Artificial Intelligence Topic 2 PDF

Topic 2 - Artificial Intelligence **What is Artificial Intelligence** - Artificial General Intelligence: a universal algorithm possessing characteristics such as self-awareness and consciousness that can create, reason, and adapt as animals and humans do. - Is the goal simply to repro...

Topic 2 - Artificial Intelligence **What is Artificial Intelligence** - Artificial General Intelligence: a universal algorithm possessing characteristics such as self-awareness and consciousness that can create, reason, and adapt as animals and humans do. - Is the goal simply to reproduce/mimic human beings, or to surpass them by achieving 'optimal' performance towards certain outcomes? (Russell & Norvig, 2009: 5) - "The science and engineering of making intelligent machines" (McCarthy, 1955) - Machine learning is a subset of AI that *trains* machines to automatically produce outputs on unseen data - ML is used "to predict or classify with the most accuracy" (Schutt & O'Neil, 2013) - "There is no such thing as an algorithmic decision; there are only ways of seeing decisions as algorithmic" (Seaver, 2018) **Threat of Value Mis-Alignment** - "Suppose we have an AI whose only goal is to make as many paper clips as possible. The AI will realize quickly that it would be much better if there were no humans because humans might decide to switch it off. Because if humans do so, there would be fewer paper clips. Also, human bodies contain a lot of atoms that could be made into paper clips. The future that the AI would be trying to gear towards would be one in which there were a lot of paper clips but no humans" (Bostrom, 2002) **Is AI an Agent or Artefact?** - ECHR Article 7: 'No one shall be held guilty of any criminal offence on account of any act or omission which did not constitute a criminal offence under national or international law at the time when it was committed' - Who is accountable for the actions of AI? Corporate liability? Distributed responsibility? - AI acts as a proxy for mens rea **Criminal uses of AI** - 91% of cyber-attacks start with a phishing email (Bahnsen et al., 2018) - 'DeepPhish' AI automatically learns from other phishing attacks to avoid spam filters and improving success rates - 96% of 14,600 online deep fake videos involved the forging of non-consensual pornographic material (Ajder et al., 2019). - Massive potential for disinformation and propaganda - Reverse-engineer system vulnerabilities to 'hypnotise' AI systems -- without human comprehension on either side - Hackers' adversarial inputs create 'cognitive landmines' amidst the 'Internet-of-Things' systems and posing problems for the 'smart city paradigm' (Scharre and Horowitz 2018: 15) - AI as an 'independent' criminal intermediary problematises actus reus, mens rea, knowledge threshold, foreseeability, liability (King et al., 2019: 6--7) - Proliferation of out-of-the-box AI-Crime toolkits. **How The Police Conceive of AI** - "The program is analogous to laser surgery, where a trained medical doctor uses modern technology to remove tumors... Extraction of offenders takes place in a 'non-invasive' manner (no task forces or saturation patrol activities)... by extracting offenders surgically, recovery time of the neighbourhood is faster" Operation LASER, 2012 - "People don't get upset when doctors use technology to prevent illness", Ex-LAPD Chief Bill Bratton in 2015 - \"The behaviours that a hunter-gatherer uses to choose a wildebeest versus a gazelle are the same calculations a criminal uses to choose a Honda versus a Lexus"(Brantingham, 2010) **A Brief History of Predictive Policing** - Chicago School (Shaw and McKay, 1942); Routine Activities Theory (Cohen and Felson, 1979); Environmental criminology's Crime Pattern Theory (Brantingham and Brantingham, 1984) - "Problem-oriented policing" (Goldstein, 1979), aimed to identify and resolve problems directly connected to increasing crime risks, mostly in areas characterized by high levels of crime (the so-called "hot-spots") - "Evidence-based policing" (Sherman 1998) empirical evidence is systematically used to formalize guidelines and evaluate agencies and officers (1998). Era of crime mapping platforms, computerized data-recording systems such as CompStat, a means to cope with consistent budget cuts - In 2008 LAPD Chief William Bratton (who implemented broken window policing and CompStat in Guiliani's NY), along with the National Institute of Justice sponsored the benefits of predictive policing **Predictive Policing Today** - [Predpol]: ML uses crime type, location, and time of the crime, to assess what areas (and in what hours) are highest risk. 'Self-exciting point process' originally developed for predicting earthquakes; crime conceived as concentrated in space and time, and creates aftershocks - [Risk Terrain Modelling]: spatial characterization of a specific area is crucial to identify places at high risk of crime distribution; favours of user-friendly data models based on raster and spatial layers - [HunchLab]: adds weather info, major events, contextual variables, to forecast high risk areas; tailors policing strategies for different types of crime and communities; prioritises crimes more effectively addressed by direct police intervention Criminals/victims follow common life patterns; overlaps in patterns increase likelihood of crime - Geographical/temporal features influence the where and when of those patterns - Criminals make rational choices about whether to commit crimes, considering the area, target suitability, risk of detection. **Operation Laser** - LAPD use Palantir to run automated profiling, hotspot maps, on data from the surveillance inputs- CCTV, Automatic License Plate Readers (ALPR), Body-Worn Video, Stingrays and Digital Receiver Technology, Homeless Management Information System, Field Interview cards; along with data from other government agencies as well as commercial brokers - Identifies friends, relatives, colleagues pulling people into the LAPD's surveillance system who otherwise wouldn't have been, track vehicles using ALPR, and examines social networks - Officers target using attributes like race, gender, physical features, recipient phone number, date, duration of the call, latitude and longitude of cell towers LAPD recently launched Area Crime and Community Intelligence Centres, hyper-local spy garrisons produce daily mission maps, perform social media monitoring and surveillance camera video pulls. ACCICs produce daily hotspot maps marking locations as 'gang related', 'gang member', 'suspect', 'gang member' 'victim', 'transient-victim', 'transient-suspect', 'domestic violence'. Diagram, schematic Description automatically generated  A close-up of a map Description automatically generated Quételet's (1836) average man theory, a form of "social physics": "an individual who, within themselves, in a given era, sums up all the qualities of the average man, would represent all that is great, beautiful and good"... "such an identity can hardly occur, and it is generally within humans' reach to resemble this type of perfection only by a greater or lesser number of its dimensions"