Abstract

In the following text, the application of profiling techniques used with the objective to depict the most precise image of an individual is investigated. The exploitation of user knowledge by businesses and institutions to manipulate their audience with personalized contents is discussed. Ethical issues related to profiling, as well as ways for users to gain awareness about the examination made possible by their data, are addressed. The core of the thesis elaborates on a proposal for a fictitious tool designed to inform users about and empower them against classification. This tool is embedded in a near-future scenario in which, similarly to the General Data Protection Regulation (GDPR), online data is made accessible to their users, is erasable and regulated to a greater extent. Finally, the constructed prospect is evaluated and a personal reflection and outlook formulated.

1. Foreword

2. Comprehension of consumers is the new oil

‣ 2.1 Criticism

‣ 2.2 Protection and alternatives

‣ 2.3 Anonymity is not the answer

The aforementioned proposal cannot solve the core of privacy issues, as it relies on anonymity to protect user-integrity. Anonymity does protect the user from being identified, however, it also has various disadvantages. “It allows some users to break the law or violate the rights of others, such as defamation, cybercrimes, bullying, propagation of racist or anti-Semitic ideas, involving damage to other individuals or to society as a whole” (24).

Furthermore, anonymity cannot be guaranteed in every domain. Agreements and contracts concluded with a bank or insurance – most of them now being concluded online – require identity verification.

A chance to transition back to fully anonymous cyberspace is unrealistic and not intended. Due to fears of exclusion or lack of alternatives, removing an account is not an option.

As Rainer Rehak declares in his talk at the 2018 Chaos Communication Congress (35C3) (25), the core of the issue constitutes finding ways to protect the ‘weak’ actors (i.e. citizens or consumers) by regulating the ‘strong’ actors (i.e. businesses and governments). Let alone, regulating is not enough. In the physical world, people reveal their true identity and take responsibilities for their acts. When interacting in public contexts, social conventions prevent negative judgment and attracting attention. Similarly, conventions of appropriate online behavior should be developed to diminish the damages of manipulation.

‣ 2.4 Manipulation

In March 2018 the whistleblower Christopher Wylie released a cache of documents prompting the Facebook–Cambridge Analytica (CA) data scandal (29). Wylie worked for CA as a data consultant and was able to give detailed information on how CA collected and analyzed data to enable psychological manipulation by political parties.

In 2014, a psychologist at the Cambridge University, Aleksandr Kogan, published an app on Facebook called „This Is Your Digital Life“. Each visitor of the app was paid $3 to $4 to complete a survey (30). The app did not only analyze the user’s profile with all of its likes, posts, comments and more, but also those of its whole network. That is how Kogan was able to collect data from approximately 87 million Facebook users (31). He developed and used techniques such as the OCEAN model, which is explained later, to create psychological profiles.

Data and analyses enabled Kogan to list relations between factors such as habitat, sex or comments to estimate political views, life-satisfaction or fears. CA advised on how to influence and manipulate the targeted audience. Donald Trump used CA services (32) to shape people’s thinking and to manipulate the votes. He conducted campaigns that discouraged potential Clinton-supporters to question their choice (33). Ted Cruz, spent around 5.8 million US$ (34) to influence attitudes, fan fears or to communicate with each user in a personalized way. According to election forecasts, Ted Cruz’s percentage grew from 5% up to more than 35% (35). Selecting the target group and inspecting specific demographics, psychographics and personality traits facilitate the communication with the group (36).

The practice of targeting single individuals by analyzing their personality is called microtargeting (37). Specifically prepared and personalized ads, statements or fake news are spread into the interface of each user.

CA is not the only company who does and sells microtargeting and profile analyses. Companies like TargetPoint (38) or Grassroots Targeting (39) use similar techniques.

As Antonio García Martínez said, “none of this is even novel: It’s best practice for any smart Facebook advertiser. Custom Audiences was launched almost six (!) [sic!] years ago, marketed publicly at the time, and only now is becoming a mainstream talking point. The ads auction has been studied by marketers and academics for even longer. The only surprise is how surprising it can still seem to many” (40).

The scandal is particularly disturbing, not only because CA ‘stole’ and analyzed information, but also since it counted on a psychological approach to manipulate.

‣ 2.5 Psychological profiles

Methods for psychological profiling are attempts to classify individuals in holistic categories for which cultural, social and economic backgrounds have no influence. As the American psychologist Gerard Saucier declares, “an optimal model will be replicable across methods, cross-culturally generalizable, comprehensive, and high in utility” (41).

The aim of developing systems to categorize individuals arose a long time ago. 300 BC already, Aristotle defined different kinds of human beings (42). The psychological models described below have been established around 100 years ago.

The DISG Theory of William Moulton Marston is based on the assumptions that every ‘normal’ person (without mental illness) can be categorized using four criteria that relatively define a person's traits: dominance, influence, steadiness, and conscientiousness (DISG) (43).

The Enneagram of Personality (44) relies on nine terms to categorize base personality types in a subjective way. Each of these types is combined with ten different perspectives such as ‘Basic fear’, ‘Ego fixation’, ‘Passion’ or ‘Stress/Disintegration’ (45).

As opposed to the abovementioned model, Carl Gustav Jung’s methods divides human beings into rational and irrational people. Each can be split into four subcategories (‘thinking’, ‘feeling’, ‘sensing’ and ‘intuition’ (46)), producing eight different kinds of personalities. Additionally, each person is classified as either extraverted or introverted.

The Big Five personality traits model or also called the OCEAN model was used by Cambridge-Analytica for microtargeting. One of its advantages is its ability to visualize an infinite amount of personalities, as it is not limited to binary classifications and produces individual and unique results. Another difference is the sort of user data that constitutes the source of the analysis. While in most models, users are interviewed or observed, the OCEAN model takes advantage of existing textual data (47).

Martin Gerlach, a former physicist at the Northwestern University, states that we “[…] don't have enough empirical evidence [to show] that something like this [the concept of personality types] really exists” (48). But first results and evaluations also showed that factors for these types “are at least partially genetically predetermined” (49) and thereby enable the estimation of personalities. “It is not yet clear that this is the ‘optimal’ model“ (50) says Saucier, but Richard Robins, a personality researcher at the University of California, argues that “this is by far the most valid estimate we have of how people cluster into types” (51).

The ability of the OCEAN model to express the uniqueness of people reinforces the effectivity of classification and its manipulating character.

‣ 2.6 The limits of classification

In the case of content personalization, the classification of users is inevitable, because marketing experts are required to use understandable terminology in order to define their target groups. When creating an advertising campaign on Facebook, marketers are invited to select among a list of classifications that characterize the target audience (52). Thus, a user with a ‘complicated’ or ‘engaged’ relationship can be picked out of the mass and addressed with tailored advertisements.

As a result of semantic selection, the user is reduced to a description which is limited by the boundaries of the vocabulary and is interpreted subjectively. The selection of words for categorization constitutes a source of marginalization. The word negro, for instance, was first used around 1440 by the Spanish and Portuguese to describe people of dark-colored skin, as the word literally means ‘black’ (53). However, the term is associated with colonial history and is found offensive (54). In 2010, the word was dropped from the U.S. census and was limited to ‘black’ or ‘African American’ (55).

The classification of gender is another example of intensive social debate, in which vocabulary plays a central role. ‘Woman’ and ‘man’ as gender descriptors are considered a source of discrimination (56). In the 70s, the LGBT community has questioned the immutable perception of the sex, both in its social/cultural and physical definition (57). From this time forward, a series of different terms defining diverse gender forms and sexual orientations have emerged. Sam Killermann, a social justice advocate, lists on his website itspronouncedmetrosexual.com a list of more than 50 LGBTQ+ vocabulary definitions, such as androsexual, bigender or skoliosexual (58). Those recent terminologies are attempts to establish norms including people that do not feel represented fairly by the heteronormative gender definitions.

Although the latter might be more appropriate to describe the variety of gender perception or sexual orientations, these definitions are still limited and remain free of interpretation. Likewise, they are charged with cultural values.

All the GAFAM (Google, Apple, Facebook, Amazon, and Microsoft) are US based companies that carry western values and use English to describe their data, despite the globality and multiculturality of their audience.

Both the vocabulary used for classification and the manipulation they enable represent complex issues. The selection of labels will certainly evolve and become less discriminatory, yet labeling will still be crucial in online profiling.

‣ 2.7 Avoiding the classifications

We assume that by making labeling unreliable, we can minimize the risks of manipulation. Personalized content that relies on target groups should become inaccurate and inefficient. However, we do not believe that this inaccuracy should be provoked by chaos, because it could have negative effects on the user experience. Rather, the risks should be minimized by unsharpening the precision of the classifications and by reducing the amount of available data.

Unsharpening and limiting classification require internet users to realize that all data is a source of labeling. To gain awareness, relatability to individual classification is necessary.

The possibility to undertake changes is given, provided the necessary consciousness is reached. Yet knowing how to enable action requires knowledge about how individual classification mechanisms work. Unfortunately, because they mostly rely on machine learning, there is no generic answer. Reverse engineering the services’ mechanisms might help to foresee how they function. Accordingly, online conventions that suggest actions could be developed. They could consist of the following principles:

• Reducing the amount of data limits the source material for classification
• Ensuring that classifications become as vague or as uncertain as possible, reduces their effectiveness
• Diminishing the match-level of classifications can avoid the assignment into target groups
• Countering a classification by strengthening a contradictory classification, could help to dull the match-level of both classifications

In the proposal that follows, attention is focused on the first three principles, as the last one deteriorates the user experience.

3. The App - Labelless

‣ 3.1 Initiation and access grant

The purpose of the tool is delicate and its approach complex. The users can be expected to instantly comprehend neither all the intricacies of the issue nor the visual codification of the tool. Therefore, they will be initiated to the application and its purpose in a step by step introduction. Once they are familiarized with the concepts, the visitors are invited to connect service for which an analysis should be created. Using a ‘single sign-on’ mechanism such as the typical ‘login with Facebook’ procedure, tools are linked with the services’ APIs. Once the connection is established, the results of the processed data are displayed.

‣ 3.2 Visualizing the classifications

The proposal uses a visual language that codifies various aspects of the user-classification.

Each individual categorization is represented by a circle or, in other words, a bubble. Each of them communicates two vital types of metrics:

The size describes the match-level

The size of a circle describes how strongly the subject is associated with a category. To indicate the level of correspondence with a descriptive term, the bubble with the highest match-level appears the largest and vice versa. Additionally, each bubble is constituted of three circles, that reflect the uncertainty of the estimation. The outer circle defines the maximum value, the middle circle represents the average and the smallest the minimum estimation. The distances between those three circles suggest the precision of the classification. A bubble with tightly close circles appears sharper and should alert the visitor that his data makes precise classification possible. On the other hand, bubbles with circles of dissimilar radius are considered less precise and therefore less reliable.

The color describes the amount of underlying data

We assume that large quantities of data constitute a greater source for classification. Therefore, we consider crucial to convey the amount of information on which the estimations are based. We use a color scale going from blue to pink to color-code each bubble. Less saturated blue circles are based on fewer data and highly saturated pink circles on more information.

These two representative dimensions are key to the understanding of our interface. The use of both size and color is replicated in the different visualizations of the tool.

‣ 3.3 The views

The tool offers two views: The so-called cluster-view gives an overview of the average match-level for each category. The so-called grid-view explains with greater detail the data-based origins of the estimations.

The cluster-view

As its name suggests, the cluster-view visualizes each classification in organized and understandable groups that form clusters and are placed in cloud-like formations on the canvas. Within each cluster, the bubbles are physically attracted, as if they were magnets or charged with gravitational force. As a consequence, the biggest bubbles occupy the center of the groups and the smallest reside on the margin.

Hovering a bubble displays a tooltip that indicates numeric metrics about the match-level and the amount of underlying data.

The grid-view

The grid-view offers details about the estimations’ source accessible through two perspectives.

Data perspective:

This representation allows the most precise examination. It displays what estimations are made in each data node and with which match-level.

In the left sidebar, the same groups as in the cluster-view are listed. When a group is expanded, its children classifications are revealed. On the main area, right of the sidebar, a grid of thin grey lines represents the data-nodes such as posts, likes, website visits, uploaded pictures, or any piece of data used for classification.

On some of the intersections between horizontal (classification) and vertical lines (data element), a blue bar indicates the degree at which the data element is associated with the corresponding category. As in the bubbles of the cluster-view, the bar is constituted of three shades that illustrate the uncertainty of the match-level. Hovering on either a category, a data-node or on a bar displays a tooltip filled with detailed metrics.

On a group’s row, the blue bars are radially oriented, summarized in a star-like shape. The lopsided distribution of the star’s spikes can express if a data element is the source of an intensive classification, and reveal the sharpness of the estimation.

Time perspective:

An additional inspection angle is provided by the time-based summary of the categories. This view is organized in the same way as in the data perspective. However, instead of showing the individual data nodes, the x-axis displays the classification over time. Each term is illustrated by a line chart. Like the bubbles, the chart is colored to demonstrate the amount of underlying data. The gradient colorization is due to the progressive transition from a month’s color to the next’s. Like in the data perspective, group rows summarize child classifications with a star-like form. Yet on this mode, each branch of the star is colorized to illustrate the amount of underlying data.

Wave chart:
The choice for the particular form, as opposed to a simple line chart, is justified as an analogy to the bubble representation. In fact, the bubble and its three circles can also be seen as a mirrored bar chart (like in the data perspective).

‣ 3.4 Filtering

Familiarized users can filter to concentrate their focus on classifications that match particular criteria. The two filters currently proposed are:

• Filtering by size:
  Filtering by size can be very useful to concentrate on bubbles of higher match-levels by excluding the smaller ones.

• Filtering by color:
  Filtering by color can be useful to identify the bubbles relying on vast amounts of data and potentially reliably usable for manipulation.

• Combining filters:
  Both filters can be combined to identify which estimations can be minimized easily. Those appear blue, big and sharp as they are base on few data are estimated with confidence.

‣ 3.5 Setting goals

For each category, users are invited to set a match-level goal. By dragging an elastic slider, a bordered circle representing the well-aimed objective is displayed on the bubble. While the interaction is ongoing, tips progressively slide in a sidebar on the right. Each tip is accompanied by an indication of the potential effect it has on the bubble.

The more challenging the goal, the more impactful the tips. These can consist of either the deletion or creation of data. Modifications that are executable via the services’ APIs can be performed without leaving the tool. The creation of alternative content, however, must be done by the users themselves. Once set, the objective circles are displayed on the bubbles and visible from the cluster-view. This way, the evolution of the classification can be observed in a later visit.

‣ 3.6 Design Evaluation

Strengths

• The choice of the visual language of the tool allows a detailed analysis of the user’s classification.
• The representation of each category with circles and the color coding, once comprehended, facilitates their investigation.
• The visualization of the match-level, as well as its uncertainty, are visualized consistently throughout the whole design.
• The form selection is adapted to the summarization of individual classifications, both over time and in total.
• The interactions, because of their visual appropriation, feel natural and contribute to the familiarization with the complex concepts of the tool
• The two views enable different depths of investigation

Improvement potential

• The choice of size- and color-scale is not well suited for comparison and only permits a rough approximation.
• The list of provided classifications is also far from being exhaustive.
• The choice of vocabulary used in the groups’ titles, as well as the selected terms for the categories inside those groups, are inspired by Facebook’s ad-settings. However, deeper research could help to appropriately justify this selection.
• The color-scale is certainly not accessible for color blind people and is potentially hardly differentiable.
• The use of light shades results in low contrast that can disadvantage people with limited eyesight.
• The relatedness between various categories is not extensively visualized.
• The indication of which service is associated with an estimation is not emphasized.
• The tool could alert the users more precisely on high priority risks.
• Currently, the designs do not support a mobile investigation.

Next steps

• A prototype that implements an explorable introduction and schemes the underlying purpose of the tool can be developed.
• Extended filters (data-type, sharpness…) can enable a more precise inspection.
• To improve the understanding of the personal classifications, a comparison with others can be supplied

4. Conclusion

Profiling is complex and finding solutions that seem to fully protect the users seem impossible. Our thesis relies on the quite pessimistic assumption that manipulative and dishonest profiling will continue to happen. Yet we did not want to do Critical Design in a way that either depicts a utopic future or a dystopic one, by just pointing an issue by the finger, without offering any kind of solution.

We identified labeling as an effective influencing method. As a result, the app focuses on visualizing and modifying the extent of those classification estimations. There are certainly other ways to protect users, not covered by our tool. There are also other forms of manipulation that are not limited to the classification of users for which our tool provides no solution and that might have greater manipulative impacts. However, we decided to focus on user labeling because it is understandable by and accessible to most and because we assume that all practitioners of profiling rely on terminology to classify their audience. Yet in spite of profiling, we do not think that all forms of manipulations are nocive to the users. Although, in our opinion, even a manipulation that serves an honest goal should not be done in the first place when not communicated transparently and comprehensively.

We are aware that by suggesting an approach to a hypothetical and relatively effective solution, we make ourselves subject to critique. For instance, we do not clarify who is to develop the tool we propose and which interests those could have. Would the tool be commercial and if so, would it not be for the ‘rich’ instead of for the ‘weak’? Moreover, it is not clear whether the reverse-engineering mechanisms behind the metrics of our proposal would work, and whether the produced classifications would be reliable enough to produce actionable tips. Observing the app, the services could find alternative, non-identifiable strategies to counter-effect the algorithms.

The evolution of technology has made the world smarter, smaller and more connected. Due to the effectiveness of this evolution, it is our responsibility to take control of individual data. Technology has and will continue to evolve. Keeping abreast with this development is important in order to avoid manipulation.

Our reason to study interface design originates from our fascination for technology. Yet the pace at which it has evolved is so high, that we doubt our capacity to be able to follow it, to understand it and to educate the following generations in a way that prevents great disbalances of power between those who know and those who do not. We guess that this problem is as old as humanity, universal and not limited to technology. Yet our thesis relies on the necessity we see to address this topic with different solution-based approaches. We hope to stimulate reflection and to nourish the debate with a visual rather than a technical approach.

5. Personal reflection

The realization of this thesis has enabled us to deep dive in the very present topics of privacy, big data scandals, profiling, targeted advertisement, microtargeting, or technologies that can offer alternatives to the world wide web. Despite our initial, very limited knowledge and naive grasp of the topic, we were able to learn a great deal and to form the dawn of our opinion. We only scratched the surface and are frankly overwhelmed by its immensity. However, we are neither discouraged nor weary. On the contrary, we developed a great interest and are confident that we will keep cultivating related knowledge.

As mentioned in the prologue, we live in a technology-versed surrounding that doubtlessly influences our perspective. The degree at which we consider profiling to be an issue might also originate from our social bubble and professional situation. Probably, many people neither consider profiling an issue nor have the luxury to question it. Nevertheless, hence the increasing dependence of humanity on technology, and because as interface designers we participate in its development, we believe that it is our role to question such topics. As a matter of fact, whilst our studies at the University of Applied Sciences, we have been encouraged to do so. Consequently, we wish to pursue our life, as professional designers but also as users, aware and critical.

6. References and Sources