即将到来的AI黑客.docx

《即将到来的AI黑客.docx》由会员分享，可在线阅读，更多相关《即将到来的AI黑客.docx（49页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、Table of ContentsIntroduction 1Hacks and Hacking2The Ubiquity of Hacking 7Ais Hacking Us10Artificial Intelligence and Robotics 11Human-Like Ais14Robots Hacking Us18When Ais Become Hackers 21The Explainability Problem24Reward Hacking 26Ais as Natural Hackers31From Science Fiction to Reality33The Impl

2、ications of Al Hackers 36Al Hacks and Power39Defending Against Al Hackers 41Belfer Center for Science and International Affairs | Harvard Kennedy Schoolthree dimensions: speed, scale, and scope.Computer speed modifies the nature of hacks. Take a simple concept-like stock trading-and automate it. It

3、becomes something different. It may be doing the same thing it always did, but its doing it at superhuman speed. An example is high-frequency trading, something unintended and unanticipated by those who designed early markets.11 Bruce Schneier (24 Aug 2006), uWhat the terrorists want/ Schneier on Se

4、curity, s: schneier. com/blog/archives/2006/08/what_the_terror.html.12 Robert L. Leahy (15 Feb 2018), How to Think About Terrorism； Psychology Today, s:/ psycholo- gytoday /us/blog/anxiety-files/201802/how-think-about-terrorism.Scale, too. Computerization allows systems to grow much larger than they

5、 could otherwise, which changes the scale of hacking. The very notion of “too big to fail“ is a hack, allowing companies to use society as a last-ditch insurance policy against their bad decision making.Finally, scope. Computers are everywhere, affecting every aspect of our lives. This means that ne

6、w concepts in computer hacking are potentially applicable everywhere, with varying results.Not all systems are equally hackable. Complex systems with many rules are particularly vulnerable, simply because there are more possibilities for unanticipated and unintended consequences. This is certainly t

7、rue for computer systems-I ve written in the past that complexity is the worst enemy of security Bruce Schneier (19 Nov 1999), UA plea for simplicity/, Schneier on Security, s: essays/archives/1999/1 l/a_plea_for_simplicit.html.-and its also true for systems like the tax code, the financial system,

8、and Ais. Systems constrained by more flexible social norms and not by rigidly defined rules are more vulnerable to hacking, because they leave themselves more open to interpretation and therefore have more loopholes.Even so, vulnerabilities will always remain, and hacks will always be possible. In 1

9、930, the mathematician Kurt Godel proved that all mathematical systems are either incomplete or inconsistent. I believe this is true more generally. Systems will always have ambiguities or inconsistencies, and they will always be exploitable. And there will always be people who want to exploit them.

10、Ais Hacking Us 14 Paul Robinette et al. (Mar 2016), Overtrust of robots in emergency evacuation scenarios, 2016 ACM/IEEE International Conference on Human-Robot Interaction. s:a cc.gotech.edu/cilcinwags/pubs/Robi- nette-HRI-2016.pdf.Artificial Intelligence and RoboticsWe could spend pages defining A

11、L In 1968, AI pioneer Marvin Minsky defined it as “the science of making machines do things that would require intelligence if done by men：5 The US Department of Defense uses: the ability of machines to perform tasks that normally require human intelligence?916 The 1950 version of the Turing test-ca

12、lled the imitation game” in the original discussion-focused on a computer program that humans couldnt distinguish from an actual human. Marvin Minsky (ed.) (1968), Semantic Information Processing, The MIT Press. Air Force Research Lab (18 Jun 2020), Artificial intelligence. s:/7afresearchlab /techno

13、logy/artifi- cial-intelligence. Graham Oppy and David Dowe (Fall 2020), “The Turing Test, Stanford Encyclopedia of Philosophy, s:/ . For our purposes, AI is an umbrella term encompassing a broad array of decision-making technologies that simulate human thinking.One differentiation I need to make is

14、between specialized-sometimes called narrow”-AI and general AL General AI is what you see in the movies. It?s AI that can sense, think, and act in a very general and human way. If its smarter than humans, its called artificial superintelligencer Combine it with robotics and you have an android, one

15、that may look more or less like a human. The movie robots that try to destroy humanity are all general ALa science-fictional android. What makes something AI often depends onthe complexity of the tasks performed and the complexity of the environment in which those tasks are performed. The thermostat

16、 performs a very simple task that only has to take into account a very simple aspect of the environment. It doesnt even need to involve a computer. A modern digital thermostat might be able to sense who is in the room and make predictions about future heat needs based on both usage and weather forec

17、ast, as well as citywide power consumption and second-by-second energy costs. A futuristic thermostat might act like a thoughtful and caring butler, whatever that would mean in the context of adjusting the ambient temperature.I would rather avoid these definitional debates, because they largely dont

18、 matter for our purposes. In addition to decision-making, the relevant qualities of the systems Ill be discussing are autonomy, automation, and physical agency. A thermostat has limited automation and physical agency, and no autonomy. A system that predicts criminal recidivism has no physical agency

19、; it just makes recommendations to a judge. A driverless car has some of all three. R2-D2 has a lot of all three, although for some reason its designers left out English speech synthesis.Robotics also has a popular mythology and a less-flashy reality. Like AI, there are many different definitions of

20、 the term. I like robot ethicist Kate Darlings definition: physically embodied objects that can sense, think, and act on their environments through physical motionr Kate Darling (2021), The New Breed: What Our History with Animals Reveals about Our Future with Robots, Henry Holt & Co. In movies and

21、television, thats often artificial people: androids. Again, I prefer to focus on technologies that are more prosaic and near term. For our purposes, robotics is autonomy, automation, and physical agency dialed way up. Its “cyber-physical autonomy”: AI technology inside objects that can interact with

22、 the world in a direct, physical manner.Human-Like Aisn Oxford Internet Institute report from 2U19 found evidence ot bots 19 Joseph Weizenbaum (Jan 1966), ELIZA: A computer program for the study of natural language communication between man and machine/ Communications of the ACM, s: linguist238Zp36-

23、weizenabaum.pdf.n Oxford Internet Institute report from 2U19 found evidence ot bots 19 Joseph Weizenbaum (Jan 1966), ELIZA: A computer program for the study of natural language communication between man and machine/ Communications of the ACM, s: linguist238Zp36-weizenabaum.pdf. 20 James Vincent (22

24、Nov 2019), Women are more likely than men to say please to their smart speaker, Verge, . 21 Clifford Nass, Youngme Moon, and Paul Carney (31 Jul 2006), “Are people polite to computers? Responses to computer-based interviewing systems, Journal of Applied Social Psychology, sonlinelibrary.wiley. com/d

25、oi/abs/10.1111/j.l 559-1816.1999.tbOOl 42.x. 22 Youngme Moon (Mar 2000), Intimate exchanges: Using computers to elicit self-disclosure from consumers, Journal of Consumer Research, 1. 23 Alessandro Bessi and Emilio Ferrara (Nov 2016), “Social bots distort the 2016 U.S. Presidential election online d

26、iscussion,H First Monday, s:firstmondoy.org/cirticle/view/7090/5653. 24 Carole Cadwalladr (26 Feb 2017), Robert Mercer: the big data billionaire waging war on mainstream media? Guardian, s:. being used to spread propaganda in fifty countries.Samantha Bradshaw and Philip N. Howard (2019), “The global

27、 disinformation order: 2019 Global inventoryof organised social media manipulation； Computational Propaganda Research Project, 。clds/s计es/93/2019/09/Cybelroop-Rep。什19.pdf. These tended to be simple programs mindlessly repeating slogans. For example, a quarter million pro-Saudi uWe all have trust in

28、crown prince Mohammed bin Salman tweets were posted following the 2018 murder of Jamal Khashoggi. Chris Bell and Alistair Coleman (18 Oct 2018), Khashoggi: Bots feed Saudi support after disappearance, BBC News, .In 2017, the Federal Communications Commission had an online public-comment period for i

29、ts plans to repeal net neutrality. A staggering 22 million comments were received. Many of themmaybe half-were submitted using stolen identities. Jacob Kastrenakes (31 Aug 2017), The net neutrality comment period was a complete mess, Verge, . These fake comments were also crude; 1.3 million were gen

30、erated from the same template, with some words altered to make them appear unique. Jeff Kao (22 Nov 2017), “More than a million pro-repeal net neutrality comments were likely faked, Hack- ernoon, s:hackernoon /more-than-a-million-pro-repeal-net-neutrality-comments-were-likely- faked-e9f0e3ed36a6. Th

31、ey didnt stand up to even cursory scrutiny.Efforts like these will only get more sophisticated. For years, AI programs have been writing news stories about sports and finance for real news organizations like the Associated Press. Ross Miller (29 Jan 2015), APs robot journalists are writing their own

32、 stories now, Verge, s: theverge.eom/2015/l/29/7939067/ap-journalism-automation-robots-financial-reporting. Bernard Marr (29 Mar 2019), Artificial Intelligence Can Now Write Amazing ContentWhat Does That Mean For Humans?, Forbes, s: fobes /sites/bemcirdmQn720】9/03/29/Qi*tificiol-intelligence-ccin-no

33、w-write- amazing-content-what-does-that-mean-for-humans/?sh=868a50450ab0. The constrained nature of those topics made them easier for an AL Theyre now starting to write more general stories. Research projects like Open AFs GPT-3 are expanding the capabilities of what Al-driven text generation can do

34、. Tom Simonite (22 Jul 2020), “Did a person write this headline, or a machine?1 Wired, s: wired. com/story/ai-text-generator-gpt-3-learning-language-fitfully/. These systems can be fed actual facts and write true stories, but they can just as easily be fed untruths and write fake news.It doesnt take

35、 much imagination to see how AI will degrade political discourse. Already, Al-driven personas can write personalized letters to newspapers and elected officials, leave intelligible comments on news sitesing. Systems like GPT-3 will make it easy for those Ais to mine previous conversations and relate

36、d Internet content and appear knowledgeable. 31 Will Heaven (21 Jan 2020), “IBMs debating Al just for a lot closer to being a useful tool Tecoiogy Review, . 32 Max Weiss (17 Dec 2019), Deepfake bot submissions to federal public comment websites cannot be distinguished from human submissions/ Technol

37、ogy Science, s:/techscience.Org/a/2019121801/. 33 Adam Rawnsley (6 Jul 2020), “Right-wing media outlets duped by a Middle East propaganda campaign, Daily Beast, . 34 Philip Sherwell (5 Jan 2020), “Chin。uses Taiwan for Al target practice to influence elections/1 Australian. s:)y. 35 Ian Sample (13 Ja

38、n 2020), What are deepfakesand how can you spot them?M Guardian, s：A theguardian /technology/2020/jan/13/what-are-deepfakes-and-how-can-you-spot-them.Then, once in a while, the AI posts something relevant to a political issue. Maybe its an article about an Alaska healthcare worker having an allergic

39、 reaction to the COVID-19 vaccine, with a worried commentary. Or maybe its something about a recent election, or racial justice, or anything thats polarizing. One persona bot cant move public opinion, but what if there were thousands of them? Millions?This has been called computational propagandar M

40、。甘 Chessen (Sep 2017), The MADCOM Future/ Atlantic Council, wp-content/uploads/2017/09/The_MADCOM_Future_RW_0926.pdf. and will change the way we view communication. AI will make the future supply of disin- formation infinite. Renee DiResta (20 Sep 2020), “The supply of disinformation will soon be in

41、finite, Atlantic, s: theatlantic /ideas/archive/2020/09/future-propaganda-will-be-computer-generated/616400/. Persona bots will break the “notice-and-comment rulemaking process, by flooding government agencies with fake comments. They may also break community discourse.These systems will affect us a

42、t the personal level as well. Earlier I mentioned social engineering. One common hacker tactic is phishing emails that purport to be from someone theyre not, intended to convince the recipient to do something she shouldnt. Most phishing emails are generic and easily tagged as spam. The more effectiv

43、e phishing emailsthe ones that result in people and companies losing lots of money-are personalized. For example, an email that impersonates the CEO to someone in the finance department, asking for a particular wire transfer, can be particularly effective. Seena Gressin (16 May 2020), “CEO imposter

44、scams: Is the boss for real?1 Federal Trade Commission, . Voice can be even more effective. Keumars Afifi-Sebet (2 Sep 2019), Fraudsters use Al voice manipulation to steal 200,000/ IT Pro, s:. The laborious task of customizing phishing attacks could be automated by AI techniques, allowing marketers

45、to send out personalized advertisements, and phishing scammers to send out individually targeted emails.Its not that being persuaded by an AI is fundamentally more damaging than being persuaded by another human, its that Ais will be able to do it at computer speed and scale. Todays cognitive hacks a

46、re crude: a fake newspaper article designed to fool only the most gullible, or a persuasive nudge designed to affect only the most desperate. AI has the potential forRobots Hacking Usa face without any trouble. This is why even minimalist illustrations are so effective. If something has a face, then

47、 its a creature of some sort: with intentions, feelings, and everything else that comes with real-world faces. If that something speaks or, even better, converses, then we believe it has intentions, desires, and agency.Robots are no exception. Many people have social relationships with their robot v

48、acuums, even complaining when the company would offer40 Brett Frischmann and Deven Desai (29 Nov 2016), The Promise and Peril of Personalization；1 Center for Internet and Society, Stanford Law School.41Kate Darling (2021), The New Breed: What Our History with Animals Reveals About our Future with Robots,Henry Holt & Co.to replace rather than repair “their” Roomba. Ja-Young SungLan GuoRebecca E. GrinterHenrik I. Christensen (2007), ,uMy Roomba is Rambo: Intimate home appliances/ UbiComp 2007: Ubiqui