it was at dartmouth in $956.00 that mccarthy organized a summer long research project focusing on machine intelligence. it would become part of ai folklore and the place where the term artificial intelligence was 1st coined it would also bring it to the attention of the u.s. military. in those days the military was the principal source of funding for computer science research and if you went into the founders you said you know we're going to make these machines smarter than people some day and whoever isn't on that ride is going to get left behind and big time so we have to stay ahead of this and boy you got funding like crazy as the defense department took over more of the funding the question started to being asked you know but what can we do with it now . the u.s. department of defense had its own research arm called the advanced research projects agency arpa the push for developing a i was driven by the logic of the cold war any technological advantage america

could get over the soviet union was pursued through our. the new field of artificial intelligence was flush with money and confidence about what it could achieve. but there were competing ideas about how i would lead the way to this brave new world. from the very 1st day there were 2 big approaches one of them was we're going to figure out the rules and we're going to teach the rules to the machine right we're going to say this is how you do this 1st or this 1st or that 1st of that so the 1st one is the so-called expert systems where you just codified old rules you can and say go for it the 2nd one was we're going to show it things we're going to feed it data and it's going to learn from the data you feed it the data and expected to build what people call a neural network which is that it looks at the data and the results and says ha ha

this this this goes together and then does it again and does it again and till it builds almost like a brain like structure. what happened was a book was published actually marvin minsky one of his colleagues that basically showed that these neural networks could not really learn certain things. which typically happened in ai is. something comes along that makes us start to doubt a piece of technology and that closes to sort of go underground for a while while the other stuff gets more attention. development of neural networks slipped into the shadows of ai research as expert systems took center stage and all the money but by the airlie 1970 s. the great advances promise by artificial intelligence had failed to materialize. a clumsy robot aptly named shaky and rudimentary processing machines fell way short of what the early ai visionaries had sold to their backers. for the us military ai

had lost both its appeal and its purpose. it cost what you call the ai winter which stopped for quite a while. the so-called ai winter would freeze state sponsored development of artificial intelligence for over 2 decades. chasse the number of bored possibilities is just astronomical. there's so many possibilities in chess that by the 20th move a chess board there are more possible ways the board could look than there were molecules of the universe. chess a measure of human intelligence for centuries had become a benchmark for how far computer technology had progressed and

a proxy for how smart computers could be. with state money frozen private companies such as i.b.m. funded their own development of ai. ringback and we did have this fall. in 1975 b m engineers built a computer the took on world champion garry kasparov in a series of chess matches. they called this computer deep blue. this is amazing engineering i thought it was so good 64000 processors going it really high speed through chess millions of noosa seconds. between the 1st game of the 2nd game the computer was trained on lots and lots of casper of games so it wasn't just becoming a good chess player but it's becoming tuned to playing against that particular person i was there actually at that at the time here in this building where the

game was actually the machine was and where the machine was it was amended and i was looking on the systems aspects of it it was search algorithms but search algorithms that were intelligent from the point of view of thinking about one other . over a series of 3 test matches deep blue beat gary kasparov at outmaneuvered out thought and out played the greatest chess grandmaster of his day. artificial intelligence had burst out of its winter and now looks set to blaze a revolutionary trail. very. high. that was a really key moment was in that when when i.b.m.'s deep blue garry kasparov to me that it still gives me goosebumps that it's back to war 3 broke through at that point in time it was the pinnacle of intelligence for anybody to be able to play

chess at the level that the grand master of the play and beating the grandmaster itself is is i think its importance cannot be overstated. but the capabilities that vent into it brought together algorithms. infrastructure which is hard and data sort of 3 of these aspects came together at that point and i would say it was the precursor to our latest revolve for ya i. i b m had built deep blue using the established expert systems model of ai the chess victory over kasparov was its greatest achievement to date. but to beat a person at a game involving set patterns and strategy was one thing beating humans at a game of general knowledge was another. the next big mark in the public

imagination was i.b.m. swats that was a machine that could play jeopardy. jeopardy's a game in the united states we give the answer to a question and then people have to try and work out what the question is. and became clear you couldn't win a game like jeopardy by just building an expert thing in each category there's just too much so they started moving to a different direction and they brought together bunch of people eventually from all parts of the company. by 2011 i.b.m. engineers were. working with new tools one key challenge was understanding human language using advancements in the new field of natural language processing they build layer upon layer of algorithms mathematical structures that allow the machine to learn human language through a mass input of data language and communication is the essence all us as being human beings. it's one of the hardest

tasks and hardest barriers for it to have crossed after we did. this u.s. president negotiated the treaty of portsmouth ending the russo-japanese war watson who is theodore roosevelt good for $800.00 what i.b.m. did was they played the best players in the world they made the t.v. came of the players who don't the best the 2 very top players and beat the. life on t.v. . i.b.m.'s watson computer like deep blue had triumphed in the battle between human and machine but unlike deep blue watson was not strictly an expert system its novelty was machine learning a branch of artificial intelligence that had been driven underground decades earlier. the machine learning side it

almost died out but from ninety's on when you started having the internet and all these digital devices in the mountains and mountains of data and you started feeding this huge amount of data to these machine learning systems they were uncannily effective. far from being abandoned machine learning had continued developing away from the mainstream of artificial intelligence. alongside technological advancements personal computers laptops mobile phones high speed microchips. and then came the world wide web search engines tech giants social media smartphones machine learning now had the 2 ingredients it always needed massive computer processing power and data masses and masses data. machine learning was now set to take off.

the united states government has been seeking to use machine learning to isolate targets for drone attacks for many years because i'm afraid we americans have a bit of a tin ear for irony it calls its program skynet. in which it is going to seek to try and find targets for attack when we talk about ai and warfare the other thing i want to be clear about is this isn't about the terminator right this isn't about killer robots what i want to talk too about is humans and targeting and the way that human intelligence analysts relate to information that comes out of semi-automated processes basically from people's cell phone metadata and we try to determine their so-called pattern a life where they're going to they know. who they talk to here's the problem i

think that the united states has the n.s.a. learned to collect it all well before they were able to understand it all faster is faster absolutely machine learning an artificial intelligence would permit the defense department to accelerate the process of target selection faster and always better. i want to make sure that people understand actually drones have not caused a huge number of civilian casualties for years the drone wars were an open secret in washington but it wasn't until 2012 that president obama officially acknowledged the program for the most part they have been very precise rescission strikes against al qaeda but that was the same year we started to hear about signature

strikes where the cia or the defense department would target not a name or an identity but essentially a phone this is a targeted focused effort at people who are on a list of active terrorists. they would collect signals data especially telephone data and put it the receipt algorithm they would decide you were a threat based on who you talk to where you went to your friends were but that's not precise at best it's an educated guess so i started to wonder what if finals brother in law and his nephew were killed because of an algorithm. mornings and snowden time 29 years old i worked for booz allen hamilton as an infrastructure analyst for n.s.a. . in 2013 edward snowden blew the whistle on the u.s. national security agency's mass surveillance program he revealed that foreign intelligence gathering tactics were now being deployed at home casting

a nation wide net to catch a few bad fish. so while they may be intending to target someone associated with a foreign government or someone that they suspect of terrorism they're collecting your communications to do so. a year later michael hayden the former director of the cia and n.s.a. stopped short of disclosing the sky net programs existence but admitted the targets were being identified for so-called signature strikes using metadata collected from everyday communications we kill people based on. a signature strike would be selecting someone who looks like other people that you think are the people you're looking for you know one critique of this is demography is destiny in some sense. so we don't we can't know because of the nature of the beast right exactly what's going on whether skynet was ever used but it seems likely to me that given what we've said about signature strikes that some kind of algorithmic targeting process

is being used to isolate targets in places like yemen and pakistan using partly signals intelligence to be clear we have no evidence that that model was ever used in practice i would be shocked if it isn't being used machine learning is pretty good at finding elements from a huge pool of non elements when president obama took over the so-called war on terror he favored drones because they meant fewer boots on the ground and as the drone wars expanded the method of selecting targets changed. metadata was not a key source of intelligence but there was far too much of it for human analysts to process the machine learning would be used to identify targets and sift the good guys from the bad at least that was the idea machine learning is good at this problem called binary classification so is someone part of a terrorist network are they not part of a terrorist network machine learning is also pretty good it asymmetric problems

where there are very very few things you're looking for in a sea of things you're not looking for a few people who are part of terrorist networks in a sea of civilians for example so where do you think that kind of statistical targeting is likely to go wrong humans to binary classification all the time ok we have to decide if we're looking at you know a bunch of for example a bunch of dogs at the dog park and which ones are dangerous and which ones aren't when we say no that dog is not dangerous but it is that's called a false negative if we made a mistake we can go the other way we can have a false positive oh my gosh that dog's really dangerous but it's not so people might take action on those false positives and will get innocent people killed i'm also worried about false negatives the thing about machine learning is that it assumes that the future is like the past it assumes that the things it's looking for in its prediction are like the things that it saw in the training ground so

if we have a problem where people are actively trying to disguise their activities. then a machine learning model is likely to make a lot of false negative mistakes it's going to miss a lot of people who are likely terrorists that said i'm not sure that there's any other way to do it but the risks are very substantial and almost certainly mistakes that lead to people's deaths are going to happen here. we always knew the strike on final family was a mistake. we met congressman we talked to senators we haven't spoke to people in the white house national security council but while everyone expressed regret for pfizer's loss no one could explain why it happened or how it happened.

talk to their own. problems and besides that instability is corruption we listen. who are pushing the united states and president trump into conflict we meet with global news makers and talk about the stories that matter just 0. we have a news gathering team here that is 2nd to their all over the world and they do a fantastic job when information is coming in very quickly all at once you've got to be able to react to all of the changes and as you know we adapt to that. my job is is to break it all down and we held the view on the stand and make sense of it. a natural resource that's gone untapped for more than 2 decades

alison is. off course years before israel has found its al-jazeera world tells the untold story of gaza as an exploited gas fields gaza loggin is only minimal as of the palestinian so it's a lot of money and how this valuable resource could have transformed palestinian lives. regard the gas deal an al-jazeera. piece it will be in doha with the top stories on al-jazeera the u.s. president donald trump says his government has reached a migration deal with mexico perspiring trade tariffs that would do to come into effect on monday under the agreement mexico will strengthen its migration laws and curb human smuggling operations so the flow of migrants into the u.s.

can be stemmed the deal will also allow migrants currently in the u.s. awaiting asylum to be sent back to mexico until their cases are resolved mexico's foreign minister spoke to journalists just after the deal was announced by donald trump in a tweet. because 1000 years the. united states will immediately abide by section $235.00 and will deploy along their southern border this will imply that those who cross the u.s. southern border to request asylum will be returned to mexico where they can wait the resolution of their application for asylum in its part of mexico for humanitarian reasons and in compliance with international obligations we authorize the entrance of those people to await their asylum applications if europeans prime minister has called for a speedy democratic transition in sudan as he tries to mediate a political crisis in the country. in khartoum where he met the ruling military judge and protest leaders days after a military crackdown left more than 100 protesters dead the opposition says is open

to talks but only if certain conditions are met. first the military council needs to recognize that a crime was committed secondly there needs to be an international investigation into the dispersal of the sit in thirdly all political detainees and all political prisoners held by the previous regime need to be released there needs to be freedom of speech and the media the military needs to be pulled from the streets and the internet ban needs to be lifted until all the demands are met we will not hold talks on a future political process the u.s. has raised the stakes in a standoff with turkey over ankara's plans to purchase russian air defenses washington is giving anchor until the end of july to backtrack on the deal if that doesn't happen is threatening to expel turkey from its f. $35.00 fighter jet program the 2 nato allies have sparred for months now over turkey's order of russia's s. $400.00 defense system those are the headlines the news continues after the big

picture i'll have the news for you in a little under 30 minutes. my client faisal banally jobbers brother in law and nephew were 2 victims among hundreds caught up in the american drone more in yemen. protested the loss of civilian lives few knew that these were victims of computerized targeting built on data gleaned from phones and surveillance apparatus fed into algorithms and narrowed down from possible terrorists to probable terrorists to definite terrorist . neither finals brother in law or nephew was a terrorist. as a human rights lawyer i had been investigating civilian casualties and drone

attacks in yemen and pakistan and over time we came to understand that the people we work with who lost loved ones were killed as a result of a semi automated algorithmic targeting process and there was a kind of article that came out about an n.s.a. slide deck that said well this this is how we're going to use machine learning to find an isolate targets in pakistan and somebody phoned you up as a statistician and asked for your take on this and you were like well look this is just this is just totally unsound yeah it was pretty unsound i mean is it just basically kind of racial profiling it's cal how do you say it well it's more complicated than that because the kind of information that we're working with in that case is to lessening that if the problem in that particular example that the n.s.a. analyst were trying to figure out was ok we identified the cia had identified a small number of people who were couriers for terrorist organizations were carrying around flash drives and messages among groups of terrorist organizations.

and the question is can you use the. phones calls that these people make and the places that their phones check in with the souls assuming we've got those people right but anyway yeah i think i think that the court i'm willing to believe that they got these people right the questions how many more are there that they didn't get yet but can we use these people's information to disambiguate them to tell the difference between these people and all the other people in a city they live in and you tell how different they are so that you can just use that to leslie meditator to predict which of the people who we know are actually are terrorist couriers and how few other people can you include in that list as few false positives can you get in that classification and i was very skeptical about the particular model that was used in that case do you see some resonances here and you know we're talking about communities abroad who have basically designated

a threat and then targeting them for armed attack does that chime at all with some of the policing and other targeting of communities of color in the states that you've thought it. you know i i i resist the temptation to kind of essential lies these as the same in every part of the world in every context we do know though that. technologies that predate these kinds of automated technologies about classifying and figuring out who might be a threat. disproportionately or targeting people who might have politics that are say more to the left people who are more interested in civil rights people who are advocates of human rights labor organizers for example so i think the question is again always underneath these projects where the values and the politics of who who's being assessed who's being classified for what

purposes and those are fundamental questions that are going to be with even with the next version of the technology top l.a.p.d. spying coalition did research on drones being deployed in l.a. in 2015 they produced a report and they actually drew very explicit lines between the use of drone technology in the middle east and the proliferation of drone technology in police forces across the united states through the urban area security initiative and they have been making these connections right the military is ation of the police. the increase in 3rd party private services that are bought by public institutions so what how do you see it you know the defense community affecting the development of these technologies is that the kind of original sin of the thing i think that particularly in an academic contexts. the big chunks of funding have been for decades and continue to be from from defense i don't think that most

of the funding that comes out of the fence establishment that's for pure research looks like it's going to be used to kill people the larger problem is that it creates a kind of. philosophical or even ideological framework that it's ok to produce things for the defense establishment and that it makes sense to produce things for law enforcement because they're here to help us that failure of having a critical understanding of what military and police institutions do to the communities that are on the receiving end of their business i think that's the larger problem the residents i kind of see it and the reason i guess i asked the question is somebody who did work with those communities in yemen in pakistan is that they talk about feeling scrutinised and over police in a way that i hear when i hear you guys talk about these other communities you know the sharp end of the policing so it's almost as if the language that comes out of them that says we feel our whole community are suspect it's never point it is that

white collar crime the containment the prediction the assessment always seems to be at a community who we fear in some way in that we want to contain well and who the we is in that is very important because we also don't profile and track for example white supremacists nazis neo nazis in the same way we don't talk about them as domestic terrorists for example in the u.s. or in other countries so the framing of who is the threat is ultimately always the value question i think that's on the table that we have to be thinking about and of course the more you ought to me the profile of who the threat is who the threatening other is the more you flatten these conversations about values the harder it is to actually talk about what people are struggling for what struggles for justice around the world look like i mean and then some of the very companies who develop these technologies expand into other areas don't you see you've got

talent here the security and intelligence firm who developed tools for counterinsurgency in iraq then sold some of their kit to the l.a.p.d. recently inking a deal with the u.n. world food program seeking to help them spot fraud i mean. where then it makes you think mass knows no context we can sell our product anywhere as long as we're finding us to testicle relationship. people can take 500 pictures to get that perfect shy. which is something when i was 20 could not have done because of what it on film and it was and you got what you got. one of his. vast swathes of the world with cameras on smartphones billions hooked

up to social media sharing images online gave machine learning engineers access to a massive data billions of images these images could now be used to train algorithms to teach themselves to detect particular features recognize particular forms this new breakthrough was called deep learning very soon ai would use deep learning to teach itself to recognize the form of a cat just by looking at millions of cat images on the internet. from there that made the leap to recognizing people. artificial intelligence now claimed the ability to pick out a specific face in a busy crowd to find the needle in the haystack. so what you have in france fear is a very typical surveillance saint's as a camera looking at a public street and people clearly want us to come on face recognition go to work out that it's seen a person and then it's got to work out with a person's face and it's got to take

a biometric caution about face in essence something that it can pat can compare against watch list if there is a match then that will alert and you'll see thoughts come up on the screen stop person is now going to the system has been identified clearly you can set your system up to alert the right people when that happens right so if you wanted to kind of test it out on my face for example how would we make that happen so very simply we would have a surveillance photograph or it could be something taken from social media or just from the internet that would be loaded into the watch list as you can say here such that if you walk past a camera that is linked to that watch list then you should be attacked if you have been seen and then send it to the appropriate place so we give it a try are. you

great crowded street saying lots of people coming towards the camera. each time the camera attacks that same person but they will biometric hotshot person and say if they're on the watch list so you can see here for example guys cliff faces on the person against the watch less than on their back consequent it up as ago or so as if the camera doesn't say no it's ok. the facial recognition engine is working hard right now checking all of these people when clearly. you know coming now it's more screen and you know just a cultural view of your face an initially just new york it will continue to identify the level of confidence will change depending on your angle to come up but the minute in one single for a player on the list. these facial recognition technology is should not exist for what purpose are they being

brought into existence again who are they being pointed at and what kinds of protections should we have against the common critique of it is while it doesn't seem black faces very well but is that is that the whole problem with facial recognition i think for the people away joy. their position would be these technologies exist and black people will be ensnared in the especially black women the failure rate is greatest on black women are going to be misread by these technologies and therefore potentially harmed because they will now be able to kind of fight back against these technologies that are pointed out that. join one of them when the of the massachusetts institute of technology showed how facial recognition programs were unable to recognize black faces particularly the faces of black women. the problem was data and bias engineers who were

mostly white males trained algorithms with a data set of images that were themselves overwhelmingly of white males the resulting algorithm struggled to recognize non white non male faces. the algorithm simply couldn't compute what they saw with what they'd been trained on the same racial and gender bias that existed in the wider world was trained into the system itself. one of the other key parts of the debate has to do with the extent to which the technology works on different faces so whether it's as effective on people of color or women is that changing does a kind of depend on the technology. is only them how it's trying so saw technology for example we have trained it on different emma graphics in different parts of the world but certainly in terms of if you. kind of strip the debates away the on the underlying author.