al weapons in the hands of rogue nations in the hands of terrorists We actually benefit from a healthy competition in this area There s a whole other dimension to this though and that s the dimension that Paul Allen my boss had in mind when he created the Allen institute for Artificial Intelligence That's to set up a mission of Al for the common good So at Allen Al or Al2 as we call ourselves we're not trying to use Al to create weapons to violate people's privacy not even to sell you things We're really trying to use Al to make the world a better place I want to give you two key examples just to highlight the potential benefits of Al one example we re working on and the second example we aren't but many other people are The first example has to do with AI based scientific breakthroughs The number of papers and I think i don't need to tell anybody in this audience this fact the number of scientific papers is growing explosively It's actually doubling every few years or so There are thousands and thousands of papers on cancer alone being published every week Let me tell you they're not getting easier to read either The number of papers that any of us can read in our lifetime is relatively small and the number of papers is growing 80 we have a problem There are no Renaissance men and women anymore What we need is some kind of tool to help us be better scientists to help us be better engineers We have a project at the Allen Institute of Al called Semantic Scholar What it s doing is using natural language processing the ability for computers to understand certain things in text - not understand Shakespeare and nuance connotations but try to extract the basic facts from turgid scientific prose Map that into knowledge that the computer can use and that scientists can use They don't have to read all these thousands and thousands of papers What if the cure for an intractable cancer is right now hidden in all these thousands if not hundreds of thousands of different papers and different studies Can machines help us by teasing apart this information it and presenting it to a medical researcher to help him or her make progress on that cancer Without going into the technical details here our vision is to say What if the cure for an - intractable cancer is right now hidden in all these thousands if not hundreds of thousands of different papers and different studies Can machines help us by teasing apart this information it and presenting it to a medical researcher to help him or her make progress on that cancer That's some of the potential that we re working on every day at the Allen Institute for Al My colleague Eric Horvitz likes to say it s the absence of Al technologies that s already killing people It's not that Al is going to be terminal-like and kill us It's quite the opposite He's not just talking about Semantic Scholar and the information in scientific text The third-leading cause of death in American hospitals is some kind of doctor error Information systems Al systems that can analyze what s happening in a hospital that can detect potential mistakes that exhausted and overworked doctors make could really save an enormous number of lives -A whole other arena this is one that we're not working on is driving Frankly human drivers worry me I have a 17 year-old He s texting and driving I know this because hetext from him I have this moral dilemma do text him back and say stop texting or do not text because I don't want him to read my text You worry about your kids They're texting and driving Even if your kids aren't texting and driving other people s kids are and they're going to run into your kids It's a huge problem Certainly some of his classmates are drinking and driving is a major problem for us it doesn t even have to be this nefarious A while ago a friend of mine in Seattle was jogging and he was hit by a car Thankfully he's okay the car was going very slowly It was driven by a 96-year old driver It was not a great idea but at the same time as we get older and we have parents we don t want to limit their mobility We want them to continue to be independent and that and still keep the rest of us safe Obviously self driving technology can make a huge difference here a very positive one There are more than 30 000 highway deaths each year close to a million accidents A lot of these can be prevented if we have better technology using Al techniques Again remember Al is a tool here it s really not that different than anti-lock brakes or automatic transmission It s not like these cars that are going to become increasingly safe it s not likea hundred cars are going to band together and say we're going to take over the White House They don't decide where to go They're tools at our disposal It s important to remember that again this distinction between intelligence and capability safe driving and autonomy which still remains in our hands The question then becomes Am I suggesting that everything s fine Al's going to be beneficial Sure we have to worry about jobs but we ll leave that to the economies But let Al blossom unconstrained unfettered No I'm not saying that either It s a complex issue We need to think about how we prevent Al from harming us What are negative uses of Al When you think about that there s actually been relatively little written or thought about this One almost naturally goes to Asimov s Three Laws of Robotics How many of you have read some of Asimov s stories Okay most of you but not everybody to understand what s Let me quickly review because even for you it s been a long harmful or what s Gt time His three laws are 1 A robot may not allow a human being to come to harm it seems like a good idea 2 A robot really requires common must obey its orders so long as it doesn't conflict with the first sense_ Remarkany law so you can't order a robot to harm somebodyrand then that s been one Of the 3 A robot should protect itself again so long as it doesn't conflict with the first or second laws These laws are really hardest things for US to quite elegant and they ve survived for more than 60 70 years give to the machine At the same time they're quite ambiguous Asimov had all these stories that you read where he showed that there's contradictions There are problems It's not simple to enforce or even understand what these laws arewith what is this notion of harm You can say harm but how do you communicate that to a computer All the computer understands is a programming language Imagine that have Alexa on my laptop or Cortana or whatever it is I tell it 'Reduce utilization on my hard drive It says Yes Oren done say 'What did you do It says took your dataset that took you years to assemble It s not backed up anywhere There's about 15 gig i deleted it That proposal you wrote for funding I deleted that too and all the copies I've saved you a lot of space 80 it was obeying my command but in doing that it created see people shuddering It creates a huge amount of harm The problem is that to understand what's harmful or what's not really requires common sense Remarkably that s been one of the hardest things for us to give to the machine There really are no machines today with even a modicum of common sense like to represent that with what I call the Al car wash This is a guy washing his car in the rain It's not a great move This is what Al systems are like today While they can play Go very well or even do speech recognition they have no common sense That's a huge problem if you want them to be able to do things and at the same time avoid harm Basically Asimov's laws are fantastic but they're not very practical In trying to be more pragmatic about it I wrote an Op-Ed piece for The New York Times It just came out in September I tried to suggest a regulatory framework Again not all the answers by any means but some ideas for thought How can we think about constraining Al First of all to those of you who raised your hands being scared of Al I do believe that we ought to put an impregnable Off switch on any Al system This is a picture from 2001 A Space Odyssey where HAL the computer says can t do that Dave When the computer is killing Dave and says can't do that we need to be able to just turn the computer off That's a fundamental principle Another thing too to think about is the fact that it is very hard to regulate or constrain the research field itself lt's The line between computer technology and Al technology is actually very unclear Instead my suggestion is let s regulate Al applications Let s not try to regulate the research fast moving lt s amorphous The line between computer technology and AI technology is actually very unclear Instead my suggestion is let's regulate Al applications Let's not try to regulate the research When we have cars that have Al and of course have the potential to cause accidents when we have toys that have Al and have the potential to violate privacy when we have robots in the workplace and We should have a rule that a computer system should engage in full disclosure and disclose that it s an Al system we don't want our Al that s privy to more and more information about us to reveal that information potentially in the home those are things that we can regulate and we should What might that look like The first thing we have to adopt is the notion of responsibility An Al system ought to be subject to all the laws that apply to its human operator its human manufacturer If my Al car crashes into yours I can't say Don't blame me It was my carexcuse We have to take responsibility for our intelligent cars the same way we have to take responsibility for our unintelligent cars That's an important legal principle It needs to be elaborated but it will help prevent irresponsible use of Al A second thing that's becoming increasingly important is full disclosure lt's easier and easier That s going to change even more so in the future for an Al to pretend that it's actually a person We see this in Facebook We see the Twitter bots We now found out that in the most recent election there was more and more of bot activity masquerading as human activity etc We should have a rule that a computer system should engage in full disclosure and disclose that it's an Al system Another remarkable thing that's happening is our technology is violating our privacy more and more That s even before we have Al This is from a recent article It turns out that in certain instances Google actually gives a user a number of options including saying don't like this ad because it knows too much about me This is not science fiction this is not a proposal this is a real thing that Google has unveiled because some people are Just getting more and more upset by how much our technology knows about them If you think of systems like the Amazon Echo which records audio In your house Al Barbie we have these Barbie dolls with chips in them engaging in dialogues with our kids Who knows what information our kids are telling Barbie I find that more scary than funny Even the Roomba you think l've got this little hockey puck robot It s cleaning dust in my house What's the big deal It turns out that in the process of doing that it's building a map of your house Apparently lRobot the company that manufactures this was actually considering selling this information to third parties They didn't do it but they were consrdering it It's not We want our Alto avoid bias Al actually has a great potential something that you thought that your Al robot was doing Imagine more sophisticated robots They pick up the phone and say He needs a new carpet It's really fraying on the to catch human blaS edges Help me out here This is not something you want WhEthBl' it's in jUngS happening without your approval 80 we don t want our Al that s privy to more and more information about us to reveal or in people to alert us to say loan processing gone awry that information Another topic this I have to confess wasn t covered in my Op Ed partialfy for space reasons I'm adding one more regulatory principle We want our Alto avoid bias Al actually has a great potential to catch human bias whether it's in judges or in people to alert us to say loan processing gone awry We have studies that showed that when judges get hungry they tend to produce more negative decisions So AI could alert to Your Honor maybe it's time for a snack your blood sugar is dropping There s a lot of benefit here There's a really interesting problem This is a bit technical with Al It goes like this The data that we give our machine learning system is typically culled from the real world It may contain some bias in it all kinds of bias What's remarkable about machine learning technology is that it tends to generalize It attempts to compress the data into some general principle When it does that it can actually amplify the bias that's in the training data which is very negative So whatever bias we have in the training data let's say loans were denied to people of a certain gender or a certain race some percentage of the time the last thing we want is Al to automatically say l get the pattern here Let's be more aggressive on denying that That s what my data is telling me This by the way is a problem that does have a technical answer One of the research scientists at just won a best paper award on work to avoid amplifying bias even beyond what s in the training data That s another important issue We would agree we want to avoid bias I could go on and on but I want to leave us time for questions and discussion just want to conclude by highlighting what I - consider the most important point I started with this question Is Al good or evil My answer is it's neither lt s neither good nor evil It's a tool It s a technology More than anything it's a pencil lt's a fancy pencil one that we can draw amazing pictures with But a pencil is a tool that we use We get to choose Do we draw nice pictures or do we draw unpleasant horrific pictures The choice is ours I hope you'll join me both in the conversation and in working as a society to make sure that Al is used for good not for evil Thank you very much ls Al good or evil My answer is it s neither It s neither good nor evil It s a tool It s a technology More than anything it s a pencil It's a fancy pencil one that we can draw amazing pictures with But a pencil is a tool that we use Committee on Oversight and Government Reform Witness Disclosure Requirement Truth in Testimony Pursuant to House Rule XL clause and Committee Rule 16 a non-governmental witnesses are required to provide the Committee with the information requested below in advance of testifying before the Committee You may attach additional sheets if you need more space Name 0mm 1 Please list any entity you are representing in your testimony before the Committee and brie y describe your relationship with each entity Name of Entity Your relationship with the entity Pnum me N r r Armc lm inwil'qen 09 Chief Exec 3 OF onjtegsor 2 Please list any federal grants or contracts including subgrants or subcontracts you or the entity or entities listed above have received since January 1 2015 that are related to the subject of the hearing Recipientof the grant or Grant or Contract A enc Pro ram Source Amount contact you or entity above Name 3 Please list any payments or contracts including subcontracts you or the entity or entities listed above have received since January 1 2015 from a foreign government that are related to the subject of the hearing Recipient of the grant or Grant or Contract A one Pro ram Source Amount contact you or entity above Name 'on above and attached is true and correct to the best of my knowledge Date lb Page I certify that the infor Signature - Oren Etzioni CEO Allen Institute for Artificial Intelligence Dr Oren Etzioni is Chief Executive Officer of the Allen Institute for Artificial Intelligence He has been a Professor at the University of Washington's Computer Science department since 1991 receiving several awards including Seattle's Geek of the Year 2013 the Robert Engelmore Memorial Award 2007 the IJCAI Distinguished Paper Award 2005 Fellow 2003 and a National Young Investigator Award 1993 He has been the founder or co f0under of several companies including Farecast sold to Microsoft in 2008 and Decide sold to eBay in 2013 He has written commentary on Al for The New York Times Nature Wired and the MIT Technology Review He helped to pioneer meta search 1994 online comparison shopping 1996 machine reading 2006 and Open Information Extraction 2007 He has authored over 100 technical papers that have garnered over 1 800 highly influential citations on Semantic Scholar He received his from Carnegie Mellon University in 1991 and his BA from Harvard in 1986