The Long Road Toward Tracking the Trackers and De-biasing : A Consensus on Shaking the Black Box and Freeing From Bias

Automated decision making is both promising and threatening. Processing the biggest data possible may lead to societal advances but also violate human rights. There is, then, an acute need to protect individuals without impeding major benefits. Non-human agents may be biased; and they may not lend themselves to easy explanations. Instead of focusing on interpreting models, there seems to be a shift toward a concept of risk assessments. Opaque systems are aimed at predicting, or forecasting, future situations. This challenges human values and ethical principles. Even though incorporating ethics in machines is an old subject of legal discussion, consensus has not yet been reached; for theories and values may be controversial. This paper examines whether there could be an agreement on fundamental principles. A commonly understood basis could allow for fair and proportionate mechanisms to address crucial aspects of partiality and opacity in automated decision making. It could trigger a shift toward a concept of ̳tracking the trackers‘ and a discussion on a ̳right to an unbiased decision maker‘.


Introduction (To Automated Decision Making -In Health and Elsewhere)
Since the middle of the 20 th century, the use of archives of human specimens, collected in the course of healthcare, has led to important advances in medical knowledge.These advances have, in turn, produced crucial benefits for human kind (Korn, 2000).Although automated medical decision making is not new (Gleser & Collen, 1972: 180-181), the speed of information growth is said to have exceeded Moore's law (Chen & Zhang, 2014: 314, 318, 320) and its value is believed to have exceeded oil's (The Economist, 2017: 7).Therefore, large amounts of data are being processed and another ‗V', for Value, is attributed to Big Data (Chartier, 2016).
Despite success stories in medical research (Gonzá lez-Ballester, 2018), personal data concerning health constitute ‗special categories' of information and merit specific protection (GDPR, 2016: Article 9(1)).Namely, the genomic sequence can reveal sensitive data, like ancestry, one's ability to metabolize drugs or diseases she might develop (Azencott, 2018).Hence, there is a growing need to protect privacy without impeding scientific and societal advances.
Ethical principles for medical research demand -amongst others-informed consent, a sine qua non condition, and confidentiality (World Medical Association, 1964).Informed consent is a challenge.In many cases, physicians need to share patient's data to draw ‗everyday' conclusions; it might be unreasonable to obtain informed consent each time data were transmitted between professionals to make a diagnosis.Thus, it could be argued that, in some cases, when confidentiality is maintained, data processing without the data subject's consent constitutes a theoretical harm; it may not be a real danger (Capron, 1991: 189).
But this could be an exception.Healthcare is one among the many fields where non-human agents perform their role as decision makers; and health-related data, like genome, are some among the many sources that can reveal sensitive information.For instance, consumers' buying history can predict health status (Robinson, Yu, & Rieke, 2014: 6).Contemporary algorithms (or ‗traditional' dating from the ‗70s) make automated decisions of utmost importance in any domain (Veale & Edwards, 2018: 399;Edwards & Veale, 2017: 25).They do so, camouflaged by the ambiguity and technical jargon of a firm's terms of service; such terms may not guarantee confidentiality.Needless to say, informed consent (Papadimos et al., 2018(Papadimos et al., : 1845) ) is rarely met; it is often obtained via rental-car-contracts-like forms, which make it certain that people will not read them (Hilts, 1995).A number of cases (Lipworth, Mason, & Kerridge, 2017: 486) and numerous jokes (Chesterman, 2017) can prove such claims.This raises tremendous concerns; a system may deny loans and visas, target citizens for scrutiny (Kroll et al., 2017) and anorexics with emetics (Veale & Edwards, 2018: 401) or score customers, job applicants and criminals (Chander, 2017).
Yet, major benefits are still being promised.Eco-innovative technologies, like smart plant systems (Smart-Plant, 2018), can foster efficiency, enhance green economy and liberate consumers from household chores (European Commission, 2013).Speech to text applications can simplify life and help people with special needs (Fearn, 2018).Smart lights can optimize energy use and guarantee sustainability (Ritchie, Kessler, & Sargent, 2018).Real-time price monitoring and tracking can offer consumers the best deal (Competitoor, 2018;Prisync, 2018;Competitor Monitor, 2018).The list goes on and it seems that applications can be endless.
The ethos has become ‗the more data, the better' (Citron & Gray, 2013: 262) and juggernaut-like firms process the biggest data possible (Dixon, 2018).Societal goals can be -and are being-achieved, but also human rights may be -and are being-threatened.Therefore, the challenge seems to be straightforward: how to protect fundamental rights without impeding societal, scientific or other advances of major importance (Žliobaitė, 2017: 1065).
This paper focuses on automated decision making.Like human minds, machines may be biased.Even though justifying decisions could be a way to diagnose bias and avoid partiality, non-human agents do not always lend themselves to easy explanations.Further uncertainties can occur, for people, forced by the will to know, tend to work against future uncertainty.They focus on a notion of risk assessments; opaque and unintelligible systems are aimed at forecasting future scenarios.A general shift toward regimes of control-by-risk challenges human values and ethical principles.Although humans, assisted by technologies, could perform a teaching role and embed such principles within systems, debates on controversial values render ‗agreement on ethics' impossible.A perfect ethical theory to address every case and scenario may never come, albeit, a free and willing consensus on fundamental principles could be achieved.Fair and proportionate strategies, built upon a commonly understood basis, could address important aspects of opacity and partiality; they could lead the way toward a notion of ‗tracking the trackers' and a potential ‗right to an unbiased decision maker'.

Algorithmic Partiality
Decision theory may be concerned with identifying the best decision to make based on the assumption that the decision maker is fully rational -or bounded rational (Wang, Xu, Hamido, & Liu, 2016).It can deal with the question of how an agent may rationally choose among available courses of action (Cooper, 2003: 43).Moreover, decision-making can be regarded as a process that leads to a certain choice.It can be seen as a procedure that comprises the phases of finding occasions for making a decision, finding possible courses of action and, finally, choosing among courses of action (Simon, 1960: 1).And Big Data involve(s) a number of steps, starting from finding instances (such as data acquisition, collection, capture or storage) and resulting in a certain choice, a decision (Janssen, Voort, & Wahyudi, 2017: 339).Therefore, decision making (to some, rational and, to others, biased) may exist in many procedures of Big Data, from collection, compilation or consolidation to data mining and use (FTC, 2016: 3;Wang et al., 2016: 748;Voort, Klievink, Arnaboldi, & Meijer, 2018;Chen, Chiang, & Storey, 2012;O'Malley, 2014).
Even though machines do make decisions based solely on automated processing (Geist, 2016), human decision makers have not yet become has-beens.Systems, such as the legal one in some jurisdictions (Harvard Law Review, 2017;Katz, 2013), may be based solely on human judgment.But people are not perfect.Although human imperfection is said to be a core dimension of freedom (Benkler, 2013), in many fields, like healthcare, some accuracy would be needed to benefit society as a whole.
Human decision making has not yet been fully understood.It is not very clear what influences people's decisions.To some, people are self-interested and, thus, make decisions that provide the greatest benefit or satisfaction (Smith, 2005).From this perspective, decision making could be simple; one may perform a cost-benefit analysis and pick the option that would maximize the benefits and minimize the costs.Yet, as imperfect (but free) beings, humans do not always seek to maximize benefits; ‗prisoner's dilemma' has proved that they are not always rational (Bravetti & Padilla, 2018).Therefore, human decision making process can depend on several factors and the outcome may not be predicted by a mere cost-benefit analysis.Biases, like imperfect information, processing capabilities, norms or resistance to change, could limit rational decision making (Kahneman, 2011: 255, 417, 430).Thus, people's brain, with its well-hidden bias, may not be transparent; one does not always know how she, herself, makes decisions.As some have aptly put it, it would be universal madness to think that one generally knows what she is doing (Katsafanas, 2018: 123-124).
Biases in people's decisions call for objectivity; and some claim that algorithms could fulfil this need.It has been argued that machines are different from human minds (Freitas, 2012;Vogel, 2008) and that data driven systems have the potential of avoiding biases (Rudin, 2016).According to this argument, systems could be less subjected to biases and, thus, come up with more accurate decisions (Rudin, 2016).Rationality and its instances, being captured in mathematical terms via, among others, game theory (Cooper, 2003: 14), could be involved.If this were the case, some levels of objectivity could be met (Kumar, 2017: 11).

Explaining Automated Decision Making: A Failure?
People may use arguments to justify their claims or beliefs.For instance, a lawyer can present her claim at the trial, describe evidence to support it, detect laws that apply, call experts or refer to previous cases.Similarly, a physician may use arguments to justify a diagnosis.
But could machines use explanations to justify their conclusions?Explanation techniques have long been proposed (Shankar & Musen, 1999).Explaining why a car stopped when the system detected a red light might not be that meaningful.However, knowing how the machine reached an outcome, when deciding whether one's life should be terminated, would be important.
A good example can be found in Ms B v. An NHS Hospital Trust (EWHC, 2002).A 43-year-old woman suffered from haemorrhage into the spinal column in her neck.She became paralyzed from neck and down and unable to breathe without ventilator support.After unsuccessful treatment, she requested that the ventilator be switched off.Yet, physicians felt they were being asked to kill her and refused (EWHC, 2002: 57).Ms B exercised her right to refuse treatment and brought her case to court.The ventilator ceased and Ms B died (Slowther, 2002).
What is of great interest is that physicians claimed the issue of desirability of her survival as ethical in nature (Foster & Miola, 2015: 509).Were an algorithm in charge, it might not see such human point of view (Taylor, 2015: 514); claim could be different.But, whatever the agent's thoughts and biases, it would be meaningful and important to know and understand its decision making.Same would be the case, when deciding if one should try a specific medical treatment or whether one would be able to pay for a loan or be eligible for parole (Angwin, Larson, Mattu, & Kirchner, 2016;O'Neil, 2016).Yet, explanations may not be possible.
Namely, there could be obstacles of legal nature.Firms would not reveal their secrets (Reddix-Smalls, 2011).A non-human agent could be secret and its controller could have undertaken steps required to keep it secret; and, virtually, any system may impart value to its owner as a consequence of this very secrecy (Jandoli & Dani: 46).Therefore, trade secrecy laws can apply (Directive 2016/943: Article 2(1)); and they cover a wide array of information, including drawings or calculations relating to, inter alia, processes (Jandoli & Dani: 49).It seems that such information, well hidden in firms' back-offices and guarded like the Coca-Cola formula (Harcourt, 2005: 5), would not be easily subjected to audit or review (Reddix-Smalls, 2011).Besides, the -recent but underdiscussed-European Directive on trade secrets does not allow for exceptions as regards researchers -who could scrutinize processes.It only excepts journalists (Directive 2016/943: Articles 1(2)(a), 5(a); Recital (19)).But journalists cannot decipher.
Moreover, explanations could be impossible due to lack of knowledge.Even if firms revealed their secrets, people would lack literacy to comprehend relevant procedures.What Bhargava et al. (2015) describe as the desire and ability to engage constructively in society through and with data, the literacy that enables individuals to understand principles and challenges of data, is most probably missing.Were decisions explained, people might not understand.
In addition to lack of knowledge, there is also asymmetry of knowledge, an information advantage enjoyed by machines (Matthias, 2004: 182).If people had skills, the very system might not lend itself to easy explanations.There is lack of interpretability (Lisboa, 2013) and experts may not be capable of comprehending how an outcome was reached (Edwards & Veale, 2018: 49).
Besides, if machines were ‗generous' and lent themselves to easy explanations, this would not per se mean de-biasing.Biases can arise for a number of reasons.They may exist independently of and prior to the creation of a system; they may emerge from resolving issues in design; or, most importantly, they can arise in a context of use, as a result of altering knowledge, population or values (Friedman & Nissenbaum, 1996: 333).Data collected may have been preferentially sampled (Crawford, 2013); they may be inaccurate, failing to represent reality; they can reflect human culture (Caliskan, Bryson, & Narayanan, 2017: 183); they may be hidden in the code; or they can emerge due to a shift in context (Friedman & Nissenbaum, 1996: 335).
Therefore, diagnosing and avoiding bias may be impossible (Veale & Binns, 2017).And it seems that challenges go beyond failure to de-bias.Tempted by the will to know, humans tend to focus on predicting technologies.This triggers a shift toward a concept of risk assessments that may, in turn, result in further uncertainties.

Working Against Uncertainty
Non-human agents tend to diffuse information in ways difficult to decipher.They may, then, be as opaque as human minds (Castelvecchi, 2016).To many authors, the problem lies in the black box function of AI (Pasquale, 2015).Although transparency has been regarded as the key to audit or monitor algorithms (Allnutt, Dickson, & Webster, 2016), the question may not be ‗how to whiten the black box'.Rather, it may be ‗what would transparency mean, when what comes into light, what is rendered transparent, escapes comprehension?'.In other words, the very value, or even usefulness, of transparency may be challenged, when what is under scrutiny is being updated; it is a later version of what was supposed to become transparent.
It might take some time for people to understand what algorithms can perform in seconds.Humans lack both time and skills; and there is a tremendous discrepancy between cyberspace and cybertime, an incompatibility of space with time in cyber-terms.As Berardi (2014: 242) has put it, the expansion of cyberspace is boundless, albeit, cybertime is not infinitely stretchable; for it is composed of time of attention and subjected to the latter's physical limitations.
But, in human-terms, time has been understood as a dimension where the past is, now, known and the future is, presently, unknown.Future is something that will be, that is not here.It portrays uncertainty (Esposito, 2015: 95).Working with the idea that the future is unknown can allow preparing for the future (Langley, 2013: 55, 56, 60, 69).Working against the notion of uncertainty can allow for techniques aimed at predicting the future (Esposito, 2015: 99).
If the beauty of the future is in its uncertainty (Esposito, 2015: 106), it seems that, today, future is aimed at becoming ugly.For, now, data may be used with a view to predicting future situations (Pearsall, 2010: 16).Yet, with the objective (or instead) of predicting, agents may render future uncertainties actionable today.If future can be seen as the performative consequence of actors and actions (Latour, 1988: 165), algorithms may produce a ‗desired' future.Worse, they can reproduce past -and its injustices.
To oversimplify, let us assume that a system collects data from areas ‗A' and ‗B'.In 2018, 500 crimes were totally committed in area ‗A'; 400 of them were committed by whites; and 100 by colored.But only the latter 100 were recorded.
Here, the ‗unfair past' would be in the belief that colored people -whose crimes were recorded-are more dangerous than whites -about whom data are missing.The ‗desired' (yet unfair) future would be in focusing on colored; for data recorded suggest that they be high risk.
Let us further assume that in area ‗B' 10,000 crimes were committed in 2018.All of them were committed by whites, but no crime was recorded.In the absence of data, ‗B' would be thought of as a safe place to live in.
The computer science adage goes ‗garbage in, garbage out' (Vayena, Blasimme, & Cohen, 2018).The system spits back what was put into it (Ferguson, 2017b: 47;Llenas, 2014).Even though more crimes were committed in ‗B', it will consider area ‗A' as high risk.Although whites committed more crimes, colored people will be scored as dangerous.The algorithm may, then, suggest that police forces draw attention to area ‗A' and the colored.This way, non-human agents not only perpetuate existing injustice, but also create new; a ‗desired', and unfair, future.This scenario is far from being ethically neutral.But it is real and happening.
The development of technical knowledge in predicting technologies, but also its driving force, i.e. the will to know, triggered the shift in people's conception of just punishment from notions of reform or rehabilitation to concepts of risk assessment (Harcourt, 2005: 32).Therefore, people have come to believe that it is just for punishment to primarily relate to the statistical probability of reoffending (Harcourt, 2005: 33).Yet, biased inputs and outputs transform ‗predictive policing' into ‗crime forecasting'.Like weather forecasting systems, these agents may become more accurate (Ferguson, 2017a(Ferguson, : 1137(Ferguson, , 1144)).But this focus on (potential) utility seems to ignore fairness and justice; it prioritizes predictive performance over interpretability (Edwards & Veale, 2017: 26).This work against uncertainty, occurring in various areas, allows for a general shift toward risk assessments, ‗razor sharp segmentation games' or a regime of control-by-risk (Poon, 2009: 656-659).
‗Slowing down' algorithms would come counter to what they are being created for.Systems need to be ‗fast' to provide benefits.As Mander (1978: 43-45) has put it, if one accepts the existence of cars, she also accepts the existence of roads laid upon landscape, oil to run the cars or firms to pump and sell oil.In the same vein, one could argue that, if people accept algorithmic speed, they also accept inability to control; accepting automated decision making would, then, mean accepting both its benefits and its risks.But there may be values higher than speed and efficiency (Perry, 1977: 412).So, could there be another way?
It seems impossible to understand a system that remains unintelligible, even when transparent.These agents consume more and more data to become even more powerful; ever-growing, yet invisible to humans, they are black holes, rather than boxes.In this arena, the unpredictable can be forecasted.People may be ‗punished' for behaviors they may never express, mistreated for entirely internal attitudes they may never manifest or denied for future scenarios that are not known and may never occur.
Being colored and living in area ‗A' (of the above scenario) does not necessarily mean one is dangerous; nor does it necessarily mean that police forces should focus on her.Same can be the case with other domains where algorithms perform their forecasting role.Namely, in health, the mere possession of a deficiency or a trait does not necessarily mean one will later become ill or disabled (Matthewman, 1984(Matthewman, : 1188)).Denying loans, jobs, paroles and any opportunity because in some years an individual may, but also may not, die from smoking, alcohol, cancer or junk food could be unfair or unreasonable.
Laws can transform messy phenomena into ordered, knowable and calculable cases (Dove & Özdemir, 2015: 530).They may regulate, among many others, activities (Friedman, 2016), behaviors (Kim, 2018), markets (Mugarura, 2016: 603, 605), populations of living species (United Nations, 1992c) or the management of the dead (Gaggioli, 2018).They can regulate expressions, like actions through which a state of mind, a belief, a mood, an emotion, an attitude or an idea may manifest itself (Anderson & Pildes, 2000: 1506).If one, holding the belief that ‗war is wrong', attacks an embassy for a nation's involvement in war, he can be punished for attacking, for expressing his belief in such way.But he cannot be punished for his very belief that ‗war is wrong'.Laws (in jurisdictions we know) do not regulate -nor do they punish people for-entirely internal thoughts or ideas.What regulators left unregulated, people's deepest wish or fear (Battelle, 2005: 1-9;Tene, 2007), is increasingly becoming the ‗input', the reason for denial or unfair treatment.
It seems that the attempt to eliminate and work against uncertainty of the future is alarmingly diminishing certainty of today, or people's certainty about themselves.It is challenging people's values, what it means to be transparent or fair.It is, then, challenging ethics.

Ethics in Machines? A Need for Consensus
Analyzing algorithmic processing without discussing ethics may be naï ve and, to some, deceptive (Mutlu, 2015(Mutlu, : 1002)).
Ethics refers to values and ways to define right and wrong actions (O'Leary, 2017: 14).Dealing with duty, what people ought to do (Merrill, 2011: 3), it can be understood as the study of one's responsibility to discern the rightness or wrongness of her actions; the art or science of one's efforts to live in proper harmonious relationship with others (Kruckeberg, 1989: 11).As such, ethics involves thoughts designed to bring about good within a society (Ray, 1996: 48).
Decades ago, some principles were suggested to -bring about such good within information and technology societies and-promote, among others, transparency or oversight (U.S.Department of Health, Education and Welfare, 1973).But consensus has not yet been achieved; over the years, several legal instruments have proposed a number -each one its own number-of principles (OECD, 1980;A29WP, 1998;FTC, 2000).The failure to implement these recommendations may have been due to their legalistic nature (Cate, 2006: 343, 355, 366).However, regardless of their nature, the very principles these recommendations suggested could be seen as the building blocks of information law (Schwartz, 1999(Schwartz, : 1670)).Therefore, were consensus achieved, it would be desirable that ethical principles (or ethical features; Kraemer, Overveld, & Peterson, 2011: 252) be embedded within design specifications of technologies (Dove, Knoppers, & Zawati, 2014).
The idea of incorporating principles in design has been developed in a number of frameworks, such as the ‗Appropriate Technology' (Schumacher, 1973), the ‗Participatory Design' (Asaro, 2000) or the ‗Value Sensitive Design' (Manders-Huits, 2011;Friedman & Kahn, 2003: 1186).Design choices have historically had ethical import.Moses's bridges are a good example.In 1930, Robert Moses was asked to design overpasses for the highway connecting New York to Long Island and Jones Beach (Winner, 1980).Moses designed very low bridges; cars could pass, albeit, the twelve-foot tall buses (that is, public transport oft-used by minorities) could not get through the overpasses (Winner, 1980: 124).Regardless of whether low bridges were intentionally designed to prevent minorities from getting to the beach, Moses and his bridges are often quoted to exemplify potential ethical (or even political) import of design choices (Manders-Huits, 2011: 272-273).
Yet, design is just one among the many shape-defining aspects of technology (Flanagan, Howe, & Nissenbaum, 2008: 330).Other crucial steps may include, to name but a few, production, deployment, adoption or distribution.For instance, in case of Stallman's (2010) ‗free as in freedom' concept, it could be fair to argue that what determines free software's shape, its magnitude, and what is of utmost importance (without prejudicing its design) is in its use, adoption and (re)distribution, as described in Stallman's (2015: 3) freedoms zero to three (to run a program for any purpose; to study and change it; to redistribute copies of it; and to distribute copies of one's modified versions).
A concept that seems to focus on multiple steps of technology, its whole life cycle, is privacy-by-design.Even though its ‗name' refers to design, its interpretation, as provided in small print, footnotes, soft laws, press releases (European Commission, 2010: 17;2015;EDPS, 2018: 4, 10) or literature (Diver & Schafer, 2017), suggests that the relevant principle (here, the protection of privacy) be embedded throughout the entire life cycle of technologies, from the early design stage to their deployment, use and ultimate disposal.Therefore, it would be reasonable and fair to suggest that ethical principles be promoted and incorporated in the many shape-determining steps of technologies.This could resemble ways people teach ethics.This idea of human teaching could be supported by non-humans; for it is argued that, when people and machines cooperate, results can be better (Kraft, 1964: 100).Or, to cite Katz (2013: 929), the equation could be simple: Humans AND Machines > Humans OR Machines.
Coined some decades ago (Samuel, 1959) and similar to human learning (Informatics Europe & EUACM, 2018: 6), machine learning can make AI respond in different situations (Jordan & Mitchell, 2015;Veale & Binns, 2017;Hastie, Tibshirani, & Friedman, 2009).An agent may, then, become better capable of reaching an outcome.For instance, in fields of healthcare, the ethical goal might be the prevention of a disease.Hence, AI could learn the settings that would be the most effective to prevent, e.g., a possible outbreak of a disease.A disease may have multiple symptoms and similar or same symptoms can be attributable to many diseases (Malmir, Amini, & Chang, 2017: 95).This could confuse even the most experienced physicians (Adlassnig, 1986;Conejar & Kim, 2014).AI could correlate data (King & Mrkonich, 2016;Mayer-Schonberger & Cukier, 2014: 68;Podesta, Pritzker, Moniz, Holdren, & Zients, 2014) and provide patterns that human physicians would never reach.Machines could, then, exclude unrelated symptoms, deliver a more accurate decision and, hence, reach the ethical outcome (this, of course, is a mere example to simplify ways in which machines may learn; it does not take -nor does it aim to take-into consideration further criteria, such as impartiality of input data, that would need to be met to reach accurate outcomes).
Yet, some have questioned this idea of teaching or learning ethics (Lorenc, 2015).It may be true that humans are not always the best teachers; for selflessness and bringing about good may not be what people desire.For example, few might want to buy a smart car that would look after others and prioritize general good over the driver's own well-being (Metz, 2016;Bonnefon, Shariff, & Rahwan, 2016).Were such preferences taught, machines would acquire, not ethical but, common behavior.Therefore, if teaching/learning ethics were the answer, the question would remain ‗which principles should be implemented?'.
No agreement has been reached (Ware, 2015) and contemporary dilemmas seek for an answer (Birnbacher & Birnbacher, 2017).Should a machine maximize happiness by terminating the life of an old man to save ten young women?What if the old man was a genius lawyer and the ten were criminals?Disagreements can emerge for ethical principles may be controversial; each one may have exceptions or counter examples (Friedman & Kahn, 2003: 1178).For instance, utilitarianism, an influential version of consequentialism, seeks for the greater good (Harris, 2008(Harris, /2009: 68): 68).But it may permit actions that some would consider unjust, unethical or unfair.The ends could justify the means and the means could get ugly (Grau, 2006: 52).Furthermore, by focusing on greater good, utilitarianism is said to ignore individual identity, to not take seriously the distinction between persons (Rawls, 2009: 24).In contrast, deontologists focus on the very action; the right action may not be the one that maximizes good.There could be constraints on what one may do, some instances in which maximizing the good may not be right (Rawls, 2009).To commentators, deontological positions are not always clear; for they do not state what these instances are (Harris, 2008(Harris, /2009: 69): 69).It is further argued that there is no consensus on the very notion.Some deontologists suggest ‗requirement', ‗permission' or ‗prohibition' as basic deontic terms, albeit, others opt for ‗obligation' or ‗blameworthiness' (Nottelmann, 2008: 325-327).
Further disagreements can arise when theorists attribute values to entities (Davidson, 2013).For instance, a knife may have utilitarian value; people use knives to eat.But it can also have intrinsic value; it may have been handed down to an individual from her parents.Therefore, she might need to care for what the knife is, not for what the knife does.Moreover, a ten-year-old kid, doing chores, might have some kind of utilitarian value; and the kid has intrinsic value, even if she does nothing.
Therefore, one could claim that entities can have utilitarian but also intrinsic value.Yet, others argue that there is either utilitarian or intrinsic value, either price or dignity, not both.According to this argument, if something has price, it can be substituted; if it has dignity, it cannot (Wood, 2018;Vöneky & Wolfrum, 2013: 96).From this perspective, humans have dignity, not price.But still, things are not always clear.Namely, governments are said to assign economic value to people (Davidson, 2013: 174) and firms claim they can measure the economic value of personal data (Malgieri & Custers, 2018).
Many theories and perspectives can lead to endless debates.Were humans the teachers, the ones who would embed ethical principles within systems, uncertainties might occur.Same would be the case with machine learning.Enabling AI to learn and perform ethical actions to reach desired goals would also require some consensus.Humans would need to agree not only on situations to which agents would be exposed, on which they would be trained, but also on the very desired ethical actions and goals.
Instructing machines to do things at which people are bad or things that require some intelligence -in narrow tasksmay not be that ‗hard'.Namely, with chess (Dickson, 2017;Bostrom, 2014), the machine has to win; people know what winning looks like (Metz, 2016).Yet, aiming to program things that come more naturally to humans, but not to machines, is admittedly challenging (Deng, 2015: 26).
Scientists estimate that AI singularity may soon create its own ethics.Even though these are experts estimating, not dreamers dreaming (Hauer, 2018: 104, 106), such estimates have been regarded as the ‗rapture of the nerds' (Popper, 2012).In any case, machines have not yet become super-intelligent (Bostrom, 2014).Nor are they truly autonomous; they will be the day they are instructed to go to work and they instead go to the beach (Markoff, 2015: 333).But algorithms collect and use data to make decisions and guide their ‗actions' in line with their goals.This way, they may effectively display some capacity for choice and intentional actions, an almost or quasi free will (Malle, 2016: 249).Occupying the position of subject in sentences, they go beyond chess or Go (Silver et al., 2017); they intervene in real life and are embedded within larger social systems (Ziewitz, 2016: 5, 7).They, thus, carry some form of what humans used to monopolize; agentic power (Beer, 2017: 5).
Seeking for an object, some principles to agree on, in the teaching/learning argument and revisiting the dilemma ‗should a machine maximize good by killing an old man to save ten young women', one could argue for utilitarianism.To many, machines could make felicitous correlations to ‗calculate greater good'.Yet, such ‗good' may have many interpretations (Emanuel & Wertheimer, 2006: 854); and deontologists might be right: killing may not be ethical.
Providing rules is not always for solving dilemmas; in some cases, giving ethical rules for solving dilemmas could be like giving rules for telling jokes (Wijnberg, 2000: 333).Ethical battles may never result in the perfect theory that will provide an explicit list of all desired principles, their exceptions, constraints or instances.
But this cannot be a reason to give up on trying to do what we believe is ethical, right, fair or just.Nor can this be an excuse to stop teaching ethics to our kids.For, whichever theory is correct, there seems to be some consensus, an agreement on some form of responsibility to make the world a better place, to bring about good or, at least, to do no harm (Otto, 2007: 20;Hunt, 2014).We may lack complete knowledge and may never discover the perfect ethical theory, the perfect fit for each case and scenario; we may, thus, not know exactly how to advise our children.But we do (Sutton, 2010).Because they are developing and they have to develop in some ethical way.They cannot just ‗exist'.Nor can they simply ‗act and behave'.For their actions will affect the world; and the world needs to get better; harm needs to be avoided.
It could be argued that, unlike children, machines cannot have ethics.But, like children, they seem to be developing.Like kids, they seem to have the potential to ‗act and behave' in ways that affect the world in domains of major ethical, societal or economic significance, from healthcare and education to trade, business, entertainment or security (Malle, Scheutz, Forlizzi, & Voiklis, 2016: 125;Bendel, 2016: 104, 106, 107).These developing beings can influence and shape people's decisions and lives; they can provide major benefits, but also threaten human rights.Therefore, they cannot simply ‗exist'.
There has to be some kind of consensus on responsibility.For, even though machines have been referred to as thoughtless (Meyer, 2014), those who craft them (when humans) are (expected to be) thoughtful.

Toward a Free and Willing Consensus
Each culture may have its own laws, norms, god or time.Diversity is widely acknowledged (Melé & Sá nchez-Runde, 2013: 682), not as a condition that will pass away but, as a permanent and desired feature of democracies and, in general, societies, communities, groups or families and their members.
But, in a number of cases, needs have to be satisfied.When shared, such needs can lead to agreements.Namely, people drafted agreements on coffee or sugar (United Nations, 1968;International Coffee Organization, 2007;United Nations, 1992b), for they recognized the importance of such goods to economies that are dependent upon them for the achievement of social goals (International Coffee Organization, 2007: 1).Furthermore, situations, involving unpleasant or extremely cruel acts and behaviors, can result in common needs and, thus, agreements.For example, barbarous acts created the need for common understanding of human rights.That is, an acute need that resulted in agreement on freedoms, immunities and benefits that everyone should enjoy in the society in which she lives (Skinner-Thompson, 2018: 10.05[A]; Universal Declaration of Human Rights, 1948).Humanity had to develop on the common basis that some fundamental rights are not created, but deriving from human dignity (United Nations, 1966b).
It was not only sugar, coffee or the human rights.Having so far won the struggle for survival, human beings share common adaptive characteristics (Remoff, 2016: 884, 885) or the intrinsic need to adapt to changing environments.Therefore, agreements have been reached on the Ozone Layer (United Nations, 1985; United Nations Environment Programme, 2018), climate change (United Nations, 1992a), desertification (United Nations, 1994), health (United Nations, 1946) and a number of other issues (United Nations, 2018): from diplomatic relations to the prevention of genocide (United Nations, 1961;1948); from economic, social, cultural, civil and political rights to rights of children, workers, refugees and people with special needs (United Nations, 1966a;1966b;1989;1990;1951;2006); from opium smoking to traffic in women and children (United Nations, 1931/1946;1921); and from trade, transport and communications to education (United Nations, 1947;1949a;2005;1949b).Such consensus, satisfying common needs that derive from external factors or the human nature itself, has allowed humans to survive and thrive, to cooperate and become better capable of getting things successfully done.Shared needs led to shared views; shared views, then, performed the vital function of enabling communications to proceed on a commonly understood basis (Anderson & Pildes, 2000: 1518).
And it seems that views could be shared in the arena of automated decision making.
Like coffee and sugar, algorithmic procedures are crucial; they are important not just to economies but to people's lives.This importance could be commonly acknowledged, for people depend on non-human agents to achieve societal advances.
Furthermore, algo-harms may not involve blood or barbarous behaviors (Bartow, 2006: 62).Yet, unfair or unethical -and, to Meyer (2014), cruel-situations can arise.Namely, not all individuals can enjoy what fundamental human rights aim to guarantee, such as the equal and fair prospect to access opportunities available in a society (European Union Agency for Fundamental Rights, 2018: 42).The latter, often referred to as ‗digital discrimination' (Edelman & Luca, 2014), is not tied to a specific territory and, thus, calls for universal measures.Same seems to be the case with unfair situations that may occur in other non-territorial areas of cyberspace.
There is also the (intrinsic) need to evolve and adapt to rapidly changing technological environments.This has been repeatedly highlighted by policy makers and organizations (A29WP, 2017;European Consumer Consultative Group, 2018;FTC, 2016;Informatics Europe & EUACM, 2018).Automated decision making is not just a scientific challenge.Leading to debates over what it means to be human, it can be understood as a political, economic, societal, cultural and philosophical challenge.Therefore, it calls for consensus, plans for humans to thrive when machines decide (Larus & Hankin, 2018).
If people do not agree on good life, they could agree that making the world a better place or at least removing pain are goods to be defended (First, 2018: 554).There could be a plenitude of shared views and those within differing areas of ethical taste could be discussed (Kruckeberg, 1993: 30).Thus, there could be a common understanding based on freedom, the necessity for ethical action and moral agency (Cooley, 2013: 367;Merrill, 2011: 8-9); a consensus to create a more symmetrical atmosphere (Hunt & Tirpok, 1993: 5-6).This could be an overlapping consensus (Rawls, 1987: 2-5) on conventional morality (Perry, 1977: 388); a free and willing agreement whose object and grounds would be in an independent ethical philosophy beyond controversial parts of theories (Baier, 1948).Instead of relying on positions at a given time or fragile circumstances that may render agreements instable, it could focus on principles and significant aspects of society (Rawls, 1987: 15;Perry, 1977: 388).Having its roots in ethics, it could be a consensus on responsibility to discern rightness and wrongness; to specify and prioritize fundamental principles; and to support measures that would guarantee effective enjoyment of rights and liberties associated with these principles (Rawls, 1987: 18).

Tracking the Trackers and a 'Right to an Unbiased Decision Maker'
Automated decision making, promising or threatening, may raise ethical questions, challenge legal certainty or be a cause for societal concern.Few might give their informed consent to data processing for medical research, albeit, millions would scream if they were denied access to benefits deriving from such processing.Tremendous challenges emerge from the acute need to protect human rights without impeding societal, scientific and other advances of utmost importance.
Thus far, it has been argued that human decision making may not be accurate; people are biased.But so are automated agents.This creates a need for explanations, to which machines do not easily lend themselves.Like human mind, they may be opaque and unintelligible.Moreover, in the absence of skills and time, it may take years for people to understand tasks run by machines.But this can also apply to tasks performed by humans.An automated agent and a human being can be biased and opaque, albeit, the former is seen as thoughtless while the latter enjoys monopoly over ethics.
However, it seems that we, the ‗thoughtful', challenge our own values by focusing on risk assessments and prioritizing (our will to know the future and thus) predictive performance over (the need to explain past or present and hence) interpretability.Then, it has to be humans the ones responsible for striking a fair balance; for reaching an agreement on well specified principles and mechanisms to enforce them.
That lack of transparency may be attributed to both human mind and AI is hardly surprising.The very goal of AI has been -from an engineering perspective-to make machines do things that would demand intelligence, were they done by humans, or -from a cognitive science perspective-to design systems that work the way human mind does (Chopra & White, 2011: 5;Citron & Pasquale, 2014: 6).Therefore, techniques, analogous to what humans do when asked to explain their decisions, could address some important aspects of opacity in automated decision making.To the extent that humans are considered interpretable, it may be this kind of interpretability that could apply (Preece, 2018: 67).Such models have already been proposed.When transparency is unachievable, post-hoc interpretations can explain predictions; instead of revealing how a model works, they perform in a way similar to, for instance, verbal explanations provided by humans (Lipton, 2018).Opaque models may then be interpreted after the fact and without sacrificing predictive performance (Lipton, 2018).However, such techniques are not always perfect; interpretations are given after the fact -or the harm.But, in so far as they may accurately interpret procedures, they could be a way to better understand opaque systems.Therefore, instead of targeting investments solely in designing opaque models, people could target labor, efforts, time and money in developing systems to interpret these models.People could invest in tracking the trackers.
Experimenting in such techniques could result in a more symmetrical atmosphere.It could strike a balance between programming to predict and instructing to interpret; between potential utility and desired fairness.Importantly, such measures could reflect general and commonly accepted principles, such as proportionality; the most far-reaching ground for review (Chalmers, Davies, & Monti, 2010: 368) and one of the oldest general principles (Court of Justice, 1956).Namely, if a predicting system restricted a human right, it would be fair to demand that such restriction should be proportionate in relation to the public interest pursued; and that any limitation, even a proportionate, should not undermine the very substance of a human right (Schütze, 2015: 98).Such automated procedures could be allowed, if they were appropriate and necessary to achieve important objectives legitimately pursued (Cheyne & Alder, 2007: 181-182).Proportionality could put automated decision making to the test; it could analyze its suitability and necessity and determine whether a procedure was suitable but also the least restrictive means to reach a given objective (Schütze, 2015: 204-205).Therefore, it could strike a fair balance between conflicting rights or values (Wright, Hurles, & Firth, 2016).Aiming to track the trackers, to examine their suitability and necessity, proportionality could be a way to address crucial aspects of opacity.
In addition to lack of transparency, bias seems to be another common feature shared by humans and agents who aim to work the way people's mind does.In both cases, de-biasing may be impossible.In human-situations, a person bound by the veil of secrecy that protects her deliberations (Herná ndez, 2014: 128), yet required to be unbiased or impartial, is a judge (Neimanis & Matjusina, 2011;Cooper, 2006).However, judges are not empty vessels that litigants fill with content (Meron, 2005: 365); they carry prior experience and, hence, bias.As some have aptly put it, in the best of all possible worlds, even the slightest possibility of bias would suffice to disqualify a judge from hearing a case; but, under such scheme, there would be no one left to adjudicate anything; reality, then, forces people to tolerate some bias (Redish & Marshall, 1986: 492).Thus, authors have suggested that the concept of judicial impartiality could be seen as some kind of awareness, an understanding by a judge that she carries bias that she has to rationalize when making decisions (Herná ndez, 2014: 154-155).
In the same vein, scholars have argued that AI could be instructed to acquire some kind of awareness of what it does not know; something close to consciousness (Raskin & Taylor Rayz, 2016: 393).Regardless of whether a quasi consciousness could be attributed to AI, systems could be programmed to diagnose factors that diminish accuracy.Namely, instead of demanding people to detect bias in automated decision making, it might be easier or more reasonable to program a system to identify, for example, missing data (Taylor, 2015: 515).
Furthermore, measures safeguarding rights associated with fundamental principles, such as fair or due process, could to some extent address partiality.Experts could focus on bias minimization as a way toward a right to an unbiased and accurate decision making.In human-scenarios, this right is a key fair process value, the floor of due process (Redish & Marshall, 1986: 479;Poole, 1998: 236).It can allow for procedural rationality that can enable people to engage in rational planning about their situations and make informed choices among options; to have a better chance of ‗knowing what is going on', what is happening to them and why (Summers, 1974: 26-27).
In this context, due and fair process could demand that automated procedures satisfy some standards (Crawford & Schultz, 2014: 93-128).Mechanisms could require systems to confirm their fairness and accuracy in various steps of decision making.This could be a way to introduce a ‗right to an unbiased decision maker', to guarantee the basis of due process.Interestingly, aspects of what has been referred to as ‗technological due process' (Citron, 2008(Citron, : 1301) ) are provided in existing regulations (GDPR, 2016).For instance, certification mechanisms -already established, yet, on a voluntary basis-(GDPR, 2016: Article 42(3)) could further determine design specifications, expertise involved or performance expected.They could also specify information about the experts who designed a given system (such as field of expertise or years of experience) or the nature of data collected (Edwards & Veale, 2017: 79).Importantly, disclosure of these data would not violate trade secrecy; nor would it run counter to privacy expectations.Thus, such information could be publicly available (GDPR, 2016: Article 42(8)) to allow for review and oversight.
It seems that measures aimed at striking a fair balance between conflicting values or rights but also mechanisms focused on strengthening fair processes could rely on principles already proposed, accepted and analyzed by regulators, policy makers and the academia.Such measures and mechanisms may reveal that the problem can often lie, not in the absence of procedures, rules or principles but, in their (application or) understanding.Namely, proportionality, applied to trade secrecy on a commonly understood basis, could resolve a number of contemporary uncertainties (Directive 2016/943: Articles 1(2), 7(1)(a), 11(2); Recitals 21,26,28,36).Undoubtedly, trade secrecy can have its place (Dixon & Gellman, 2014: 7).But who would argue for secrecy that denies due process, violates privacy or prevents justice?When secrecy mistreats or threatens, proportionality and its strict ‗tests' could protect public interest and fundamental rights and freedoms.

Discussion
Humans have long been prioritizing themselves over ‗others'.Conservation projects are a good example (Doak, Bakker, Goldstein, & Hale, 2014).Setting population goals for non-human species in the low thousands, while condoning for humans a population in the billions, raises the question ‗how many humans can, in fairness to other species, live on earth?' (Mathews, 2016: 142).
From this perspective, prioritizing AI over humans (and their rights) seems ‗interesting'.Yet, if this prioritization is to satisfy mere economic interests, it may be problematic; if this emphasis is because AI may resemble human mind and, thus, ‗must deserve' some particular attention, an anthropo-like-centric view, then it may be hardly surprising -it would most probably be following the trend to prioritize humans or those who (aim to) ‗think alike'; and, if this is because AI may serve human needs but also desires -the will to know what will be-, it could be a cause for concern.
Regardless of whether AI is prioritized for profit or its ‗human-like mind', for people's wills or their needs, such emphasis and focus can often be unfair, unjust or unethical.People have to think both seriously and creatively.Fair and proportionate approaches may not always or in any scenario enable people to understand, scrutinize or control everything; nor would they per se guarantee absolute accuracy.But they would require that processes be structured to foster appropriate control and reasonably accurate determinations (Perry, 1977: 392); they could work as a tool for achieving long lasting improvements (Behrens, 2005: 6) and, thus, address important aspects of automated decision making.Based on a free and willing consensus, built upon common understanding, such strategies could allow for sharing good practice, for incentivizing changes in behaviors (Behrens, 2005: 6).Instead of supporting solely the regime of control-by-risk, where numbers are aimed at reflecting people's internal thoughts and ideas to forecast scenarios and judge them, there could be a shift toward, or some particular focus on, a concept where one's actions would be regarded as key determinants of her identity and character; a notion where choosing what to do would be seen as choosing who to be(come) (Lebacqz, 1985: 83;Johannesen, 1988: 62).Some have argued that if people were able to control everything they would not even wake up in the morning (Kim & Routledge, 2018).This may be true and in compliance with working with the idea that the future is unknown.People may never achieve ultimate control over everything, always and everywhere -such control may not be desirable.
But an appropriate and adequate level of control is desired and could be reached.People, working on a commonly understood basis, could take their responsibility to discern rightness that many theories of ethics aim to study.Under a fair and proportionate scheme, one who controls her behavior in a suitable sense (Fischer & Ravizza, 2000: 12-14) could then be rightly subject to reactive attitudes (Oshana, 2002: 263;Strawson, 1993); she could, thus, be rightly held responsible.Such scheme might be a way to bridge responsibility gaps (Mittelstadt, Allo, Taddeo, Wachter, & Floridi, 2016: 11) emerging from complete absence of control over algorithms' behavior.
In the past, courts rightly suggested that holding a firm responsible for selling cellular phones to ‗teenage punks' who use the phones to set up drug deals would be unreasonable (U.S. Court of Appeals, 1992).Today, at risk of ‗falling' into responsibility gaps, claims suggest that those using automated systems be held responsible for decisions made, even if they cannot explain (and, perhaps, control) the relevant process (Association for Computing Machinery, 2017).Instead of constantly looking for someone on whom to pin the blame, attention could be drawn to strategies that would ensure appropriate control over procedures and behaviors; policies that would guarantee there would be the prerequisite for rightly holding someone responsible; schemes that would promote a heightened sense of responsibility or accountability (Perry, 1977: 428).
This article has argued that there may be a number of fundamental principles, on which people could agree, to address algorithmic uncertainties.Strategies to effectively exercise rights related to these principles could range from a due process approach and a ‗right to a fair, accurate and unbiased decision maker' to proportionality tests and investing in ‗tracking the trackers', in favoring the rights of the judged over the judger (Bambauer & Zarsky, 2018: 35).People could build upon existing principles and mechanisms.They could reach a free and willing consensus that would be the criterion of proportionality (Perry, 1977: 407); an agreement that would respect diversity.
As noted above, each culture, community or family may have its own ‗rules, god or time'.But they do have some kind of, or something similar to, ‗rules, god or time'; those without ‗rules' have something similar to what others call ‗rules'; those without ‗god' have something to believe in, something close to what is referred to as ‗god'; and those without time, if any, know that something happened and expect something to be.If people cannot agree on ‗time', they could agree on ‗day', ‗week' or ‗past, present and future'.
The perfect ethical theory that would address any scenario may not come soon.Instead of waiting for it, there could be some consensus, some experimenting in fair processes, before we come to realize that muddling through would have been better than not acting at all. /papers.ssrn.com/sol3/papers.cfm?abstract_id=1926431