Profiling is a highly evocative term with multiple meanings, used in both specialist and non-specialist contexts. Whereas the literature on statistics does not pay specific attention to definitions and tends to focus on technical aspects (e.g. data mining techniques and predictive models), providing a definition appears an issue among socio-legal scholars and policy makers. However a widely shared definition has not yet emerged.

Gary T. Marx gave one of the oldest definitions of profiling in a paper that analyses systems of data searching. Profiling (defined by the author in contrast to “matching”) is defined by stressing the logic behind it: “the logic of profiling is more indirect than that of matching. It follows an inductive logic in seeking clues that will increase the probability of discovering infractions relative to random searches. Profiling permits investigators to correlate a number of distinct data items in order to assess how close a person or event comes to a predetermined characterization or model of infraction”.⁷ According to the author’s background, this definition is strictly related to the law enforcement domain.

Almost 10 years later, Roger Clarke defined profiling as a “dataveillance technique (…) whereby a set of characteristics of a particular class of person is inferred from past experience, and data-holdings are then searched for individuals with a close fit to that set of characteristics”.⁸

A legal scholar, Bygrave again stressed: “profiling is the inference of a set of characteristics (profile) about an individual person or collective entity and the subsequent treatment of that person/entity or other persons/entities in the light of these characteristics”.⁹

Later on, Mireille Hildebrandt was the one who put the best effort to precisely define profiling and its distinctive features and the working definition proposed here has built on her work. She defines profiling as “the process of ‘discovering’ patterns in data in databases that can be used to identify or represent a human or nonhuman subject (individual or group) and/or the application of profiles (sets of correlated data) to individuate and represent an individual subject or to identify a subject as a member of a group (which can be an existing community or a discovered category).”¹⁰

Profiling creates a new form of knowledge that makes visible patterns that are otherwise “invisible to the naked human eye”.¹¹ They are based on correlations found in data sets, and cannot be “equated with causes or reasons without further inquiry; they are probabilistic knowledge.”¹² Profiling represents a shift from the idea that knowledge is the result of tested hypothesis. It generates hypotheses: “the correlations as such become the ‘pertinent’ information, triggering questions and suppositions”.¹³ Consequently profiling fosters new forms of generating and applying knowledge. Due to the growing capacities of databases, and capabilities of advanced analysis profiling procedures become increasingly complex. In this context the human role in interpreting data changes significantly.

As pointed out by Hildebrandt, profiling can be categorized into non-automated, automated and autonomic profiling. Non-automated profiling is a form of reasoning that does not rely on any process of automation. Automated profiling is based on “automated functions that collect and aggregate data” and develop into “automation technologies that can move beyond advice on decision-making, taking a load of low-level and even high-level decisions out of human hands.”¹⁴ Differently, autonomic profiling describes the process whereby the human role is minimized and the decision making process is entirely driven by the machine.¹⁵ Autonomic profiling “goes one step further than automated profiling.”¹⁶ The machines drive the decision making process, providing for a readjusted environment based on their profiling and without calling for human intervention. Besides their degree of automation profiling methods can be distinguished by their object and application. Profiling can be applied as group profiling or individual profiling: the techniques that identify and represent groups can also focus on individuals.¹⁷ Moreover profiling relies on data collected from one single person or group to apply the information derived from data processing to the same person or group – direct profiling – or it relies on categorization and generalisation from data collected among a large population to apply it to certain persons or groups – indirect profiling. Group profiling can also be classified as distributive group profiling or non-distributive group profiling.¹⁸ A distributive group profile identifies a certain number of people having the same attributes. All the members of the group share the same characteristics. In contrast, a non-distributive group profile identifies a certain number of people who do not share all the attributes of the group’s profile.

These distinctions give an idea of the different types of profiling and their application. The forms of profiling, which are subject of this article are automated and autonomic profiling and their various forms and fields of application.

The following proposed definition takes into account the preceding evolution of technologies in which profiling is embedded and focuses on the purpose profiling is being used for. It will be the basis for this paper:

Profiling is a technique of (partly) automated processing of personal and/or non-personal data, aimed at producing knowledge by inferring correlations from data in the form of profiles that can subsequently be applied as a basis for decision-making.

A profile is a set of correlated data that represents a (individual or collective) subject.

Constructing profiles is the process of discovering unknown patterns between data in large data sets that can be used to create profiles.

Applying profiles is the process of identifying and representing a specific individual or group as fitting a profile and of taking some form of decision based on this identification and representation.

Advanced data analysis tools have established new social practices of knowledge production and have created new types of knowledge. We argue that the practices of profiling have facilitated and are part of a broader societal paradigm of prevention. We will elaborate on the societal implications of changing social practices through emerging profiling technologies as a ground for the examination of threats for fundamental rights and values of European societies in Sect. 1.4.

Observations made by human beings need to be written down to be made explicit. The written documentation of observations can be regarded as a first step to enable a generalized and objectified way of keeping information and exchanging it between individuals and institutions.¹⁹ Digitized information, however, can be processed and analysed automatically so that information is easier and cheaper to store, process and analyse. An illustrative example of how exhaustive and expansive the detailed documentation of people’s activities and behaviour has been, is the comparison between digital data the NSA stores with the amounts of files the Stasi – German Democratic Republic’s domestic secret service – produced. All the information captured throughout the Stasi history would fill about 48.000 cabinets covering approximately 0,019 km². The NSA’s planned data centre in Utah will host about 5 zettabytes of data which could roughly be converted in about 42 quadrillion file cabinets covering 17 million km² – bigger than the European continent.²⁰ The example also shows the differing efforts needed to collect and archive data depending on whether using analog or digital data processing. While the Stasi needed to install microphones, hire staff to monitor and document people’s behaviour to gain information about their habits, attitudes and social networks, in a digitized world a lot of that information can be monitored and stored on the fly through sensors, log data or user generated content. This shows that the digitization of communication and transactions does not only produce more data but also provides new kinds of information²¹ which can be used to extract knowledge about individuals: their social relations, interests and activities. Once stored and made accessible via computer networks, data becomes easily exchangeable worldwide. At the same time it becomes hard to grasp how data is exchanged, which information is gained and by whom. Furthermore the specific mediums can store specific data. Certain elements which can be archived on paper cannot be archived digitally and vice versa. Moreover certain information can hardly be digitized respectively digitally analyzed, e.g. hand-written information, and smells. By that, archives have a filtering function which shapes the accessibility of information as data. But simplified storage and exchange of data are only one aspect of the ongoing process of digitization of everyday life. Beyond that advanced methods of data analysis have fundamentally changed the procedures of knowledge production through automation.

Another effect of the digitization of data becomes evident when we think of the different haptic and cognitive perceptions of digital versus analog files and folders. Different items and elements can be put in an analog or digital file, and at the same time, the availability of and the access to certain kinds of information fundamentally changes. In other words: accessing information at a (real) desktop is very different from accessing information when sitting in front of a computer screen. Paper folders can be touched and felt, digital files are browsed on a screen and can be searched by keywords. Consequently, the way of reasoning changes, as first findings of one of the case studies conducted in PROFILING show.²² More interaction of the analyst is oriented towards computer interfaces and thus influenced by the way user interfaces are designed, information is presented, and how searches can be conducted.²³ The transformation of the human role in knowledge production processes is even more significant when it comes to examining large-scale databases. Learning algorithms are trained on specific data sets to build categories or to find patterns in the data. Assumptions or hypotheses made by the analyst play a minor role during data processing, they are to a certain degree hidden in the process of writing algorithms and training the algorithms. Finally, hypotheses are derived “from the material”.²⁴ As a consequence implicit assumptions driving the actors during the selection of training data, preprocessing target data and suitable algorithms become invisible and the outcomes produced by “the data” seem objectified. Subjective assumptions and social norms are hidden in the technology during the process of automatization, while outcomes based on computed models and databases are often perceived as solid statistics and thus more objective than human interpretation.²⁵ This perception as objectified knowledge of computer-generated models supports the thesis of a general tendency of technology to make social norms more durable²⁶ and more specifically the thesis that social sorting becomes strengthened if mediated through technology.²⁷ Profiles, as mentioned above, can be seen as hypotheses. These hypotheses are inductive as they are not necessarily developed on the basis of a theory or a common sense expectation, but often emerge in the process of data mining. This can be regarded as a shift from a more traditional, rather assumption-driven approach to a discovery-driven approach to knowledge generation.²⁸ This shift results not only from growing data capabilities and advancing technological methods. Lyon argues that the conceptualization of social threats as actuarially identifiable and addressable risks and the desire for intelligence-led management of populations play a key role in the spread of profiling technologies.²⁹ In this context data mining is considered a key technology for risk assessment in various fields of application such as eHealth, airport security, and policing. Profiling techniques are used to identify categories and groups in order to assess risks and probabilities of certain future developments. The generated profiles can then be used to sort individuals, groups, events or processes in order to make them addressable for specific practices.³⁰ In this regard profiling is a technology to structure potential futures in order to make them governable in the presence. Therefore profiling is an important practice of a broader societal preventive paradigm, which is based on probabilistic knowledge used to manage social processes in the form of risk management.³¹ By that profiling technologies provide means of control, which can be exercised for care and protection or coercion and repression.³²

Even though the results of data mining are often limited reliable,³³ proponents claim that the potentials for managing social and technological processes in more efficient ways through data gathering and analysis are immense. They expect that the growing amount of data and increasingly advanced tools for examination will provide information which will allow organisations to identify, target, and act upon undesirable developments at an early stage – preferably before they occur. Preemptive policing, early detection of pandemic risks, and the prevention of tax fraud are examples of the societal benefits of the use of sophisticated data mining methods. Yet there is a downside to these opportunities implied by the technological evolution of digitization: it threatens key aspects of fundamental citizen rights, such as the rights to privacy, data protection and non-discrimination, and core values of European societies – democracy, the rule of law, autonomy and self-determination. As societies rely more and more on profiling methods to steer social and technological processes the urgency of dealing with these threats grows.

In the current EU data protection legislation the word profiling does not appear. However. Article 15 of the Directive 95/46/EC (hereinafter, Data Protection Directive, DPD) concerns ‘automated individual decisions’ and thus is closely related to profiling. According to article 15(1): “every person has the right not to be subject to a decision which produces legal effects concerning him or significantly affects him and which is based solely on automated processing of data intended to evaluate certain personal aspects relating to him, such as his performance at work, creditworthiness, reliability, conduct, etc.” At the same time, article 15(2) states an exception: “a person may nevertheless be subjected to an automated individual decision if that decision is taken: (a) in the course of the entering into or performance of a contract, provided the request for the entering into or the performance of the contract, lodged by the data subject, has been satisfied or that there are suitable measures to safeguard his legitimate interests, such as arrangements allowing him to put his point of view; or (b) is authorized by a law which also lays down measures to safeguard the data subject’s legitimate interests”.

In the light of Article 15 of the DPD, it is relevant whether the processing is meant to evaluate a certain aspect of the person’s behavior, character or identity on which a decision can be based. A decision based on a profile can comply with the law, but a natural person has to be involved in the process. To sum up, Article 15 does not take the form of a direct prohibition on a particular type of decision-making; rather, it directs each EU Member State to confer on persons a right to prevent them from being subjected to purely automated decisions in general.⁶⁸