typically learn statistical models from many users and apply them to a single user. This may render inconclusive evidence which may be right on average but not for that single individual. A new privacy risk of learning algorithms is that they can also reveal new knowledge (van Otterlo, 2013; Schwartz et al., 2013; Youyou et al., 2015; Kosinski et al., 2013), predicting personal traits from language use, Facebook like's or just a photo.77 Such algorithms obviously have effects on privacy, but certainly also transformative effects related to autonomy. A more general consequence of adaptive algorithms is that we move in the direction of "the end of code" (Tanz, 2016). In the near future, increasingly many algorithmic decision-making tasks will be learned from data, instead of hardcoded by programmers. This has consequences for society, and for people, who will more often be assigned the role of trainer, instead of programmer. Algorithms that optimize The third class of algorithms consists of algorithms that optimize, incorporate feedback, and experiment. These typically employ reward functions that represent what are good outcomes, which can be, for example, a sale in a web shop, or obtaining a new member on a social network. Reward definitions tell an algorithm what is important to focus on. For example, advertising algorithms on webpages get +1 reward for each time a user clicks on an offer. Optimization algorithms will, based on all that is known about statistical aspects and based on all data about a problem, compute the best expected solution. The most prominent system currently comes from Google's DeepMind. It combines reasoning, learning and optimization, beat the world best Go player (Metz, 2016b) and is now tackling the complex computer game Starcraft-2.78 Optimization algorithms feature two kinds of rewards. One is used by the algorithm to optimize and represents clicks, sales, or other things which are valuable. The other type are rewards for users (e.g. a sale), with the goal of nudging79 them into doing something (e.g. buying something). Manipulating users' behavior obviously has transformative effects on autonomy. Worse, just like learning algorithms, optimization works well on average and could deliver nudges to the wrong users too, which would make the outcomes discriminating and unfair. Optimization algorithms typically iterate the optimizations by experimenting with particular decisions, through interactions with the problem (see Wiering and van Otterlo, 2012). A good example are algorithms that determine the advertisements on the web: they can "try out" (experiment) with various advertisements for individual users, and use the feedback (clicking behavior) of individuals to optimize advertisement placings. So, instead of a one-pass optimization, it becomes an experimentation loop in which data is collected, decisions are made, feedback and new data is collected, and so on. Platforms with large user bases are ideal laboratories for experimentation. For example, Netflix experiments with user suggestions to optimize their rewards which are related to how much is being watched (Gomez-Uribe and Hunt, 2015). Optimization algorithms are generally used to rank things or people. In the ranked society in which we now live everything gets ranked, with examples such as Yelp, Amazon, Facebook (likes), TripAdvisor, Tinder (swiping) and OkCupid, all to find "the best" restaurant, lover, holiday trip, or book. Also in our work life, ranking and scoring becomes the norm (called: workplace monitoring80). The ultimate example is China's 2020 plan (Chin and Wong, 2016) to rank everyone in society to find out "how good a citizen are you". Scores are computed from many things ranging from school results to behavior on social media, to credit score, and combined into one overall score. The higher that score, the more privileges the citizen gets (from easier car rental and bank loans, to visa to other countries). The ethics of experimentation has many aspects (Puschmann and Bozdag, 2014). Most important here are the choice of reward function (who decides has great power) and the fact that (especially on the internet) we often do not know we are part of an experiment, and maybe we need new forms of consent. Physical manifestations A fourth class of algorithms concerns physical manifestations such as robots and sensors (internet-of-things). These algorithms go beyond the digital world and have physical presence and agency in our physical world, which may jeopardize human safety. A first manifestation is the internet-of-things (Ng and Wakenshaw, 2017) in which many appliances and gadgets get connected and where increasingly sensors are being placed everywhere81, creating data traces of once physical activities. The programmable world (Wasik, 2013) will feature all digital (and intelligent) items around us as being one giant computer (or: algorithm) that can assist us and manipulate us. For example, if your car and refrigerator and microwave could work together, they could - with the right predictions on the weather, your driving mood and speed, and possible traffic jams - have your diner perfectly cooked and warm the moment you get home from work. The ubiquity of such systems will raise ethical issues since they will be influential, but often unnoticeable. Also, privacy concerns are raised. A similar big development will be physical robots82 in our society. "A robot is a constructed system that displays both physical and mental agency, but is not alive in the biological sense" (Richards and Smart, 2016). Many types of robots exist, ranging from simple vacuum cleaners, to humanoids (with human-like appearance83 84) to robots capable of manipulating their physical environments in hospital or manufacturing situations. Robots are not yet part of our daily lives, but the literature on the ethics of robots is rich (Lichocki et al. 2011; Smart and Richards, 2016). Steinert (2014) frames the ethics of robots into four main85 categories: robots as tools (or instruments), robots as recipients of moral behavior, robots as moral actors, and robots as part of society. The difference between the first and the latter two is mainly one of responsibility. The introduction of increasing numbers of robotic agent in society (the fourth type) will also have socio-economic consequences we can only partially imagine, most obviously for work which will86 increasingly being taken (or not87) over by robots (Ford, 2013). Robots are also expected to have (ethical) impact on things like law enforcement, the military, traffic (Kirkpatrick, 2015), healthcare and even prostitution (Richardson, 2016). archives in liquid times 282 martijn van otterlo from intended archivists to intentional algivists. ethical codes for humans and machines in the archives 77 https://www.theguardian.com/technology/2017/sep/12/artificial-intelligence-face-recognition-michal- kosinski 78 https://deepmind.com/blog/deepmind-and-blizzard-release-starcraft-ii-ai-research-environment/ 79 https://en.wikipedia.org/wiki/Behavioural_Insights_Team 80 https://harpers.org/archive/2015/03/the-spy-who-fired-me/ 81 https://www.iamexpat.nl/lifestyle/lifestyle-news/hidden-cameras-dutch-advertisement-billboards-ns-train- stations-can-see-you 82 https://en.wikipedia.org/wiki/R.U.R. 83 https://en.wikipedia.org/wiki/Uncanny_valley 84 https://www.wired.com/2017/04/robots-arent-human-make/ 85 The article also includes a fifth type which refers to the influence of robots on ethics itself (meta-ethics). 86 https://www.wired.com/brandlab/2015/04/rise-machines-future-lots-robots-jobs-humans/ 87 https://www.wired.com/2017/08/robots-will-not-take-your-job/ 283

Periodiekviewer Koninklijke Vereniging van Archivarissen

Jaarboeken Stichting Archiefpublicaties | 2017 | | pagina 143