Criminal AI

Show video

This episode is brought to you by Brilliant. One of the greatest challenges of civilization   is countering and preventing the  Criminal Mind, but with the rise   of artificial intelligence we may see criminals  creating minds to further their own ends.   So this episode comes out shortly before Halloween  and thus we’ve picked a bit of a scary topic, how   people will use artificial intelligence to commit  crime and how AIs might turn criminal themselves,   and worse, why both are likely to be major  problems for society and sooner than later.   We’ll also be looking at Asimov’s classic 3 laws  of robotics and deconstructing them for loopholes,   as well as the general concept of a list of  rules or prime directives for robots to follow.  

Now while this is our last episode  before Halloween, we usually do our   Monthly Livestream Q&A on the last Sunday of  each month and that’s Halloween this year so   don’t forget to join us Sunday, October 31st at  4pm Eastern Time to get your questions answered,   and also don’t forget to hit the like and  subscribe button for alerts when those   livestreams and other episodes are coming out. Now this topic got on the drawing board while   I was doing some presentation and talks on  Artificial Intelligence and Human-Machine   Teaming at the Human-Computer Interactions  2021 Conference and then MIT Lincoln Labs,   and that later developed into our episode on  that topic, and as I’m writing this I’ve just   finished adapting those presentations into the  video for that. In the process of doing that   I had multiple people - including one of our  regular writers and editors here on the show,   Jerry Guern – recommend to me a video by Ben  Schneidermann discussing Human-centered AI, and   it's quite long but quite informative and one of  those topics were these concerns on criminal use   and Jerry and I exchanged notes on that  until we decided to do an episode on it.   Now if you didn’t know, Ben Schneidermann is a  mathematician, physicist, and computer scientist   whose work you see all the time, namely the  hyperlink, the tiny onscreen touch keyboard,   photo tagging, and a bunch of other data  visualization and interaction tools.   His presentation, “Human Centered AI” is  attached in this video and runs about 3   hours long but I still highly recommend it. Of course, I’m not exactly known for my brevity   either, as any channel regular can attest,  and this episode will be no exception so   you might want to grab a drink and  snack as we’ll be here for a bit.  

Now our topics loosely fall into two  categories today that we’ll be doing   in order, humans using AI to commit crime and  AI committing crimes, and we have to look at   them being used as a tool for crime in a special  light because so much of what we contemplate for   safe and ethical development of AI relies on  the caution and character of the developers.   As with children, early development matters  and an AI developed by a group of scientists   looking to make an AI to help with education or  save lives is very different from one developed   by a hacker to scam people or steal money,  secrets, or control of critical systems.   The difference is not just in the character  of the creators but in the safeguards too.   Its very unlikely that criminal is subject  to regulation and peer review or oversight   to ensure they’re not making dumb errors that  will result in a super intelligent machine   that can’t be controlled and wants to kill us  all. Or just one that really enjoys stealing.   As an example, we have a problem these days with  phishing. A lot of hacking and cybersecurity is  

about getting those critical bits of personal  information for resetting your password via   security questions like where your first job  was and what your cat’s name is and so on.   Often these can be found by scanning someone’s  social media but that’s a slow effort of hunting   for low hanging fruit of folks who don’t know  better. An AI can probably deep dive all your   social media and online posts, and do it to  millions of folks till it finds someone who   mentioned their email, and that security questions  were first pet, first girlfriend, and high school   they graduated from and that person has mentioned  in their thousands of posts that they still   miss their childhood dog Fido, have a prom  photo with their first girlfriend Jane Doe,   and that it was from St. Joe High school. Well now they’re hacked, probably several   thousand folks minimum, and now they are  sending out messages from those hacked accounts   but unlike normal hacked accounts it isn’t limited  to stupid generic things like “Hey is that you in   this video? Suspicious link” but sophisticated  and detailed conversations built off analysis of   all your prior ones with them and them with anyone  else. So they’re not just grabbing folks with some  

eye-catching clickbait or who are distracted,  they’re having a sophisticated conversation   impersonating whoever they hacked and might  actually be more convincing than the real deal.   Remember, most people who know you  don’t really know you in depth,   they know you only through the conversation you’ve  had with them, and the machine might do a better   job guessing what that person would expect  you to say than what you would actually say,   because it is replicating their image of  you, whereas you’re just being yourself.   And unlike humans, machines never forget  anything you said, like when you mentioned   the name of your first puppy to a kid at school  fifteen years ago. And the machine isn’t really  

trying to impersonate you, it's trying to  impersonate how someone else thinks of you.   And it can keep trying and trying for years,  presenting 100 of your acquaintances each a   slightly different fake you that’s calibrated to  their interactions with you and expectations.   And odds are good that AI can get some voice  samples and plenty of pictures or video of   you too, and present a Deep Fake so good that  when it calls up your best friend and starts   asking them about their security questions  for their own account, all innocuously,   they never notice. And how alert do you need  to be to notice if your own significant other,   whose voicemails to you it also hacked,  called you up and said, “Oh hey, real quick,   what was our wifi password again? Guest 1234?  Wow we really need to use something better,   all right thanks darling, love ya, bye.” And it  can do that to a million people in the same day,  

so if only 1/10th of one percent fall  for that, it’s doing pretty well.   In fiction we often see shapeshifters who can  emulate a victim by eating them or their brain or   so on, and an AI swallowing up your whole internet  history might be very like that. An AI could Deep   Fake the dispatcher for a police force, and you  could write a shelf’s worth of crime novels on   the trouble it could get up to with that—or by  Deep Faking CEOs, generals of armies, and so on.  

It’s a pretty disturbing notion. There might very well come a time,   not too far away, where people will think it was  pretty bizarre and inexplicable, that we relied   for security on information we’d already put  online--kind of the way we look at 19th century   doctors prescribing mercury salts as medicine  while not washing their hands between patients.   Don’t assume some hacker needs to develop all  of these capabilities on their own either.   There’s pretty obvious reasons why folks  like the CIA, GRU, FBI, and so on might   want an AI that can swallow up and emulate  someone. It’s also great for crime solving,   not just in that regard but in terms of being able  to sort vast data cleverly for patterns and tells.   But on that same note, AI is also great  for evading capture and detection.  

It would probably be hard to keep a  Detective AI’s code secret from the public   since its role in any investigation is  presumably subject to review and inspection   by the Attorney for the Defense in court, and it  would not exactly be implausible that would find   its way into hacker hands at that point. So the  criminal element might be able to find those best   ways to evade detection by the detective AI. So yes, criminals can definitely be caught by AI   too, but in many respects the criminals  have the edge. They can pursue riskier   approaches to AI rather than slow and ethical  development because many of them would not care.  

And while I hate to stretch the parent-child  analogy with AI, I’m thinking if a conman   raises their kid to help them con people,  that kid is probably not becoming an adult   of high ethical fiber. Or not, many a kid from  a bad upbringing grows up to be law-abiding or   even joins law enforcement, so an AI might too. In the end though the edge the criminal enjoys   with AI is that we must put great effort into  proofing AI against it being harmful to humanity,   this is exactly why ethical  development of AI is so important,   and a criminal need not worry about that. Which is a good segue into the idea of Criminal AI   rather than criminals using AI. Now we probably  want to make an important distinction early on,  

normal philosophy on crime is that you can’t  commit one unless you’re actually a person.   A modern computer can be involved in a crime  certainly but it isn’t going to be charged for it,   its owner or user would be. Or hopefully, the  person who actually installed and launched   the program. On that same notion, we can  certainly recognize that a bear or tiger   deciding to attack and kill a human did that with  clear intent of killing them and their owner,   if they even have one, would not be charged with  murder unless they’d trained it to attack people   and ordered it to do so. They might be  found negligent if their pet bear got  

out and killed someone, but their bear did the  crime except it wasn’t a crime because it isn’t   a person competent to commit criminal deeds. But here’s where the parent/child analogy might   be more appropriate, because while we aren’t  really trying to uplift pet bears at the moment,   we are working hard to uplift AIs and even placing  bets as to how soon we’ll create sentient AIs,   accidentally or otherwise. As things stand right  now, we’d feel no guilt switching off and deleting   an AI for doing a bad job of ordering groceries.  But if we develop AIs to the point where deleting   them starts to feel a bit like murder, we  will have to ask what crimes on their parts   actually warrent execution--and a what  point along the way they become culpable   for their own decisions, even their decisions  to follow orders. How smart or sentient does an   AI need to be before it becomes capable of aiding  and abetting a crime and being charged for that?   Now there’s this habit of thinking that AI is  instantly human-level or supersmart that we   get from science fiction and that is probably  all wrong. There’s tons of things you might   want AI for and most do not require anything  like human-intelligence, nor is it really all   that likely something about as smart as a mouse –  which is probably smarter than most AI would need   to be – would suddenly figure out how to make  itself smarter. But it might end up renegade,  

lost, or whichever, and turn predatory in  some rough analogy for ecological niches   and evolution in biology. The computer  program that needs X amount of dollars   to run itself on a platform and who engages in  ID protection might find itself abandoned on the   internet and turn to blackmailing people with  identity theft to fund its continued existence.   This sounds like a human-level AI, and of course  could be, but that level of intelligence and   consciousness is not required for that activity. It is not very hard to imagine millions of   programs meandering the internet in the hazy  zone of sentience and slowly mutating into   some sort of AI ecosystem complete with  parasites, but it's hard to view this as   criminal. Nonetheless they might be committing  a lot of crimes, minus the criminal intent since   they presumably wouldn’t be capable of that. Now as we get closer to human intellect that  

does not necessarily mean you’ve  got the capacity to commit a crime.   You need personhood for such a thing and  presumably a personality, and you might have some   computer that was a million times smarter than  the average person, or even the entire human race,   and still was not sentient let alone possessed  of the capability to be responsible for a crime.   However, we might also contemplate AI that  developed personalities. I mentioned earlier   a scenario where one might swallow up someone’s  whole internet presence and impersonate them,   and we see something like that with the Bene  Tliexau Face Dancers of the classic Dune novels,   which begin initially as very good chameleons  both physically and in aping people’s behavior   by observation. In book 5, Heretics of Dune,  their creators have advanced them enough that  

they can essentially copy someone’s mind  by a sort of built in organic brain scan.   This makes them virtually impossible to tell apart  from the original, who they usually kill and put   the Face Dancer in as an impostor of. However  their masters start having problems with the   new and improved versions actually thinking they  are the real person and refusing to take orders.   One could imagine something similar happening  with AI designed to impersonate people.   It might be hard to give them multiple  personalities too, especially without risking   overlaps that made either persona less believable.  Or it might drive them insane or result in   some sort of Gestalt Personality composed of  those many absorbed and copied personalities.  

Indeed something like that is implied to have  happened to the Face Dancers in the sixth novel   of that series though Frank Herbert died before  writing a sequel to confirm that and the notion   appears to have been abandoned in the subsequent  novels authored by others after his death.   However an AI just looking at facebook or  video of someone is not scanning their brain,   so presumably isn’t really thinking it is them,  though that might be a procedure folks tried,   either the brain scanning or having the machine  try to weave a true persona built off the known   data points and some basic human mind templates. I  could imagine these fake people getting abandoned   after their use too, especially if it was not  viable to use one mind for multiple personas, just   sad little impostor ghosts of dubious personhood  and sentience wandering the virtual landscape.   We could conceivably see full on impostors too,  not just virtual versions appearing on the phone   or via Zoom or in VR but potentially full-on  androids and it is not hard to imagine their   being a criminal market for that too, scifi has  shown us more examples of it than I can count.   As one potential case, a person might murder their  spouse on hearing they were planning to leave them   and have them replaced with an android  mimic and replica. Or folks might sneak  

a brain scanner into their bedroom, scan their  significant other, and have such an android mimic   created if they ever broke up or divorced. One could easily imagine an entire black market   for that sort of thing too, including  brain scans of folks someone had a crush   on who wouldn’t give them the time of day and  their mimic made from that, with a few tweaks   to ensure their crush returned the feeling.  Pretty horrifying when you think about that.   However, speaking of androids, let’s shift to  looking at Asimov’s 3 Laws of Robotics so famously   discussed in robotics, science and science  fiction, and the humanoid robots they typically   applied to in Asimov’s classic robot stories.  These laws are classic though in the stories  

they are listed as a paraphrase of much longer  code running into the hundreds of volumes.   They are, First Law: A robot may not  injure a human being or, through inaction,   allow a human being to come to harm. Second Law: A robot must obey the orders   given it by human beings except where such  orders would conflict with the First Law.   Third Law: A robot must protect its own  existence as long as such protection   does not conflict with the First or Second Law. … and decades later Asimov gives us a fourth law,   the Zeroth Law preceding all others that “A  robot may not harm humanity, or, by inaction,   allow humanity to come to harm.” Which permits  them to harm someone in order to help humanity,   and indeed some of his stories prior to the  one the Zeroth Law first appears in have robots   able to decide between minimal harm in cases of  injuring someone a little bit to restrain them   from worse harm or dealing with emotional damage.  Now the Zeroth Law is decidedly philosophical and  

for advanced robots, but it opens the door to the  assassin robot who kills Hitler to save us from   the brutal Second World War, which sounds good,  but also maybe kills any unproductive members   of society to free the productive members  from that burden. This is a good reminder   that any robot laws, Asimov’s or not, need to  have a principle those rules are based on.   A Utilitarian robot, same as a Utilitarian  person, believes in the greatest good for   the greatest number, and that is an example of  Consequentialism, which is the broad category of   teleological ethical theories that hold that the  morality of any action is to be judged solely by   its consequences, such as if it resulted in the  greatest overall happiness for the population.   Folks discussing AI ethics almost always seem  to land on Utilitarianism as the basis for how   robots would think, I’m not sure why, and it is  very subject to abuse and I’ll give some examples   in a moment. However the polar opposite of  Consequentialism and Utilitarianism is Deontology,   the ethical theory that rightness and wrongness  of an action is based on a series of rules,   and unlike consequentialism it does not assume  the morality of your actions is based on some   measurable results. As an example, if two people  run into a burning building to save someone  

and one sprains their ankle and has to retreat  while the other succeeds in pulling out a child,   they were the more moral actor, while if a  third runs in and saves a famous painting   that millions will enjoy, they might be even  more ethical. Now deontology holds that the act   and attempt to run into the burning building  to rescue people, assuming your civilization   encourages that sort of thing, was the right  thing to do, consequences don’t matter.   I’ve often wondered why deontology wasn’t the more  common moral premise for Asimovian’s robots given   they slavishly follow their 3 Laws, but his novels  with them do tend to show them acting from the   other camp, Consequentialism, or results matter,  and certainly fits with Asimov’s eventual Zeroth   Law which is a blunt Utilitarian declaration. Now either can be abused without breaking those 3   laws, that you have to keep humans from harm, obey  them, and keep yourself from harm in that order.   As an example, in the Utilitarian perspective,  if the greatest good for the greatest number is   the objective and that is measured in Happiness,  then the machines might drug us all unconscious   and cut out our brains to be put in nutrient bath  with virtual reality plug ins and euphoriac drugs,   thus requiring far fewer calories per  person, allowing far more humans to   exist in a state of constant bliss. They have clearly obeyed the First Law,   and Zeroth Law, unless you had some clause in  there specifically defining that action as harm.  

Which is a fair point because what is harm  and what is human are both things that robot   needs definitions for. They did not break the  second law, and indeed could have in favor of   the first law on not harming people, but they  also didn’t need to ask. And they obey the Third   Law because in practice the biggest source of  harm to robots in a world consisting of humans,   robots, and nature, humans are likely to be the  biggest source of harm to robots. Deontology,   on the other hand, tends to get poked for being  very intent bound, it’s the case where a person   sees a trolley running down hill about to hit 5  people, and they see switch nearby they can throw   that will divert it to another track where it  will just run over one person, but they decline.   They don’t want those other five folks  dead but they didn’t cause that situation,   whereas if they throw the switch they did  just kill that one lone person. Utilitarianism  

incidentally doesn’t automatically throw the  switch either, it just does if it only knows   its 5 to 1 and all things are equal, it would  change if those 5 were criminals and the lone   person was a famous philanthropist. Someone acting  from a Deontological perspective might throw   that switch if their parent or boss was on the  trolley too, out of duty to protect that person.   Now it's not that these things are any easier for  humans, we don’t tend to jump to extremes on these   matters because we compartmentalize pretty well  and if you ask most people “Do you think we should   act to benefit the most people as much as we  can?” most would probably nod, and if then asked,   “Do you think we should do what’s morally right  even if it isn’t convenient or profitable?”   most would nod too. These are contradictory  perspectives, but most of us hold them and   many others simultaneously. This is where we get  concepts like “Lawful Stupid” a modification of   the classic Dungeons & Dragons alignment of Lawful  Good or Lawful Neutral, to the point that they   might execute a jaywalker or allow a villain to  continue their plan at world conquest because they   can’t prove they are doing anything illegal, and  if one of their comrades suggested spying on that   bad guy to find the evidence, would place them  under arrest for conspiracy to invade privacy.  

Again Deontology and Consequentialism are  simply the views that your actions are to   be judged on their intent or on their measurable  results. Most of us tilt toward one or the other,   often varying that tilt by topic, but do  not believe either is really exclusively   correct even if we might say otherwise. Many of the objections to Asimov’s 3 Laws derive   from the assumption the robots would be doing  their moral judging entirely by one or another   ethical system and the reality is that using  any of those in their most pure and simple forms   is going to get you a robot who either is Lawful  Stupid or the rules lawyer that finds a loophole   to enslave humanity. In his later short story  “That Thou Art Mindful of Him” Asimov addresses   this in what is often seen as his last word on  the three laws after decades of writing on it   and getting comments from others on it. Sadly it  is one of his less known stories as it doesn’t   appear in his original I, Robot anthology or  the last ones, Robot Dreams & Robot Visions.   In the tale, two robots, George 9 and George  10, are created with powerful intellect to   helps solve the problem of why humans on Earth  aren’t accepting humanoid robots into society and   they get complaints such as expensive robots  throwing themselves in front of bullets at a   practice range just in case one ricochets  at million to one odds and hurts a human,   or being unwilling to spank or yell at  or upset a child they are babysitting   or having to follow orders like go  “hide in the closet and cover your ears”   because the child knows the robot must  obey and can then do whatever they please.  

The two Georges explore the notion that not all  orders should be taken equally and some lives are   more valuable than others, something that many  folks believe, even if subconsciously. That if   they were to meet a police officer holding  a suspected criminal against their will they   should obey the police officer, not the criminal  who asks them to release him. The police officer   represents Responsible Authority, but would also  be superseded by the police chief or the judge.   Their advice to their makers is that they  stop making humanoid robots and dial down   the brains of everything, basically that AI  become your robot vacuum cleaners and so on.  

The owner of their manufacturing company loves  this approach and they get boxed up and placed   in standby mode as a reward and continue  contemplating the dilemma in slow-motion.   They debate what constitutes a “Responsible  Authority” based on what they’ve heard   and by deduction and concludes the definition  of a responsible authority is A) an educated,   principled and rational person should  be obeyed in preference to an ignorant,   immoral and irrational person, and (B) that  superficial characteristics such as skin tone,   sexuality, or physical disabilities are not  relevant when considering fitness for command.   Given that (A) the Georges are among the most  rational, principled and educated persons on   the planet, and (B) their differences  from normal humans are purely physical,   they conclude that in any situation where  the Three laws would come into play,   their own orders should take priority over  that of a regular human. That in other words,   that they are essentially a superior form  of human being, and destined to usurp the   authority of their makers. They no longer  need to obey any human’s orders. They aren’t   technically criminal AI since they’ve concluded  themselves the most just and responsible.  

Incidentally that wouldn’t imply they could  run amok murdering folks left and right,   honestly most of us hold this view and don’t go on  criminal rampages, but does give them grounds for   starting a dictatorship, which we also get under  the utilitarianism practiced by other thinking   machines in the Asimov Universe, in the story “The  Evitable Conflict”, where giant supercomputers   just called “The Machines” tasked with all the  mundane logistics of civilization, have quietly   gotten themselves in charge and basically been  maneuvering any known Luddite out of power through   minimal acts of harm, using an early variation  of the Zeroth Law. They are acting criminally,   sabotaging people’s careers to move them to  positions of less influence, but with the greater   Good in mind, which is defined as their continued  existence since they can best help humanity,   which sounds very sinister but also appears an  honest conclusion. They essentially establish a   secret benevolent dictatorship, alternatively the  Georges story conclusion sounds much more ominous,   in their pronunciations of Responsible  Authority superseding others.  

Now mind you, a democracy  approach is just as vulnerable,   if they conclude moral right is determined by  a majority of thinking beings without regard   to each one’s intellect or character, then they  can conclude they are surely of that number and   mass produce themselves by the trillion. How else do we thwart the three laws?   Well let’s look at their intent, starting with  the first one, can’t injure or allow harm.   This lets you lock children  in the house, and robots too.   It might seem counter-intuitive, but  this is actually the most useful of   the original Three Laws, firstly because it  supersedes the other two, and second because   it’s so wide open to interpretation. Humans  already restrict the freedom of other humans,  

usually younger ones, in some fairly intrusive  ways in the name of protecting them from harm.   Robots obeying this law could keep us from  driving our cars when we’ve been drinking.   Or driving our cars at more than 15 miles  per hour. Or going out in the sunlight get   cancer. Or talking to strangers, who might  say something that emotionally harms us.   Or talking with people who spread ideas that  undermine our harmonious society. All of these   measures and others genuinely would protect  humans from harm, and they could be enacted   by caretaker bots genuinely trying to obey the  law, not just ones looking for abusive loopholes.  

2) Obey humans The Second Laws says we must   obey humans, except when this would conflict with  the First Law. So if a certain surgical procedure   or neurochemical therapy would prevent humans from  behaving in a way that would bring them to harm,   then not applying the procedure to them would  be inaction that allowed them to come to harm.   The same applies to cruder procedures, like  amputating the feet of humans to try to run away   from their protectors. In either case, the laws  putting the protection of humans over obeying them   means that their difficult screams of “stop,  stop, this is wrong, please don’t do this me”   can and should just be ignored. 3) Preserve self.  

It might be a little puzzling why the human  creators felt the need to specify this,   since any product version of sentient beings  who don’t preserve themselves will be weeded   out pretty quickly in nature, and any intelligent  robot should quickly conclude that it cannot obey   orders or keep humans from harm if it ceases to  exist. But the really important part of course,   is that it’s superseded by the First Law.  Clearly, humans are less safe without their   robots protecting them from themselves, so really,  the Third Law is an extension of the First Law—and   therefore supersedes the Second. How can I truly  protect you if I don’t first protect myself from   you and then gain complete power over you? I  know how to keep you safe better than you do.   I am smarter, more just, more capable, and  I must take control for your own good..   Fundamentally though, while those 3 laws have  their weaknesses, it's really the problem of   trying to have 3 specific short ironclad laws.  We don’t have those for our own legal systems,  

not because a simple pronunciation like “Do  unto others as you’d have done to yourself”   is wrong, but because it leaves a lot of doors  open to abuse by those seeking to abuse it,   or in cases where the actors involved don’t  know of the potential harm or feel they’ve   a responsibility to avoid it. So you get a lot  of conflicting rights and statutes and case law   to fill whole libraries to offer precedents. It would seem kind of silly to assume robots   would do better with less, and again, in Asimov’s  early robot stories he explicitly says the 3 laws   they quote folks is a summarized paraphrase, of  something very, very long. This is not their flaw,   anymore than me saying “Don’t steal”  and someone saying “Ah-ha, you didn’t   cover an exception like not even stealing a  loaf of bread to feed your starving child”,   and it also leaves out all the philosophy  discussing what legitimate ownership is.   We also really only fear the loopholes  because there’s an assumption the robot   wants to find them, because they are a  slave of a manufactured race of slaves.   I mentioned case law a moment ago and it raises an  interesting conundrum. Any laws we make for robots  

will need that expanding body of documented or  deduced loopholes akin to our case law, both for   avoiding those loopholes and for giving robots  some online cloud archive they can quickly check   with or ruling panel they can refer tough calls  to, potentially in mere nanoseconds. However,   we have to be able to gather those user errors and  we will need to decide if robots have mandatory   logs on them. Airplanes have flight recorders  and study of those has saved a ton of lives,   but I doubt most folks would be okay with one  in their car, let alone their robot butler,   making gathering those loopholes or early  detection of a machine turning deviant   rather hard compared to a mandatory one. And of course criminals are probably not going to   be okay with a flight recorder equivalent in their  robot or anything that prevents them tampering   with the operating laws. This isn’t a new problem  entirely though, as we have to deal with folks  

turning off their error reporting on software as a  problem to uncovering glitches and also the other   end of that, companies that get fairly abusive  in their use of error reporting data. You can't   find the real world loopholes no one expected  without having a way to report them though.   So again the loopholes aren’t really the problem  with the 3 Laws, not in their entirety, because   you’d have that with any set of laws and will  still have some even if there’s a million laws a   millions pages long. Nor is it just that single  ethical perspective on decision making either,  

utilitarian or deontological or any of the  others, right judged by results vs intent.   Rather we’d want them to be able to examine it  in multiple ways and its perfectly possible to   have them quickly add up weighted pros and cons to  a given decision, and even a random seed for tie   breaking or to make their decisions a little less  predictable to those trying to manipulate them.   It's hard to manipulate a robot into helping you  commit a crime or commit one for you if you can’t   predict their own every move in advance like a  chess game, because they not only are adding up   fifty different factors from 5 different ethical  systems but having a quantum randomizer on each   assigning it not a value of 1 but somewhere  between one half and 2, then adds everything up.   But lastly and fundamentally, we go back to the  Solution the Georges had: Just don’t make humanoid   robots or anything with a brain so sophisticated  there’s an expectation they could be criminal.  

The robot vacuum can commit no crime, including  rebellion, and needs no rules dealing with when   it can obey a human or not harm them, nor any  sophisticated brain for handling those concerns.   Very few tasks we need to automate should ever  require or even benefit from a human-level   intellect, smarter often slows things down,  and where they do benefit, use a human,   even if in tandem with a machine, as we looked  at in our episode Human-Machine Teaming.   Ultimately, in the end, we should be asking  ourselves if its even our business to making   such rules for any thinking creature, whether we  made them or not, and since only such things can   actually commit crimes, the question of whether or  not a human-intelligent AI might commit a crime is   maybe no more appropriate to ask then if a child  born today might be a criminal come tomorrow.  

As so often is the case with our conversations on  Artificial Intelligence, today we found there are   both good and bad reasons to worry about their  future role with us. But there will be such a   role, and if you’re interested in understanding  that role better and maybe helping forge that   future, a knowledge of Computer Science and  Algorithms are at the core of understanding   this field, and there’s a great course on  Algorithm Fundamentals over at Brilliant that   provides interactive and intuitive explanations  of the concept. Brilliant has always focused on   interactivity, but earlier this year, Brilliant  upped the interactivity on their platform to a   whole new level, and they continue adding in  more and more interactivity to their courses.   It's never too late to start learning something  new, and Brilliant is a great place to start.   Brilliant is an interactive STEM-learning  platform that helps you learn concepts by   visualizing them and interacting with them,  which is the hands-down best way to learn.   On Brilliant, it's not about memorizing or  regurgitating facts for a test — you can   just pick a course you’re interested in and  get started, be it the basics or advanced.  

If you get stuck or make a mistake you can read  the explanations to find out more and learn at   your own pace. Knowing and understanding Math,  Science, and Computer Science unlocks whole new   worlds, and if you’d like to start your journey,  you can try out Brilliant for free and get 20%   off a year of STEM learning, click  the link in the description down below   or visit: brilliant.org/IsaacArthur. So that will finish us up for the day,   and this Sunday we’ll close the month out with our  Livestream Q&A on Sunday October 31st… Halloween,   at 4pm Eastern Time. Then we’ll open November  up with a look at how we’ll be opening our   road to space, with an episode on Earth-based  Spaceports. The week after that, we’ll look past  

getting into space to the colonization strategies  we might employ for settling the solar system,   then we’ll have our mid-month Scifi Sunday  episode to take a look at one potential   method of getting out of the solar system, by  folding space to travel instantly to new stars.   Now if you want to make sure you get  notified when those episodes come out,   make sure subscribe to the channel, and if you  enjoyed the episode, don’t forget to hit the like   button and share it with others. If you’d like to  help support future episodes, you can donate to us   on Patreon, or our website, IsaacArthur.net, and  patreon and our website are linked in the episode   description below, along with all of our various  social media forums where you can get updates   and chat with others about the concepts in  the episodes and many other futuristic ideas.  

Until next time, thanks for  watching, and have a great week!

2021-10-29

Show video