Abstract
Algorithms are the key instrument for the economy-on-demand using platforms for its clients, workers and self-employed. An effective legal enforcement must not be limited to the control of the outcome of the algorithm but should also focus on the algorithm itself. This article assesses the present capacities of computer science to control and certify rule-based and data-centric (machine learning) algorithms. It discusses the legal instruments for the control of algorithms and their enforcement and institutional pre-conditions. It favours a digital agency that concentrates expertise and bureaucracy for the certification and official calibration of algorithms and promotes an international approach to the regulation of legal standards.
Employment in the economy-on-demand and the influence of algorithms
Platform-based business models change the working world and the ongoing digitalisation presents opportunities and risks, which are currently being addressed by economists, social scientists and legal scholars. 1 It creates new job opportunities and removes market barriers, allowing businesses to recruit their workers from a larger pool of candidates. For certain categories of persons, the demand economy enables them to enter the market, in particular people providing nursing care for family members, people with disabilities, as well as people living in poorly served regions. In addition, digital labour platforms provide another tool to facilitate the reconciliation of family and working life. At the same time, critics view the platform economy as a new type of Taylorism – low-paid wage labour, barely above the minimum subsistence level, undermining the legal and social standards of good work. 2
Jurists move between these two narratives. For labour law experts, the starting point of the discussion was the employment status of these workers. The scholarly debate, in all, follows accustomed patterns: an analysis of the working conditions leads to an assessment of the necessary standards of social protection and to suggestions of remedial measures. Further, trade union initiatives (e.g. fair crowdwork) – with the support of the European Parliament 3 – are demanding guidelines for digital platforms. They have established their own rating platform, where platform providers and platform clientele can be evaluated.
On the other hand, so far, the concrete implications of algorithmic platform management and platform management by artificial intelligence are rarely discussed. Social scientists and economists have accurately described the phenomenon of algorithmic management or algocracy: 4 Workers are permanently being monitored, tracked and rated. The continuous and automated recording of usage habits, work results and customer reviews, often combined with rewards for favourable user behaviour are characteristics of the platform economy. A worker’s data track leads to a built-up of virtual reputation, which in return results in more (lucrative) assignments. Instead, if previous results did not meet the clients’ or platforms’ expectations, workers may algorithmically get rejected from future jobs. If it is known what kind of behaviour leads to more job offers and/or higher wages (e.g. work within certain time-frames, working speed or quality), algorithms trigger behavioural effects. At the same time, the comprehensive access of platforms to these user data aims at a customer-oriented selection of workers as well as an economical optimisation up to the gradual exclusion of less profitable platform users. 5 This phenomenon is not limited to platforms. Other actors, in all sectors of the economy, also implement this mechanism if they make use of algorithms in their management (management-by-algorithms). 6 Therefore, the following explanations on legal risks and regulation may also be considered in other cases.
In his influential article ‘Code is Law’ in Harvard Magazine, Lawrence Lessig considers ‘code’, i.e. the set of algorithms and their implementations, as the main regulator of cyberspace. 7 Here we address the question, whether in the light of an increasing number of autonomous decisions made by computer programmes, the ‘code’ itself requires additional and novel forms of regulation. The effects of algorithms on the working world are not limited to the intentional supervision and selection of workers (by applying a scoring function). Their design (due to its construction or the data used for training) has side effects platform operators may neither want to achieve, nor be aware of. This is particularly true, where contingencies, such as incomplete data sets or distorted information, lead to falsified rankings because the algorithm does not know how to adequately deal with misinformation. This may result in a worker’s ranking substantially deteriorating or improving for no apparent reason. The same applies, where job opportunities are only presented to a pre-defined number of platform users, selected by an algorithm from a group of workers with identical rankings.
These malfunctions of algorithms must not only interest platform managers, looking to recruit the best-qualified workers, but also jurists. Such effects result in the exclusion of workers from certain job offers or ban them from the platform entirely. This complicates or hinders market access, especially where platforms develop into dominant undertakings. Analyses by economists and legal scholars specialising in competition law suggest that these effects are occurring. In short: from an economic standpoint, a platform is particularly interesting for its users, if it connects a wide audience of people. This causes suction effects, which bear the risk of unlawful market concentration. It is already observable, that larger platforms are starting to control the market.
Legal relevance of algorithms
The fairness of algorithms is an ethical, an economic, but also a legal, category. From a legal perspective, their fairness must depend on the legal framework, which defines certain actions or results as legally relevant. Despite the different national legal backgrounds, certain issues regarding the fairness of algorithms are relevant in most jurisdictions. They originate from anti-discrimination law 8 and competition law. Both, notwithstanding their diverging objectives (protection of personality rights on the one hand, strengthening of competition on the other), have a common denominator: discriminatory as well as anti-competitive conduct complicate or limit market access and impede appropriate pricing.
The specific effects of algorithms this article focuses on, are the legally relevant effects of the design of algorithms. Firstly, this is the general rule of the algorithm (scoring function), which defines the deciding criteria for a worker’s ranking (e.g. cheaper and better). Such a general rule can result in direct as well as indirect discrimination, but in this context indirect discriminatory conduct is more interesting, as the resulting manipulations are subtle and clandestine. For example, if it proves advantageous for contractors to work on Sundays, this may be discriminatory based on religion. The same is true, where algorithms favour workers, who are available at times at which people with family commitments are less likely to be able to render services. Typically, this leads to the disadvantageous treatment of women. 9 Cases of indirect discrimination also occur where pace of work is important, as this tends to exclude people with bodily impairments. Not in all cases, unequal treatment is unjustifiable, and indirect discriminations in particular can be objectively justified. To what extent the full inclusion of people with disabilities or the reconciliation of family and working life require the restriction of such effects are valuation issues we need to debate on national and international level but should not leave them to the company and the coder.
Besides the general rule, the specific construction of an algorithm may cause legally relevant effects. Its functioning can discriminate against groups of people or limit market access. It can disadvantage people based on personal characteristics, if regulation defects result in a systematically less favourable treatment of women, persons with disabilities or people of different ethnic origins. Additionally, algorithms can be the cause of discriminatory conduct or structural abuse in the sense of the competition law concept, if they favour certain contractors without objective grounds or deny groups of workers access to the platform 10 or prevent the change between platforms by incompatibility. 11 The vertical integration of users by the platform and the lacking access to information can also distort competition (e.g. platform users receive detailed information on clients and its preferences that are otherwise not available on the market). Data possibly allow platform price discrimination with regard to the behaviour and needs. This may affect the platform’s clients and crowdworkers as well, but not every discrimination is relevant as it depends on the negative effects on the competition and the consumer surplus.
The following contribution examines, how far it is possible, to identify discriminatory effects based on the algorithm rather than its output. This would enable the optimisation of algorithms, but also allow for their supervision and lead to the question of whether algorithms must be subject to an ex ante control.
Fairness of algorithms – Bias of algorithms
Starting point – Technical part
In the broad field of e-commerce, algorithms on a very general level perform various forms of ranking tasks. Database entries (e.g. products or service providers) are ranked according to an objective function (cost function or scoring function). In our examples and in the discussion of algorithmic properties, we will restrict ourselves to algorithm-centric data processing (as opposed to the data-centric data processing, which is characteristic of machine learning approaches). This restriction allows us to be clearer and more specific. In the outlook we will address some additions to the debate arising from machine learning.
In the case of algorithm-centric data processing, deviations from fairness (and, hence, discrimination) enter this process on two levels:
The definition of the scoring function (usually either requiring the challenge of mapping a high-dimensional feature vector onto a real number OR defining a distance measure (or metric) for two such high-dimensional feature vectors).
The tie-breaker criteria that decide on the order of the output, when equal values of the scoring function are encountered.
In the following we will (A) introduce the relevant terminology using a few simple examples; (B) summarise the existing literature on algorithmic fairness, particularly in computer science; and (C) embed the ideas derived from the literature in a broader sense with a special emphasis on their legal implications.
Terminology
In addition to the ranking tasks mentioned above, algorithms also perform decision tasks and clustering tasks. As discussed above, ranking tasks use a scoring function to produce an ordered list of data entries. A decision task produces a decision about an individual data entry (e.g. a ‘yes’ or ‘no’ result or the assignment of a data entry to a category from a list of pre-defined categories). A clustering task summarises all algorithmic attempts to find groups in data. Clustering can be used, for example, to derive the categories, which are required as input to a decision task, from existing data. Currently, most implementations for ranking tasks (as encountered in the context of platform economy) are still rule-based. The application area of clustering tasks is dominated by data-driven (machine learning) algorithms, while decision tasks are typically addressed by both algorithmic paradigms. Figure 1 summarises this overview of algorithmic paradigms.
Summary of algorithmic paradigms, together with typical algorithmic tasks and the issues, which may have a systematic impact on fairness.
In a (binary) decision task, the algorithm receives an input and computes a yes/no answer. In a clustering task, the algorithm receives a large set of inputs and groups them into categories, called clusters.
It should be noted that the distinction between ranking, classification and decision tasks, though helpful, is not always clear. Mapping inputs into categories can in most cases be viewed as a classification task. On the other hand, when the same task is rephrased as finding groups in the data, which – if the initial categories are meaningful – should correspond to these categories, one can also consider it a clustering task. When formulated slightly differently, one can also regard this as a decision task: for each input one decides, whether this input is mapped to a given category or not (with the subsidiary condition that no input can be mapped to more than one category). We can also formulate this task as a ranking task, by using a scoring scheme (e.g. a distance metric from other members of the category), in order to assess, to which of the given categories the input at hand is closest.
All these algorithmic tasks revolve around the processing of available data. Even though the technical and conceptual details behind each algorithmic task may vary, it is instructive to look at the general features of such algorithms using a simple example. Imagine an online marketplace connecting customers to service providers. Based on the information she provided in her service request query a user, Alice, will receive a ranked list of potential service providers. The rank is determined by an algorithm converting Alice’s information vector and the information vector of a potential service provider into a score quantifying the agreement of the two information vectors and then sorting all potential service providers according to this score.
Here a vector means an ordered list of inputs (e.g. parameter values of the job being offered and to be evaluated by the scoring function). In the example outlined above, if we assume a home decoration job offer, Alice might characterise the job in the online platform using parameters like size of the room, time frame for the work to be completed, current surface conditions of the walls, etc. or any other type of categorical or numerical information required to characterise the offered job, often using categories and terminology provided by the online marketplace. The companies then have, for example, tolerance windows for some of these parameters, indicating, whether or not they are likely to accept such an offer. These parameters are then put together in the scoring function, which often contains features proprietary to the marketplace, and then in turn leads to the ranking of companies displayed to Alice.
Let us assume that two potential service providers for Alice’s query, company B and company C, have the same score. This could be because the information vectors of B and C are nearly identical or because the scoring function employed by the algorithm is insensitive to the differences between the two information vectors (e.g. by having very small weights assigned to the vector components, in which the two information vectors differ) or, lastly, because the positive and negative contributions to the score happen to yield the same score for both, B and C, with respect to Alice’s service request query. The latter is more frequent in the case of short vectors and mostly categorical information (e.g. simple ‘yes’ and ‘no’ entries) contained in the information vector. Such cases require a tie-breaker criterion, in order to determine, whether B or C are ranked higher in the list received by Alice in response to her database query. Typical outcomes could be that B is ranked above C, because alphabetic sorting is used in the case of an equal score. It could also be that the company, which has been in the database longer, is returned above the other (first in, first out, FIFO), or, conversely, the company, which has last joined the database, is returned above the other (last in, first out, LIFO).
Repeated application of such tie-breaker criteria over a large number of queries leaves a pattern of rank deviations. For example, the average rank of a company, whose name starts with Z, could be systematically higher than the average rank of a company with an A.
Furthermore, the scoring function itself has a strong impact on the pattern of ranks produced by the database operation and the algorithmic generation of service provider lists in response to a query. As already outlined above, it is easy to envision situations, where flaws of the scoring function are exploited by users by modifying their information vectors in alignment with the specifications of the scoring functions, in order to obtain systematically better rankings.
It is a challenge in its own right to define the fairness of algorithms. In the case of decision algorithms typical definitions of algorithmic fairness involve the expectation that the assigned categories on average are equally probable with respect to outside attributes, which do not directly enter the algorithm. This can be quantitatively assessed by computing the difference of the conditional probability of an outcome given the one characteristic, for which discrimination is expected, (e.g. ‘female’ as gender information or ‘foreign’ as ethnicity information) and the other characteristic (e.g. ‘male’ or ‘native’). 12
In the case of ranking algorithms, a fairness requirement could include the robustness of the rank against small variations of the scoring system 13 or a similar long-term distribution of ranks for all database entries, when exposed to a broad range of queries.
Literature review
There is a rich debate in computer science about the fairness of algorithms and algorithmic discrimination. Even though this debate is still far from a convergence towards universally accepted standards and methods, the recent literature can be classified into attempts to measure discrimination, 14 algorithm enhancements preventing discrimination and increasing fairness, 15 and the auditing or control of algorithms with respect to discrimination. 16 Contributions to this debate exist for both algorithm-centric and data-centric (machine learning) approaches. 17
Summarising, this debate has shaped our view on algorithmic fairness by providing illustrative examples, by contributing novel mathematical concepts and novel classification schemes of fairness enhancing methods.
Regarding the question, which elements of this debate can have direct implications for the legal framework of algorithmic decisions, a few investigations are of particular interest. Žliobaitė and Custers discuss an interesting toy model example, where they generate, and then analyse, data with an embedded ethnicity bias. 18 The fact that the data model trained with these fictitious data indirectly perpetuated the bias, even though ethnicity information was not included in the data, illustrates in a clear and transparent fashion, how discrimination enters data-centric algorithms (i.e. algorithms based on learning patterns in training data). The main result of this study is that formulating a data model (i.e. learning decision rules via data-centric data processing) requires the sensitive, discriminatory information to be included in the training data: ‘[c]ollecting sensitive personal data is necessary in order to guarantee fairness of algorithms, and law making needs to find sensible ways to allow using such data in the modelling process.’ 19
In the legal system in the US, one application area of algorithms – risk assessment in criminal sentencing – has received particular attention over the last years. 20 This development had already been anticipated almost a decade ago by Susskind. 21 Corbett-Davies et al. use the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) algorithm as the main example in a broad discussion of optimal decision rules under various definitions of algorithmic fairness. 22 The algorithm has been under scrutiny recently, because (1) a study could demonstrate that even though race is not an explicit input to the algorithms, the output shows a clear racial bias; 23 and (2) it was shown that a ‘crowd computing’ approach, where the decision is made by laypersons, achieves a similar performance as the COMPAS algorithm. 24
A trend in the literature, for example, is to find ways of excluding biasing information (e.g. race) from the feature set (e.g. the questionnaire filled out by applicants for a loan). A remaining challenging problem is then to ensure that the excluded information cannot be statistically reconstructed by other features. In contrast to this general approach, in Kleinberg et al. (2018) it has been argued to explicitly include such information in the feature set and then distinguish between an ‘efficient planner’ algorithm (that just maximises a score and hence can lead to an effective bias) and an ‘equitable planner’ algorithm (that avoids discriminatory biases by requiring equal distribution across the relevant features). 25
The debate in computer science: Discussion and outlook
The topic of algorithmic fairness needs to be distinguished from another virulent topic, algorithmic regulation, which exists in two variants: regulation by algorithms 26 and regulation of algorithms. 27
Currently we see a shift from algorithm-centric to data-centric data processing, i.e. decision procedures are ‘learned’ via training on large data sets and then, after test and calibration phases, employed to autonomously perform classification, decision and ranking tasks. The computational structures, in these ‘machine learning’ approaches are, for example, neural networks. 28
Any type of bias on this level can be linked, in principle, to two basic types:
biases in the training data; and
limitations (or even artefacts) generated by the predefined computational structure (e.g. the number of layers in a multi-layer neural network).
One challenge is the lack of interpretability of the internal representation of the data within the algorithm (e.g. within the pattern of weights in the trained neural network). This ‘opacity that arises from the characteristics of machine learning algorithms’ 29 limits the assessment of algorithmic fairness.
Certification of algorithms – from a control ex post to an examination ex ante?
Legal instruments
Legal violations resulting from the specific characteristics of algorithms call for the further development of legislation and law enforcement. From the perspective of the right owner, key barriers to law enforcement are the lack of transparency and comprehensibility of algorithms as well as the pace of technological change. The development of new algorithms and machine learning systems is likely to proceed faster than the development of high-quality test methods. As a result, the detection of legal infringements becomes significantly less likely, which – according to economic analysis – will provoke an increase in rights violations. Therefore, the legal system needs to counteract the undermining of the laws by technology.
Approaches for the improvement of legal protection must be technically possible as well as practically enforceable. Suitable legal reactions to the impact of algorithms on the economy-on-demand do have to reflect the available instruments and its enforcement but also the institution necessary in order to keep up with the technical progress and to protect the interests of enterprises and workers as well as the business secrets of the platform owner. In addition, one must consider that platforms operate internationally. Thus, solutions on national level are likely to fall short.
In cases of lack of transparency, the respect of rights and the improvement of legal protection are usually ensured by creating transparency and improving comprehensibility, in particular data protection and consumer protection law, but also employment law. However, regarding the issue at stake, an increase in transparency does not remedy the described deficiencies; adverse effects have their roots in the algorithms themselves. The disclosure of said algorithms is thinkable, but rather useless, as only computer experts can read and decipher them. This area is the preserve of experts. But even they encounter their limits when dealing with machine learning algorithms, and protected business secrets may be concerned. Hence, from the perspective of those seeking justice, solutions through data protection law are worthless. 30 Regarding platforms, the right not to be subject to a decision based solely on automated processing contained in Art. 22 General Data Protection Regulation (GDPR) is also not an adequate remedy. 31 The platform management has to respond to the worker’s request individually. Doing so, requires a solution that is more expensive and by no means more transparent.
As rule-based and data-centric algorithms are technical tools the certification or calibration by an independent institution is an alternative path to prevent intentional and unintentional unfairness of algorithms. Despite the need for more research in computer science, the legal response can and shall keep up with the technical development. Summarising the detailed technical discussion above, a suitable requirement for rule-based algorithms could be the robustness of the rank against small variations of the scoring system. In the case of data-centric algorithms one may impose the use of multiple methods and classifiers, as well as monitoring the results under variation of the training data. For both algorithmic paradigms, the long-term distribution of ranks for all database entries under a broad range of queries could be monitored and statistically assessed with respect to intrinsic biases.
Certification or calibration as an instrument for legal control of algorithms has two main preconditions. First, there is a need for a stable and reliable testing method. Second, the legal standards have to be feasibly transferable in a testing tool. The development of calibration and certification methods requires more research at the interface of computer sciences, social sciences and data analytics, embedded in an ongoing dialog with legal experts. In general, rule-based algorithms may be easier to audit and control with respect to algorithmic fairness than data-centric algorithms.
Furthermore, certification needs the transformation of legal standards in decision trees that allow the detection of discrimination and its proper justification. Whilst the detection of an intentional or unintentional unequal treatment according to one of the inadmissible criteria is very formal and therefore easy to identify if the data are available, the proper justification of discrimination is much more difficult. The main obstacle for certification as a legal instrument is the justification of discrimination that is regulated by general clauses in anti-discrimination law. When designing a certification tool one can either limit the test to unequal treatment as such, and check the detected discrimination by the administration, or integrate justifications in the testing tool as far as possible and thereby certify that the algorithm is up to some extent free of discrimination.
Despite these pre-conditions, certification or calibration is superior to the existing dummy job applications (test persons). 32 The assessment of the algorithm itself is more effective, it provides a more intensive and precise scrutiny. Evaluations by test persons can lead to mistakes. Calibration by a specific homogenous data set can have a homogenous output. Hence, big data sets allow proving biases by statistical methods.
Enforcement of certification or calibration
Formation of an agency – Independent accumulation of expertise, administrative solution
These are strong indicators that the control of algorithms should be conducted by a body or an agency, as such a specialised institution could accumulate the necessary expertise and monitor new technological developments. 33 It could participate in the development of testing standards and their evaluation and certify algorithms, whilst at the same time ensuring the adequate protection of business secrets. Legislation is no alternative means. The rapid development and the technical complexity make it impossible to have detailed legislation on the certification of algorithms. Under constitutional law it is sufficient if legislation provides the legal framework for such an agency (e.g. aim, procedure, and authorisation).
First developments in this direction can be observed in the USA, the UK and the Netherlands: In these countries, competition authorities have special departments for competition and consumer law protection in the digital economy. The House of Lords’ Select Committee on Artificial Intelligence makes additional proposals on expert guidance. 34 In late 2016, the advisory council for the German Federal Ministry for Justice and Consumer Protection proposed the establishment of a digital agency. 35 In 2017, the Federal Ministry for Economic Affairs and Industry endorsed this proposal. 36
Surprisingly, these advances have not yet been taken up in the discussion on platform-based business models and crowd work. Still, participating in this discourse is important because (most) crowd workers cannot be classified as employees, but are independent contractors. Moreover, unfair competition laws and anti-trust laws also apply to employers, at least beyond the exception for collective agreements. Therefore, addressing the specific problems of algorithms through a digital agency simultaneously offers a consistent solution for crowd work.
Such an agency is an administrative body and should be under the control of the highest administration, e.g. the Federal Cartel Office (Bundeskartellamt) is controlled by the Federal Ministry of Economic Affairs and Energy. To what extent the social partners should participate in such an agency or its administrative procedure has to be decided based on various factors, but it has to be borne in mind that only a part of the affected crowd workers are employees and organised by social partners. It is of utmost importance that when an agency exercises administrative power, it observes and respects the rule of law. Social partners can support such an agency by providing information and exercising and/or enhancing private enforcement. Anti-discrimination law already promotes that organisations with legitimate interests may engage within the legal framework of the Member States supporting complaints (see Art. 10, para. 2 Direction 2000/78/EC). Anti-trust law has its own elements of private enforcement (e.g. disgorgement of benefits by association see para. 34 Gesetz gegen Wettbewerbsbeschränkungen, Competition Act).
Administrative enforcement or Legal incentives
When creating such an agency and defining its responsibilities, one should consider whether algorithms should be controlled ex post or need to be certified ex ante. The latter would lead to a fundamental change of paradigm: As a rule, competition law, anti-discrimination law and data protection law rely on an ex-post control. A control ex ante highly depends on the threat for a person’s rights by the algorithm. From a systematic point of view one can introduce a repressive ban with an exemption option or a preventive ban with an authorisation option but if the technical development is reasonable and desirable in general it shall be allowed but certified by the administration.
All these options might come into use. As lethal autonomous weapons systems (‘killer robots’) will be repressively banned, a preventive ban is typical for areas of the law guaranteeing technical safety (product and equipment safety), where failures can lead to irreversible damage to lives, bodies and health. Only those algorithms that have potential effect on the health and safety or life of a person will preventively be banned. There might be fields of application that create comparable risks (e.g. autonomous cars, medical treatments) but there is no need for a general preventive ban with an authorisation option. In the economy-on-demand, the inherent dangers of algorithms are of a different kind. Algorithms (machine learning and rule-based) are a commonly used technique. This technical revolution has its ambiguity, but as long as we regard it as promising and desirable for society, economy and public welfare, the use of algorithms should be permitted in legal categories.
However, the lack of transparency measurably complicates ex post law enforcement. Effective remedy through associations seems unlikely, as they are facing the same technical obstacles. This is an argument in favour of a certification ex ante. Therefore, certification could be used to make legal protection ex post more effective. Legal responses are not limited to the alternatives of prohibited and allowed behaviours or products. An ex ante control may also be exercised for behaviours and products that are allowed in general but should be monitored by the administration due to their potential dangers for rights and freedoms.
Still, to some extent, a race between technical advancement and the development of testing standards is inevitable, leading to a protection gap between legislation and law enforcement. The administration has to keep up with the technical development. This can be facilitated through incentives for the companies, first, to ask for certification, and second, to keep the agency informed about the technical development. Beyond labour law, Scherrer proposed a national regulation of strict liability for the users of specified algorithms, 37 which can be reduced to fault liability, if the algorithm is subject to an ex-ante control by the national agency. This proposed regulation has two advantages: According to international private law, the law applicable to obligations arising out of a tort is the law of the country in which the damage occurred. 38 Hence, national legislation can offer protection, even if the algorithm was constructed, tested and implemented elsewhere. Furthermore, the company that implemented the algorithm has a strong incentive to contact the agency and ask for the certification of the algorithms.
This concept can be transferred to labour law issues. Even if all EU Member States and many other states world-wide allow a choice of law for the employment contract, there are overriding mandatory provisions in the national laws, which govern the employment relationship without the choice of law. 39 Anti-discrimination law in particular contains such overriding mandatory provisions. 40 The strict liability for discrimination could be a starting point for a comparable regulation. The range of damages might be too low for a strong incentive, but the sum of liability can make the regulation finally effective.
National and/or international solutions
In implementing such a strategy, attention should be paid to the coordination of national, supranational and international laws. In general, states are competent to establish obligatory, national digital agencies. As a result, platforms, which operate internationally, would be subject to multiple checks. That would create different, perhaps even conflicting standards and costs. Simplifications could be achieved through international certification standards or cross-recognitions of certificates. Within the EU cross-recognitions follow the principle of the country of origin; in the wider context of international law they are governed by international agreements.
The question, which international organisation shall be responsible, is not easy to answer, as there is the ILO for employment law, the WIPO for intellectual property law and the WTO for international trade law. One suitable player could be the OECD, since it has already enacted guidelines for corporate governance and multinational enterprises and has a digital agenda. The development of a code of conduct in form of international soft law and the establishment of a certification body would be sensible additions to the current efforts to uphold legal and ethical standards in the digital economy.
Conclusion
Our article has shown that an algorithm-based platform economy carries the risk of undermining current legislation. Law enforcement through individual legal protection should be complemented by state support, where individual protection is limited by reason of technicity and collective representation is not able to remedy this. It is our task to cooperate with physicists and computer scientists to further develop our legal system. Particularly, self-learning systems will present new challenges, not only of a technical, but also of a legal nature.
Algorithmic decisions are by now ubiquitous. 41 We would like to emphasise that the overarching topic of the fairness of algorithms goes far beyond the specific focus selected here, the platform economy. The risks of input-output biases, and thus breaches of algorithmic fairness, which have been enumerated above (choice of the scoring function and tie-breaking criteria for rule-based algorithms; choice of methods and biased training data in data-centric algorithms can and should be delineated in a wide range of domains of application.
Businesses are already aware of the necessity of such developments. In the 2017 Founders’ Letter, published on 30 April 2018, the President of the Google parent company Alphabet, Sergey Brin, wrote about algorithms (in particular in the context of machine learning): ‘However, such powerful tools also bring with them new questions and responsibilities. How will they affect employment across different sectors? How can we understand what they are doing under the hood? What about measures of fairness?’ 42
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
