Best Practices for Transparency in Machine Generated Personalization

by   Laura Schelenz, et al.
Universität Tübingen

Machine generated personalization is increasingly used in online systems. Personalization is intended to provide users with relevant content, products, and solutions that address their respective needs and preferences. However, users are becoming increasingly vulnerable to online manipulation due to algorithmic advancements and lack of transparency. Such manipulation decreases users' levels of trust, autonomy, and satisfaction concerning the systems with which they interact. Increasing transparency is an important goal for personalization based systems. Unfortunately, system designers lack guidance in assessing and implementing transparency in their developed systems. In this work we combine insights from technology ethics and computer science to generate a list of transparency best practices for machine generated personalization. Based on these best practices, we develop a checklist to be used by designers wishing to evaluate and increase the transparency of their algorithmic systems. Adopting a designer perspective, we apply the checklist to prominent online services and discuss its advantages and shortcomings. We encourage researchers to adopt the checklist in various environments and to work towards a consensus-based tool for measuring transparency in the personalization community.



There are no comments yet.


page 6


Algorithmic Transparency with Strategic Users

Should firms that apply machine learning algorithms in their decision-ma...

Uncertainty as a Form of Transparency: Measuring, Communicating, and Using Uncertainty

Transparency of algorithmic systems entails exposing system properties t...

Black Feminist Musings on Algorithmic Oppression

This paper unapologetically reflects on the critical role that Black fem...

"I had a solid theory before but it's falling apart": Polarizing Effects of Algorithmic Transparency

The rise of machine learning has brought closer scrutiny to intelligent ...

Revocation Statuses on the Internet

The modern Internet is highly dependent on the trust communicated via X....

TILT: A GDPR-Aligned Transparency Information Language and Toolkit for Practical Privacy Engineering

In this paper, we present TILT, a transparency information language and ...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Recent years saw significant increase in personalization approaches for online systems [segal2018optimizing, van2019collecting, yanovsky2019one]. Such personalization can be used to direct users’ attention to relevant content [kolekar2019rule], increase their motivation when working online [muldoon2018survey], improve their performance [jackson2016encouraging], extend their engagement [segal2016intervention] and more. These approaches rely on social theories of human behaviour (e.g. [deci2002overview, borji2012state]

) as well as on machine learning based abilities to predict human behavior and human response to various interventions  

[plonsky2019predicting, hartford2016deep].

Yet, personalization technology that focuses on maximizing system designers goals runs the risk of marginalizing users. 444 Personalized recommendations of content, products, and services usually attempt to influence a person’s decision-making. When such influences are hidden and subtly try to persuade users (maybe even against their expressed goals), this constitutes a form of manipulation  [Susser.2018]. Subverting a person’s decision-making abilities reduces their autonomy. Especially with regard to personalized advertisement, personlization can exploit users’ vulnerabilities  [Spencer.2019] and may even threaten democratic processes  [Crain.2019].

Applying transparency to the design of personalized content can help address these challenges. It can reduce power asymmetries between users and organizations, increase users’ understanding of what the system is doing, explain the reasoning for its operations, and enable users to obtain control over the system’s behaviour  [Turilli.2009, Lepri.2018, Simon.2010].

The computer science community has affirmed the importance of transparency in its profession. The ACM Code of Ethics reads: “The entire computing profession benefits when the ethical decision-making process is accountable to and transparent to all stakeholders” [ACMCouncil.22June2018]

. Also political bodies identify transparency as a pivotal principle in Artificial Intelligence based software  

[Floridi.2018, HighLevelExpertGrouponArtificialIntelligence.2019]. Especially with the advent of legal frameworks that prescribe transparency in data collection, processing, and storage  [EuropeanParliamentandCounciloftheEuropeanUnion.27April2016], system designers require increased awareness and guidance in the implementation of transparency in their systems  [Amershi.2019]. Recent and emerging scholarship on explainable AI underlines the importance of transparency in computer systems  [Hois.2019, Rosenfeld.2019, Zhou.C2018, Preece.2018, adadi2018peeking, unknown].

This paper is organized as follows. Section 2 discusses the need for transparency in Artificial Intelligence systems. Section 3 provides a definition of transparency drawing on prior art, and derives transparency best practices for personalization in online systems. These best practices constitute ethical responsibilities on the part of system designers. Based on these best practices, we specify questions that should be addressed when considering transparency in personalization processes. This constitutes a concrete first checklist that can be used by systems designers to evaluate and operationalize transparency in their systems.

Section 4 then describes a preliminary application of the checklist to existing web sites in the wild. Specifically, we look at Facebook, Netflix, YouTube, Spotify and Amazon. For each such destination, we check how the elements from the checklist are supported in the particular site. In section 5, we discuss the value but also limitations of our approach. We conclude the paper with section 6 by pointing to further research needed in this area.

We make the following main contributions to the movement of responsible personalization: first, we create a new definition of transparency in the context of machine generated personalization, second, we develop a set of best practices for the community based on this definition and on prior art, and third, we generate a concrete tool-set to help system designers asses and realize these best practices in their respective systems.

2 The Need for Transparency in AI Based Systems

The speedy uptake of Artifical Intelligence based approaches has raised concerns about their ambivalence. Personalized recommendation may help users find relevant content and items but also introduce bias and pose a risk of manipulation  [Bozdag.2013]. Similarly, algorithmic decision-making may help allocate resources in a fairer manner but also pose risks of discriminating against social groups due to bias in the system  [Zou.2018]. Such risks are especially high if the user is unaware of the processes of personalization and decision making. This is often the case with algorithmic systems, as filtering, classification, personalization, and recommendation remain intransparent or even opaque  [Pasquale.2015, Burrell.2016].

Transparency can help address the concerns voiced about AI based computer systems. First, transparency can balance power asymmetry, empowering users while curtailing the influence of companies on customer behavior. While companies have easy access to user data, users lack knowledge of algorithmic systems  [Lepri.2018]. Especially big players in the information system economy hold enormous power vis à vis users to the extent that they can shape information, knowledge and culture production  [Benkler.2006]. User empowerment by means of transparency and user control may level the playing field.

Second, transparency can increase user autonomy. Recommender systems usually filter content according to preference models that easily create a feedback loop. A classic example is the “filter bubble” in social media platforms  [Pariser.2012, Veltri.2019]. When users lack exposure to information diversity, their autonomy and ability to make independent decisions is impacted  [Mittelstadt.2016]. However, if users understand why and how an algorithm presents information to them, they can better reflect on how sources of information inform their decisions.

Third, transparency can boost privacy rights and user trust in algorithmic systems. For instance, users can only give meaningful informed consent when they understand the risks of algorithmic decision-making  [Mittelstadt.2016].

Fourth, transparency can enable fairness and non-discrimination in algorithmic decision-making. Algorithmic decision-making is becoming ever more pervasive, affecting individuals in pivotal areas of life  [Laat.2018]. While human decision-making reserves the possibility to provide a straightforward face to face explanation of why someone’s application was denied, algorithmic systems are considered too complex for operators to provide a simple answer  [Miller.22062017] (for an opposing view, see  [Zerilli.2019]). Transparency may thus increase subjects’ ability to understand the cause of decisions made by algorithms. Thereby, transparency enables users to assess whether a decision-making process is fair and non-discriminatory  [Abdollahi.C2018, p.2].

While transparency is highly relevant, it is not absolute. Calls for transparency may not always be ethical and warranted. For instance, they depend on the standing of different actors that are involved and interact in algorithmic assemblages  [Kitchin.2017, Seaver.2019, Brill.2015]. For instance, demanding increased transparency on behalf of users (in terms of sharing more data) seems inappropriate given their vulnerability to loss of informational privacy  [Lanzing.2016]. It is thus appropriate to focus attention on promoting transparency from the system design perspective, and increase users understanding of the logic underlying designers’ activities [Kitchin.2017, ONeil.2016, Steiner.2013].

3 Best Practices for Transparency

We take a three step approach to developing best practices for transparency in machine generated personalization. First, we develop a new definition of transparency for algorithmic systems by drawing on prior art. Second, from this definition, we derive best practice for the implementation of transparency in machine generated personalization. Third, we translate these best practices into questions for system designers to be used as a reflection and assessment tool, presented as an online checklist for open usage.

3.1 Step 1: Transparency Definition

To generate a list of best practices, we began by asking: What is transparency in the context of AI systems? When working with the term transparency, we should first clarify the relationship of transparency to principles of ethics. According to Turilli and Floridi  [Turilli.2009], transparency is not an ethics principles itself. Rather, transparency can enable or prevent ethics. In some cases, calls for transparency may for instance inhibit privacy rights. We thus frame transparency not as a principle of ethics but as a practice that can achieve ethics goals such as autonomy and accountability  [Turilli.2009].

We investigated views on transparency from technology ethics, the philosophy of technology, computer sciences, but also ethics guidelines and legal documents. Based on our analysis, we define transparency as follows:

Transparency is a practice of system design that centers on the disclosure of information to users, whereas this information should be understandable to the respective user and provide insights about the system. Specifically, the information disclosed should enable the user to understand why and how the system may produce or why and how it has produced a certain outcome (e.g. why a user received a certain personalized recommendation).

The first important component of transparency is the notion that information must be “understandable.” The user of a system must be able to comprehend the information disclosed to them. For instance, the GDPR [EuropeanParliamentandCounciloftheEuropeanUnion.27April2016] states with regard to data processing that information must be provided in “clear and plain language and it should not contain unfair terms”  [EuropeanParliamentandCounciloftheEuropeanUnion.27April2016, p. 8].

Here, we can see how transparency is a relational concept and a practice. Whether the information provided is transparent depends on the individual user (or data subject), their cognitive abilities, their language skills, and epistemic conditions. In a way, practices of transparency must be personalized to the user at hand, given the diversity of users’ ability to comprehend information  [Vakarelov.2019]. The system thus relates to the user and produces knowledge for them, which makes transparency a performative practice  [Ananny.2018].

Several sources stress the importance of information comprehensibility. According to Chromnik et al.  [Chromnik.20March2019], transparency is an enabling condition for the user to “understand the cause of a decision.” Ananny and Crawford  [Ananny.2018] describe transparency as a form of “seeing” and “understanding” an actor-network. The authors stress that transparency means not merely looking inside a system but across systems. Transparency thus means explaining a model as it interacts with other actors in an algorithmic system  [Ananny.2018]. Floridi et al.  [Floridi.2018] understand transparency as explainability, whereas explainability incorporates both intelligibility and accountability. AI decision-making processes can only be understood if we are able to grasp how models work and who is responsible for the way they work  [Floridi.2018]. For Vakarelov and Rogerson  [Vakarelov.2019], transparency means communication of information under two conditions: information must be a) sufficient and b) accessible. The latter means that the recipient of the information must be able to comprehend and act upon the information.

Another crucial element of transparency is information disclosure about deliberation or decision-making processes. The IEEE Guideline for “Ethically Aligned Design” states that transparency means the possibility to ascertain why a certain decision was made  [InstituteofElectricalandElectronicsEngineers.2019]. For Turilli and Floridi [Turilli.2009], disclosing information refers to communication about the deliberation process, i.e. how information came about. The rationale here is that the deliberation process reveals the values that guide organizations or system designers in their everyday practices and illustrate how they make decisions.

Similarly, for Tene and Polonetsky  [Tene.2013], transparency refers to the revelation of information about criteria used in decision-making processes. The disclosure of the dataset (or its existence) is less relevant than the actual factors (such as inferences made from the data) that inform a model and its effects on users. Also Zerilli et al. [Zerilli.2019] argue that, similar to explanations in human decision-making, a system should reveal factors in decision-making and how they might be weighted.

Dahl  [Dahl.2018] even argues that it is not necessary to reveal the inner working of a model for the user to determine whether a system is trustworthy. Rather, transparency means providing key details about how the results came about or offering expert testimony about how the system usually works. Burrell  [Burrell.2016, p. 9]

suggests that improving interpretability of models is crucial to reduce opacity: “One approach to building more interpretable classifiers is to implement an end user facing component to provide not only the classification outcome, but also exposing some of the logic of this classification.”

Finally, there can be an element of participation in transparency. The user is expected to assess the system with regard to its trustworthiness based on the information that is disclosed. Furthermore, the user may become active in choosing between different models, i.e. different options of personalization  [Simon.2010]. The user is thus becoming involved in the process of transparency which increases user control while interacting with the system.

3.2 Step 2: Best Practices

From our definition of transparency, we derived nine principles of transparency for responsible personalization. They reflect the three core elements of transparency: information provided must be understandable to users, information must be disclosed about why and how a model reaches a certain outcome, and users should have a say in personalization processes. The best practices further reflect additional needs for information about the data collection processes, the composition of datasets, the functionality of a model, the responsibilities for the model or system, and how the model may interact with other models across algorithmic systems.

Table 1 shows the list of the best practices as well as the sources on which these practices build. It also identifies the relevant system architecture components relevant for each best practice based on the Input-Processing-Output architecture model [boell2012conceptualizing]. These components include: ”Input” for transparency relating to the data used by the system, ”Processing” for transparency relating to system models and ”Output” for presenting the transparent information to the user. We extend this architecture with a ”Control” component to represent the control given to the user over the system’s personalization behaviour.

We define user control as the possibility of users to interact with the system to adjust elements thereof to their respective needs and preferences. It is important that users not only “feel” that they have control because this can put them at risk of exploitation. If users think that they have control, they might feel encouraged to share more data  [Tene.2013]. User control is thus of particular ethical sensitivity and significance as it relates directly to the autonomy of a person.

The best practices above are necessarily vague and generic, which allows system designers in the personalization community to adapt them to their context. Concrete measures for the implementation of transparency must thus be specified according to use case. The best practices may even be used by designers in other domains such as incentive design or decision-making and classification technology. Finally, the best practices described above may not be equally important to achieve goals of transparency, such as user autonomy. Based on the analysis in step 1, particular relevance can be ascribed to practices 1, 3, and 8. The emphasis on these practices also informs the questions in the checklist.

No. Component Description of transparency standard Sources
1 Input, Processing, Output, Control Disclosing accessible and actionable information, meaning that the user can comprehend and act upon the information  [Vakarelov.2019, EuropeanParliamentandCounciloftheEuropeanUnion.27April2016]
2 Input, Processing Disclosing relevant and detailed information about data collection and processing; this includes notification of data collected for personalization, information about pre-processing and possible biases in the dataset  [EuropeanParliamentandCounciloftheEuropeanUnion.27April2016, Buolamwini.2018, ACMCouncil.22June2018]
3 Processing Disclosing relevant and detailed information about the goals of the designer/system, the reasoning of a system, the factors and criteria used (potentially also how they are weighted), as well as the inferences made to reach an algorithmic decision  [Turilli.2009, Floridi.2018, Burrell.2016, Zerilli.2019, Tene.2013, Chromnik.20March2019, Srmo.2005, InstituteofElectricalandElectronicsEngineers.2019, Abdollahi.C2018]
4 Processing If appropriate and helpful, providing expert testimony (e.g. by a member of the design team) about how a system works and reaches a certain outcome, including information about the stochastic nature of a model as well as lab accuracy performance of a model  [Dahl.2018]
5 Processing Possibly disclosing information about how the algorithmic model may affect the user and how the model may interact with other models across algorithmic systems  [Ananny.2018]
6 Output Disclosing that a machine is communicating with the user and not a real person  [link_latlng_03]
7 Output Disclosing information about those responsible for the model (e.g. name of the company or designer)  [Floridi.2018]
8 Control Proposing alternative choices for user interaction with the system, e.g. different options for personalization  [Simon.2010]
9 Control Providing the user with opportunities to give feedback about personalization; providing the user with opportunities to specify their goals as these goals are expected to drive personalization  [heer2019agency]
Table 1: Transparency Best Practices for Machine Generated Personalization.

3.3 Step 3: Checklist

Based on steps 1 and 2, we can now move to define a checklist for systems designers to assess the transparancy of machine generated personalization. We map each architecture component in Table 1 (namely Input, Processing, Output, Control) to a section in the checklist. Questions for each section are then derived from the best practices uncovered in the previous steps. In this process, we prioritize some best practices that were overwhelmingly affirmed by the literature.

The resulting checklist is given at:

The checklist includes a total of 23 questions. After filling it, the system designer can download a PDF file with their responses. They can also print an empty copy of the checklist to be filled offline if needed.

To arrive at a comprehensible and user friendly checklist, we omitted some questions. If system designers wanted to attempt at particularly high standards of transparency, they could also answer the following additional questions:

  • Does the system explain to the user how the model(s) may interact with other models in algorithmic systems?

  • Does someone from the design team provide expert testimony to the users about how the model(s) works (e.g. in a video)?

We note that the checklist is supplied as an assessment tool for system designers, enabling them to identify areas in their system which suffer from lack of transparency as well as point to imbalances between the transparency aspects of a system and the control it gives users over its operation. Ideally, a system designer has implemented transparency so that they can check yes for every question. However, the goal should not be to score high on the checklist but rather to have an honest assessment and decide on priorities and next steps.

4 Case Study: Applying the checklist

We perform an initial application of the proposed checklist as a reflective and assessment tool for the following online services that use personalization: Facebook, Netflix, YouTube, Spotify, and Amazon. For each of these destinations, we take a system designers point of view, and ask ”how are the transparency elements from the checklist supported on this particular site?”, when examining the information available to registered users on the sites.

Table 2 present our application of the checklist to Facebook. As can be seen from the table, while some transparency elements are well established on this site, other elements are only partially supported or not supported at all and should be considered for future improvement.

Question Reply Details
Does the system inform the user about the purpose of personalization? Yes
Does the system inform the user who developed the technology and is liable in cases of wrongdoing? Yes
Does the system inform the user about their rights under data protection law? Partial Local law rights are not specified.
Does the system inform the user about possible risks of engaging with the system? No Risks are not specified.
Have users given informed consent about the collection, processing, and storage of their data? Partial Default data collection policies are not specified.
Does the system inform the user about the fact that data is collected for personalization? Yes
Does the system inform the user about which data is collected to produce personalized content for them? Partial Details are missing.
Does the system inform the user about pre-processing done with the data collected for personalization purposes? No Pre-processing of data is not explained.
Does the system inform the user if their data is used and shared beyond the goals of personalization? Partial Information about sharing data with partners is given without sufficient details as to the use of this data.
Does the system inform the user about the kind of data that is processed to create a certain personalized item? Partial The link between data sources and personalization is not clear.
Does the system explain to the user why they are receiving a certain personalization? Partial The notion of personalization is generally mentioned but not specified enough.
Does the system inform the user about the behavioural models underlying the personalization system? No Missing information about models used.
Does the system inform the user about possible constraints of the model such that may result from pre-processing or biases in the dataset? No Missing information about models constraints.
Does the system present information to the user in a location where they can notice it and access it easily? Partial Hard to find the links to this data. Visibility and accessibility are lacking.
Does the system provide information to the user in a comprehensible way and can they act upon this information? Partial Setting option is hard to understand and follow.
Does the system provide the user with information in a clear and simple language that avoids technical terms? Yes
Does the system make it clear to the user that they interact with a machine? Yes
Does the system provide the user with the opportunity to specify their goals which are then used for personalization? No Missing capability.
Does the system provide the user with different options as to the personalized content they receive? Partial Notification control is good. Ads control is poor. Data control is very partial.
Does the system provide the user with opt-in and opt-out options (e.g. for data collection)? Partial Complicated. Users have to control each option in separation.
If applicable, can the user adjust frequency and timing of personalized content? Partial. Is not supported for some content.
Does the user have a say in which data or models are used for personalization? Partial Users cannot fully understand the connection between data and personalization.
Does the system encourage the user to give feedback and express their opinion about the personalization mechanisms used (type, frequency, duration, etc.)? No Feedback is not strongly encouraged.
Table 2: Preliminary Checklist Application to Facebook

To perform a preliminary comparison between the different sites and between the different sections of the checklist for each site, we also compute the percentage of ”Yes” and ”Partial” replies for each checklist section, when a ”Partial” reply is weighted as 50% positive. Namely, for each checklist question, we give a ”Yes” reply a value of 1 and a ”Partial” reply a value of 0.5. We then sum these values for each section and divide it by the total number of questions in the corresponding section. This computation, while being limited and potentially biased due to the subjective filling of the checklist by the research team, may offer comparative information about the different sites and between the different checklist sections. Figure 1 presents the result of this comparison. We further discuss these results in the next section.

Figure 1: Preliminary checklist, online sites: Y-axis is the percentage of positive and partial replies in each checklist section

5 Discussion

The major advantage of the transparency checklist is that it helps system designers understand where they are strong on transparency and where improvements are needed. Looking at Figure 1 and the online systems we have examined from the designer perspective, we notice that they primarily focus on realizing transparency in the ”Input” category, i.e. with regard to data collection and the handling of user data. They are particularly weak in providing information about why and how models bring about certain personalization (”Processing”). They also lack participatory elements such as offering the user different options of personalization or allowing the user to supply feedback to the system (”Control”).

This trend to follow best practices of data or ”Input” transparency may be attributable to the rise of data protection laws such as the GDPR. System designers so far pay less attention to transparency about the reasoning and underlying logic of personalization. This finding is in line with a recent study on guidelines for human-computer interaction by Amershi et al. [Amershi.2019]. In that study the authors note that ”Guideline 11 - Make clear why the system did what it did - had one of the highest number of violations”. This is a severe shortcoming as ethics and philosophical work on transparency in algorithmic system clearly identify the need to disclose information about how a certain outcome (personalization) emerged. Making processing-related information transparent does not necessarily mean cracking open and looking inside the system, but rather providing meaningful and understandable information about the goals of personalization as well as the factors that contribute to making a personalized recommendation.

We suspect that transparency about the reasoning of a system will gain relevance in the future. In fact, there is an ongoing debate whether the GDPR even provides a legal right to receive an explanation for algorithmic decision-making  [Wachter.2016]. Legal cases in the future will shed light on such questions and eventually, disclosure of why and how a computer model caused a certain outcome may become customary practice.

Literature also points to the need for user control to fulfill transparency  [Simon.2010]. Users should be provided with different options of personalization to align best with their personal goals and increase their autonomy. Our application of the checklist points to significant shortcomings in the realm of ”user control.” As a system designer, having applied the checklist and seen some blind spots, one would now be able to make a deliberate decision about whether to increase user control in one’s own system.

Another advantage of the checklist is that it can be used as an assessment tool, not just internally for self-assessment but also as an openly accessible evaluation of a system’s transparency performance. An online service may commission a ”transparency check” by an independent organization to assess the system’s trustworthiness.

Studies show that transparency can be a competitive advantage of companies  [Buell.2019], and thus companies may have an interest in providing information about the transparency levels of their online services. The desire for independent audits may increase in the future with movements to certify ”trustworthy use of Artificial Intelligence”  [FraunhoferInstituteforIntelligentAnalysisandInformationSystems.2019].

Beyond commissioned reviews of a system’s transparency performance, users and activists may employ the transparency checklist as a means of control and oversight of online services. A comparison of online services’ transparency performance as in figure 1 exposes the brands behind them and may generate pressure to implement increased transparency.

We now point to some limitations of our approach. First, we note that the idea or ”ideal” of transparency itself has limits  [Ananny.2018]. For instance, transparency rests on the idea that something can be known  [Ananny.2018]. There is no guarantee that we succeed in understanding a model, even when transparency is in place. This can be due to lack of resources, human capital  [Mittelstadt.2016], and lack of basic digital or technical literacy  [Burrell.2016]. Disclosing information can also confuse users and not add to clarity or insight  [Bannister.2011, Ananny.2018].

Transparency may further clash with important ethical principles such as privacy. Full disclosure of input or output data may put users at risk of being re-identified, especially in areas like finance and medicine  [Lepri.2018]. “For example, facilitating access to electronic medical records may help life-saving research, but may simultaneously expose patients to fraud or a breach of privacy, as their personal information is disclosed”  [Turilli.2009, p. 106]. Business interests may also be legitimate reasons to reject full disclosure  [Lepri.2018].

Finally, there is concern that users may “game the system”  [Laat.2018, p. 528]. In the example of credit distribution or decisions-making about loans, it is argued that some components of an algorithm must be kept secret so that users cannot manipulate their input data to receive the desired outcome  [Lepri.2018].

These limitations of transparency also put a checklist in perspective. Whether transparency is appropriate or warranted depends on the unique use case. Different priorities such as legal requirements, business interests, and resources for transparency implementation must be weighted individually. The best practices and accompanying transparency checklist are necessarily vague as to the scope of transparency.

It remains an open question how much and what kind of transparency should be provided. These are questions for the personalization community or the respective design teams. Zerilli et al.  [Zerilli.2019] argue that algorithmic systems are increasingly held to higher standards than human decision-making. They advocate against “opening the black box” and instead suggest providing information about the factors weighted in decisions  [Zerilli.2019, p. 665]. On the other hand, Ananny and Crawford  [Ananny.2018] argue that algorithmic systems cannot be fully comprehended unless we reveal the complex relationships between different components and human as well as non-human actors in algorithmic networks  [Ananny.2018]. Answers to the question of scope and type of information may also vary depending on the type of algorithms and audiences involved  [Laat.2018].

The urgency of transparency in algorithmic systems depends on the context. In some cases of online nudging, transparency may not be essential to user autonomy. One could think of a software that supports users in learning languages. Nudging the user to spend more time with the tool and getting to the next level may not present ethically problematic implications. However, trying to addict a user to a certain game or activity seems rather precarious. This has to do with the implications for the user’s autonomy. User autonomy is violated when the interests of the user are marginalized vis à vis the interests of an organization  [Mittelstadt.2016]. There is a fine line between acceptable and abusive practices. While increased transparency can benefit the user in many cases, the need for transparency has to be determined case by case.

Another significant issue concerns the relationship of information to the user (and in that sense, transparency is a relational concept). The same information may make something transparent to one group or individual but not to others  [Vakarelov.2019]. It follows that transparency must be configured to the individual user. In fact, we may need a personalization technology to fulfill the transparency best practices for machine generated personalizaion [Kouki.2019, schneider2019personalized].

We should note here that, while an ethics perspective promotes user control and meaningful transparency, it is not certain that users desire transparency and control. From extensive work in the field of privacy and data protection, we know that users claim privacy to be an important issue for them but rarely take steps to protect their data (“privacy paradox”)  [Gerber.2018]. Similar dynamics may apply to transparency. Nevertheless, independent of users’ personal preferences, users should have the opportunity to take advantage of transparency. Even when users disregard information provided to them, system designers have an ethical responsibility to implement transparency best practices.

6 Conclusion and Future Work

In this work, we have developed best practices for implementing transparency in machine generated personalization. We have further created a checklist for system designers to assist designers in realizing the best practices. The proposed tool includes 23 questions in four areas: Input, Processing, Output, and Control. While transparency needs may vary depending on the use case, the checklist can be used as a supporting instrument that guides system designers in embedding transparency into their work. We have demonstrated such a use by a preliminary application of the checklist from a system designer perpective to prominent online services that use personalization.

While we propose a first transparency and user control checklist, we recognize that it may be amended in future engagement with researchers and system designers. Ideally, the items in the transparency checklist should be discussed by experts in the field and present a consensus of the personalization community  [aczel2019consensus]. This can increase the checklist’s likelihood of adoption. We therefore encourage researchers, funding agencies, and journals to provide feedback and recommendations. Furthermore, tangible design actions based on the best practices have to be developed in future work. Tutorials and workshops may invite system designers to apply the checklist and create innovative design solutions that implement transparency in their respective systems.

Several questions emerge from our definition and identification of transparency best practices. Future work can engage with transparency conceptually, its ambiguities, but also with regard to possible implementation. What is the subject and scope of transparency in different areas where algorithmic systems are in place? We are particularly intrigued by the question of how transparency can be personalized given the diversity of capabilities and transparency needs of users. Responsible personalization may thus need personalized transparency. This project has received funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No 823783. The authors would like to kindly thank PD Dr. Jessica Heesen for her valuable comments on draft versions of the paper. We further thank our colleague Michael Lohaus for insights from discussions of the paper.