Defining Profiling: A New Type of Knowledge?
In this first chapter a set of relevant distinctions will be made to explore old and new ways
of profiling, making a first attempt to define the type of profiling that is the subject of this
publication. The text explains how profiling or pattern recognition allows us to discrimi-
nate noise from information on the basis of the knowledge that is constructed, providing a
sophisticated way of coping with the increasing abundance of data. The major distinctions
discussed are between individual and group profiles (often combined in personalised pro-
filing), between distributive and non-distributive group profiles and between construction
and application of profiles. Having described automated profiling we will compare such
machine profiling to organic and human profiling, which have been crucial competences
for the survival of both human and non-human organisms. The most salient difference
between organic and machine profiling may be the fact that as a citizen, consumer or
employee we find ourselves in the position of being profiled, without access to the knowl-
edge that is used to categorise and deal with us. This seems to impair our personal freedom,
because we cannot adequately anticipate the actions of those that know about us what we
may not know about ourselves.
Profiling occurs in a diversity of contexts: from criminal investigation to marketing
research, from mathematics to computer engineering, from healthcare applications
for elderly people to genetic screening and preventive medicine, from forensic bio-
metrics to immigration policy with regard to iris-scans, from supply chain manage-
ment with the help of RFID-technologies to actuarial justice. Looking into these
different domains it soon becomes clear that the term profiling is used here to refer
to a set of technologies, which share at least one common characteristic: the use of
algorithms or other techniques to create, discover or construct knowledge from
huge sets of data. Automated profiling involves different technologies (hardware),
such as RFID-tags, biometrics, sensors and computers as well as techniques (soft-
ware), such as data cleansing, data aggregation and data mining. The technologies
and techniques are integrated into profiling practices that allow both the construction
Vrije Universiteit Brussel, Erasmus University Rotterdam
M. Hildebrandt and S. Gutwirth (eds.), Profiling the European Citizen: 17
© Springer Science + Business Media B.V. 2008
18 M. Hildebrandt
and the application of profiles. These profiles are used to make decisions, sometimes
even without human intervention. The vision of Ambient Intelligence or ubiquitous
networked environments depends entirely on autonomic profiling, the type of
profiling that allows machines to communicate with other machines and to take
decisions without human intervention.
In this chapter we will start with the identification of profiling as such, providing
working definitions of profiling and some related terms. After that we will discuss
the difference between group profiling and personalised profiling and the way they
are mixed up in practice. On the basis of this initial exploration of automated profil-
ing, such technological (machine) profiling will be compared with non-technological
forms of profiling, in particular organic and human profiling. This should enhance
our understanding of the difference between machine and human profiling, which is
crucial for an adequate assessment of the opportunities and risks involved.
2.2 Identification of Profiling
In this volume the focus will be on automated profiling, which is the result of a
process of data mining. Data mining – which will be discussed in detail in chapters
2 and 3 - is a procedure by which large databases are mined by means of algorithms
for patterns of correlations between data. These correlations indicate a relation
between data, without establishing causes or reasons.
What they provide is a kind
of prediction, based on past behaviour (of humans or nonhumans). In that sense
profiling is an inductive way to generate knowledge; the correlations stand for a
probability that things will turn out the same in the future. What they do not reveal
is why this should be the case. In fact, profilers are not very interested in causes or
reasons, their interest lies in a reliable prediction, to allow adequate decision mak-
ing. For this reason profiling can best be understood from a pragmatic perspective:
it aims for knowledge that is defined by its effects, not for conceptual elaboration.
Another way to articulate the particular kind of knowledge produced by profiling is
to see profiles as hypotheses. Interestingly, these hypotheses are not necessarily
developed within the framework of a theory or on the basis of a common sense
expectation. Instead, the hypothesis often emerges in the process of data mining, a
change in perspective that is sometimes referred to as a discovery-driven approach,
Correlations can of course be spurious (see http://www.burns.com/wcbspurcorl.htm), however,
this does not mean that non-spurious correlations necessarily establish causal or motivational
relationships between data.
According to the founding father of American pragmatism, Charles Saunders Peirce, the Maxim
of Pragmatism reads as follows: ‘Consider what effects that might conceivably have practical
bearings we conceive the object of our conception to have: then, our conception of those effects
is the whole of our conception of the object’ (Peirce, 1997:111). A pragmatic approach of knowl-
edge should not be conflated with a populist or naïvely ‘practical’ attitudes to knowledge.
2 Defining Profiling: A New Type of Knowledge? 19
as opposed to the more traditional assumption-driven approach.
‘Data mining pro-
vides its users with answers to questions they did not know to ask’ (Zarsky, 2002-
2003:8). After correlations (hypotheses) have surfaced they are tested when the
profiles are applied. This is why the construction and application of profiles are
entangled in profiling practices (complementing the inductive process of generating
profiles with the deductive process of testing them on new data).
Before supplying a working definition of profiling we need to define three
terms, which are central in the context of profiling: data subject, subject and data
controller. The central position in profiling is taken by what is called the data
subject, which we define as the subject (human or non-human, individual or
group) that a profile refers to. In the case of group profiling this means that the
data subject may be the result of profiling, not necessarily pre-existing as a group
that thinks of itself as a group. For instance, a category of blue-eyed women may
emerge as a data subject, because as a category they correlate with a specific
probability to suffer from breast cancer. This implies that we use the term data
subject in a different way than is usual in data protection legislation, as in the case
when the data subject is defined as ‘an identified or identifiable natural person’.
We define a subject as the human or nonhuman individual of which data are
recorded which are used to generate profiles and/or as the human or nonhuman
individual to which a profile is applied.
The next – equally central – position is
taken by the data controller (sometimes called data user), which we define as the
subject (person or organisation) that determines the purposes of the processing of
the data and the use that will be made of them (including the sale of data or of
the profiles inferred from them).
A simple working definition of profiling could be:
The process of ‘discovering’ correlations between data in databases that can be used to
identify and represent a human or nonhuman subject (individual or group) and/or the appli-
cation of profiles (sets of correlated data) to individuate and represent a subject or to iden-
tify a subject as a member of a group or category.
To understand the meaning of profiling, it may be helpful to add the purpose of
profiling. Besides individuation, profiling mainly aims for risk-assessment and/or
assessment of opportunities of individual subjects. This, however, cannot be taken
for granted. If the interests of the data controller and subject differ it may well be
that the interests of the data controller, who pays for the whole process, will take
precedence. Thus – in the end – what counts are the risks and opportunities for the
data controller. For this reason the purpose of profiling can best be formulated as:
Custers (2004: 46), referring to B. Cogan, Data Mining; Dig Deep for the Power of Knowledge.
Internet publication at www.BioinformaticsWorld.info/feature 3a.html.
Par. 2 (a) Directive 95/46 European Community (D 95/46/EC).
This means that in this context the term data subject is used in a different way compared to the
way it is used in Data Protection legislation. What is called a subject in this text, is called a data
subject in D 95/46/EC, meaning the subject whose data have been processed.
20 M. Hildebrandt
The assessment of risks and/or opportunities for the data controller (in relation to risks and
opportunities concerning the individual subject).
This raises the question whether it is possible to empower a human subject to make
her a data controller in her own right, with regard to profiles that can be inferred
from her data and profiles that may be applied to her.
2.3 Group Profiling & Personalised Profiling
2.3.1 Groups: Communities and Categories
Profiling techniques generate correlations between data. For instance, a correlation may
be found between people that are left-handed and have blue eyes and a specific disease
or a correlation may be found between people that live in a certain neighbourhood and
have a particular level of income or a correlation may be found between one’s individual
keystroke behaviour and regular visits to a specific type of pornographic website. To
generate such correlations in a reliable way we need to collect, aggregate and store the
relevant data over an extended period of time, perhaps by integrating different databases
that contain such data. In the examples given, the correlations concern data of certain
categories of subjects, for instance the category of people with blue eyes that are left-
handed or the category of people that live in a certain neighbourhood. Once the process
of data mining establishes the correlations, two interrelated things happen: (1) a certain
category is constituted (2) as having certain attributes. The category is usually called a
group and the set of attributes are called the group’s profile.
Another possibility is that the data of an existing group of people, who form
some kind of community, are collected, aggregated, stored and processed in order
to find shared features. For instance, the members of a local church or the students
living in a certain dormitory can be the target of profiling. In this case the process
of data mining will not establish them as a group (which they already were) but it
may generate correlations and certain attributes between them, such as a typical
way of dressing, particular eating habits or specific travel habits.
Group profiling can concern both communities (existing groups) and categories
(e.g., all people with blue eyes). In the case of categories, the members of the group
did not necessarily form a community when the process was initiated; in the case
of communities the members of the group already formed a community (however
unstructured). The fact that profiling may establish categories as sharing certain
attributes may in fact lead to community building, if the members of such a cate-
gory become aware of the profile they share. The fact that data controllers may tar-
get the members of a category in a certain way – without them being aware of this
– may of course impact their behaviour as members of this category.
Cp. Zarsky (2002-2003:9-15) on clustering and association rules.
2 Defining Profiling: A New Type of Knowledge? 21
2.3.2 Distributive and Non-distributive Profiles
To understand some of the implications of group profiling we have to discriminate
between distributive and non-distributive profiles. A distributive profile identifies a
group of which all members share all the attributes of the group’s profile. This means
that the group profile can be applied without any problem to a member of the group –
in that sense it is also a personal profile. An example of a distributive profile is the cat-
egory of bachelors that all share the attribute of not being married. A less tautological
example is the category of oak trees that all develop a certain type of leaf. Being a
member of a group with a distributive profile has potentially pervasive social and legal
implications because the profile will apply without qualification to all members.
It should be obvious that apart from groups that are defined in terms of a shared
attribute (e.g., the group of bachelors that share the attribute of not being married),
most groups do not have distributive profiles. A non-distributive profile identifies a
group of which not all members share all the attributes of the group’s profile.
instance, Hare’s checklist for psychopaths is a non-distributive profile. It contains 20
items e.g., absence of guilt, superficial charm, pathological lying and poor aggres-
sion control, that have to be checked and scored on a 3-point scale (0: does not
apply; 1: applies to some extent; 2: applies). A person whose profile counts 30 points
or more is considered a psychopath, a profile that is said to be – statistically – predi-
ctive of violent criminal recidivism of released offenders. The category of persons
who score 30 points or more on Hare’s checklist has a non-distributive profile,
because not every person in this group shares the same attributes. From a social and
legal perspective this is very instructive, because it implies that one cannot apply the
profile to members of the group without qualification (Edens, 2001).
It is important to realise that treating members of a group that has a non-distributive
profiles as fitting the entire profile may have interesting effects. For instance, if people
fit the profile of a high-income market segment, service providers may decide to offer
them certain goods or provide access to certain services, which may reinforce their fit
in the category. If the group profile is non-distributive and they in fact do not share the
relevant attributes (e.g., they may live in a certain neighbourhood that is profiled as
high-income, while in fact they have a very low income, being an au pair), they may
actually be ‘normalised’ into the behaviour profiled as characteristic for this group.
2.3.3 Actuarial Approach in Marketing, Insurance and Justice
The use of non-distributive profiles (the usual case) implies that profiles are always
probabilistic. They basically describe the chance that a certain correlation will occur
In terms of Wittgenstein, the members of the group have a family resemblance, they cannot be
identified by means of a common denominator. Cp. Custers, 2004; Vedder, 1999.
Cp. Vedder, 1999.
22 M. Hildebrandt
in the future, on the basis of its occurrence in the past. As indicated above, the cor-
relation does not imply a causal or motivational relationship between the correlated
data, they merely indicate the fact that the occurrence of one will probably coincide
with the occurrence of the other. For instance, in genetic profiling, we may find that
the presence of a certain gene correlates with a certain disease. Depending on the
exact correlation (the percentage of cases in which it occurs) we may predict – in
terms of probability – the chance that a person with the relevant gene will develop
the relevant disease. In reality, of course, the correlations may be very complex, e.
g., depending on a whole set of different factors in a non-linear way. The exponential
increase in computer power, however, enables the storage of a nearly unlimited
amount of data and allows computer scientists to develop very complex algorithms
to mine these data.
This has led to relatively new developments in marketing, insur-
ance and justice, based on the targeted assessment of consumer preferences (leading
to spam), targeted risk-assessment (concerning financial credibility) and on criminal
profiling (leading to actuarial justice). In all three fields it becomes possible to take
decisions (in customer relationship management, on refined types of price-discrimi-
nations, on categorised or even personalised interest-rates and insurance premiums,
on targets for criminal investigation and on sentencing modalities) that are based on
highly informed predictions of future behaviour. This approach to customers and
citizens can be termed an actuarial approach, because it builds on highly sophisti-
cated assessments of the risks and opportunities involved. The caveat of this
approach is that it extrapolates from the past to the future on the basis of blind cor-
relations, tending to see the future as determined by established probabilities, possi-
bly disabling potentially better solutions that lie in the realm of low probabilities.
2.3.4 Personalisation and Ambient Intelligence
Mining the data from a variety of people allows categorising them into different
types of groups, generating high rates of predictability concerning the behaviour of
categories of people. Apart from group profiling, however, a second type of profil-
ing has evolved, that mines the data of one individuated subject.
metrics is a good example of such profiling (discussed in detail in chapter 5).
Data mining by means of algorithms or heuristics (see chapter 4) works with a set of instructions
that has to be followed chronologically, this is called conventional computing. The end result of
such a process is entirely predictable, even if our brains do not have the computing powers to
apply the algorithm in as little time as the computer does. According to Stergiou and Siganos
(1996), data mining by means of neural networks works with ‘highly interconnected processing
elements (neurones) working in parallel to solve a specific problem’. They claim that one of the
advantages of this emerging technology is that problems that we do not understand can still be
solved, while for the same reason the resolution of the problem is not predictable.
In terms of the reply of Jaquet-Chiffelle this would be ‘direct individual profiling’.
2 Defining Profiling: A New Type of Knowledge? 23
For instance, profiling the keystroke behaviour of one particular person may enable
a service provider to ‘recognise’ this person as she goes online because of her
behavioural biometric ‘signature’ and allows the service provider to check her
online behaviour (discussed in detail in chapter 9) and thus to build up a very per-
sonal profile that can be used to offer specific goods and provide access to certain
services. The profile can also be stored and sold to other interested parties, or be
requested by the criminal justice or immigration authorities.
Such personalised profiling is the conditio sine qua non of Ambient Intelligence
(AmI), the vision of a networked environment that monitors its users and adapts its
services in real time, permanently learning to anticipate the user’s preferences in
order to adapt to them.
Ambient Intelligence presumes an RFID-tagged environ-
ment, and/or an environment enhanced with sensors and/or biometric appliances,
all connected with online databases and software that allows a continuous process
of real-time profiling. The intelligence is not situated in one device but emerges in
their interconnections. The online world with its seemingly limitless capability to
collect, aggregate, store and mine behavioural data thus integrates the offline world,
creating a new blend of virtual and physical reality (ITU, 2005).
ments may know your preferences long before you become aware of them and
adapt themselves in order to meet those preferences. The AmI vision promises a
paradise of user-centric environments, providing a continuous flow of customised
services by means of ubiquitous and pervasive computing.
However, one does not
need an overdose of imagination to foresee that such highly personalised profiling
engenders unprecedented risks for users to be manipulated into certain preferences,
especially in the case that users have no feed-back on what happens to the data they
‘leak’ while moving around in their animated environments.
2.4 Automated and Non-automated Profiling
2.4.1 Categorisation, Stereotyping and Profiling
Long before computers made their way into everyday life, criminal investigators
composed profiles of unknown suspects, psychologists compiled profiles of people
with specific personality disorders,
marketing managers made profiles of different
As elaborated in chapter 6, personalised profiling can be a combination of individual and group
profiling, or in terms of the reply of Jaquet-Chiffelle ‘direct and indirect individual profiling’.
See Mark Weiser’s pioneering work on ubiquitous computing, for example, Weiser (1991:
94 – 104).
The AmI vision has been propagated mainly by Philips and the European Commission, see for
example, Aarts and Marzano 2003; ISTAG (Information Society Technology Advisory Group), 2001.
On psychometrics (psychological testing), see for example, Rasch, 1980; Thorndike, 1971.
24 M. Hildebrandt
types of potential customers and managers profiled the potentials of their employ-
ees for specific jobs.
Adequate profiling seems to have been a crucial competence
of professional occupation and business enterprise since their inception, perhaps
most visible today in marketing and criminal investigation.
However, profiling is
not just a professional, business or government preoccupation. As Schauer (2003)
convincingly demonstrates in his Profiles, Probabilities and Stereotypes, profiling
is a form of generalisation or categorisation we all apply routinely to get us through
life. Habermas would probably speak of Kontingenzbewältigung, being the reduc-
tion of complexity in an environment that demands continuous choices of action,
which would swamp us if we were to reflect on each of them. Schauer professes
that categorisation is mainly a good thing, especially if it is based on a ‘sound sta-
tistical basis’ and his position has a strong appeal to our common sense. How could
we move on in life if we did not take certain generalisations for granted, if we did
not live by certain rules that are based on such generalisation – even if they do not
always apply? Schauer warns against attempts to look at each and every case in
isolation, attending to the particular instead of the general, glorifying what lawyers
in Germany once called ‘Einzelfallgerechtigkeit’. In his opinion routine assess-
ments on the basis of generalisation are not only necessary to cope with complexity
and multiplicity but they also provide just instead of arbitrary decisions, because
of the appeal to a general standard, which creates a type of predictability (essential
for e.g., legal certainty). In psychology the need to reduce the weight of recurring
decisions is thought to be the cause of ‘stereotyping’, a healthy way to deal with the
growing complexities of life. It means that we – unconsciously - group different
events, things or persons into categories in order to assess what can be expected and
to be able to decide how to act. Stereotyping allows anticipation. Following this line
of thinking, categorisation and stereotyping are a kind of everyday profiling, based
on experience and practical wisdom and if we believe Schauer, it also produces a
kind of justice. In the next section I will take this line of thought one step further in
claiming that profiling is not only a part of professional and everyday life but also
a constitutive competence of life itself in the biological sense of the word.
However, before describing the process of profiling from the perspective of the
life sciences, we need to make some comments on Schauer’s defence of categorisa-
tion and its relationship to profiling. In the introduction to his book he discusses
‘Generalization Good and Bad’. He starts by drawing a distinction between generalisa-
tions with and without a statistical or factual basis; those without he calls spurious. He
qualifies this distinction by indicating that in everyday life we may pronounce many
generalisations without intending them to be taken as absolute. For instance, when
See for example, Rafter and Smyth, 2001.
For a history and overview of criminal profiling, see for example Turvey, 1999. For a history
and overview of data mining in academic marketing research see for example Wilkie, William and
Moore (2003: 116-146). For more practical research see for example, Peppers and Rogers, 1993
(about the integration of data mining and CRM (Customer Relationship Management) to achieve
2 Defining Profiling: A New Type of Knowledge? 25
we say that ‘Bulldogs have bad hips’, this – according to Schauer – may be a good
generalisation, even though a majority of bulldogs do not have bad hips. ‘As long as
the probability of a dog’s having hip problems given that the dog is a bulldog is
greater than the probability of a dog’s having hip problems given no information
about the breed of dog, we can say that the trait of being a bulldog is relevant,
and we can say that generalizing from that trait meets the threshold of statistical
(or actuarial) soundness’ (Schauer 2003: 11).
Thus we have what he calls universal
generalisations, which denote a group of which all members share the generalised
characteristic and non-universal generalisations, which denote a group of which a
majority or a relevant minority share the generalised characteristic. Schauer then
moves on to discuss prejudice or stereotype as a kind of generalisation, recognising
that these terms are often used in a pejorative way. He seems to conclude that the use
of a non-universal generalisation must not be rejected, while admitting that - depend-
ing on the context (sic!) – sometimes such prejudice or stereotyping can indeed be
morally flawed. One example he gives is the case of racial profiling, though he
seems to suggest that in this case the generalisation is not based on sound statistical
or empirical evidence. The reason for the fact that acting on a nonspurious non-uni-
versal generalisation may – under certain circumstances - be morally wrong is that
‘equality becomes important precisely because it treats unlike cases alike’ (Schauer,
2003: 296). So, even if most ex-convicts or a relevant minority of them, are prone to
commit crimes again, we may decide we want to treat them equally when they apply
for a job, insurance or try to rent a house - equally to non-ex-convicts. A principle
such as the presumption of innocence has the same function: even if we are quite
sure that a person has committed a certain crime, government officials cannot treat
this person as an offender until guilt has been proven according to law. I am not sure
these are the examples Schauer would endorse to demonstrate the importance of the
moral evaluation that may interfere with justified generalisation but he has explained
in a clear voice how generalisation, equality and even community relate to each
other. In the remainder of this chapter we can use some of the salient distinctions he
makes to clarify the complexities of automated profiling and the implications it may
have for fairness and equality.
2.4.2 Organic Profiling: A Critical Sign of Life
After concluding that profiling is part of professional as well as everyday life of human
beings, I would like to make a brief excursion into the life sciences to highlight the
importance of profiling for living organisms. As Van Brakel (1999) writes, biology and
information theory have developed into an integrated domain, part of the life sciences.
This is an important development, which may help us to understand the way auto-
mated (machine) profiling can generate knowledge, although not human knowledge.
Emphasis of Schauer.
26 M. Hildebrandt
Both ‘organic profiling’ and automated machine profiling concerns the production of
implicit knowledge, or at least knowledge that has not reached a human conscious.
In 1987 Maturana and Varela published a little book, The Tree of Knowledge,
explaining The Biological Roots of Human Understanding.
For our purposes the
theory of knowledge argued in their book is interesting because it explains knowl-
edge as something that an observer attributes to an organism that effectively deals
with its environment. For Maturana and Varela knowledge is constituted by the
interactions between – for instance - a fly and its immediate environment, if this
interaction is successful in the sense that it sustains the life of the fly. Their under-
standing of knowledge is enactive (knowledge and action ‘cause’ each other): only
by acting, an organism finds out about its environment and in that sense even per-
ception is a form of – entirely implicit - action.
To be more precise one could say
that all living organisms, in order to survive, must continuously profile their envi-
ronment to be able to adapt themselves and/or to adapt the environment. Profiling
in this case means the process of extracting relevant information from the environ-
ment. However, what counts as information depends on the knowledge the organ-
ism has built on the basis of continuous interaction with its environment, because
this knowledge determines what type of information is relevant and valid. This
means that what counts as information at one point in time may be noise at another
point in time and what counts as noise for one individual (organism) may be infor-
mation for another. It also means that knowledge depends on both the environment
and the organism and must be understood as fundamentally dynamic and context-
dependent. Knowledge in this sense is always local knowledge. This does not mean
that generalisation is out of bounds, quite on the contrary. To be able to act in an
environment adequate generalisation is necessary but the question of which gener-
alisation is adequate will depend on the context (and on the organism).
What is crucial at this point is that (1) profiling the environment happens without
involving a conscious mind (2) profiling provides feed-back necessary to survive
(3) profiling extracts information, depending on knowledge that allows one to dis-
criminate between noise and information (4) profiling transforms information into
knowledge and (5) information and knowledge always depend on both the organ-
ism and its environment, there is no view from anywhere.
2.4.3 Human Profiling: The Meaning of Autonomous Action
The small excursion into profiling by nonhuman organisms allows us to develop
a keener eye for what makes knowledge human knowledge. If perception, information
Revised edition of 1998. Matuna and Varela coined the term autopeiosis in 1973 to describe the
process that constitutes living organisms. In The Tree of Knowledge they expound on their theory
of biology by investigation the relationship between living organisms and their environment.
Their theory of knowledge thus combines pragmatism and embodied phenomenology, rejecting
both mentalism or naïve empiricism.
2 Defining Profiling: A New Type of Knowledge? 27
gathering, feed-back and even knowledge are not specific for the human animal,
what is? Could it be that consciousness is the discriminating attribute, and if so,
what difference does this make for profiling? Compared to a plant, a dog seems
to have a different kind of awareness of the world. We may be inclined to call
this awareness a consciousness. This is not the case because the dog is aware of
being aware but because it seems to embody a unified self that is absent in a
plant. The philosopher Helmuth Plessner (1975) described the difference by
pointing out that all mammals have a central nervous system that seems to allow
for a centralisation of the awareness, giving rise to a conscious presence in the
world. The difference between humans and other mammals, according to
Plessner, is the fact that a human is also conscious of being conscious, conscious
of herself. This reflective attribute, which is often thought to derive from the fact
that we use language to communicate with each other, is absent in other mammals
or present to a different degree.
To assess why this difference is relevant for our study of profiling we need to
connect our capacity for reflection with our capacity for intentional action
(which we suppose to be less evident in other mammals).
that we can look back upon ourselves, which also implies that we can consider
our actions as our actions, as it were, from a distance. Such reflection can be
incorporated into our actions – even before we act. We may thus consciously
reflect upon different courses of action and intentionally prefer one alternative
to another. This is what allows intentional action and this seems to be the pre-
condition for autonomous action: an action we have freely decided upon, an
action within our own control. Auto is Greek for self, nomos is Greek for law so
human autonomy implies intentional action and conscious reflection, two condi-
tions for positive freedom.
Before moving on to the relevance of intentional action and conscious reflection
for profiling, we need to keep in mind an important fact. Most of our actions are nei-
ther intentional nor conscious. We can move around freely in this world because we
have acquired habits that are inscribed in our bodies, allowing us to act in a number
of ways without giving it thought. However, the very small amount of actions we
actually consciously intend, are distinctive for our moral competence – taking into
account that conscious reflection is the incentive to create new habits which will
again move out from the zone of intentional action, but did originate from it.
2.4.4 Machine Profiling: The Meaning
of Autonomic Machine Behaviour
In 2001, Paul Horn, IBM’s Senior Vice President, introduced the idea of autonomic
computing. Interestingly, he chose a term that refers to biology, to the autonomic
We cannot be too presumptious here, see for example de Waal, 2001.
28 M. Hildebrandt
nervous system, because it ‘governs our heart rate and body temperature, thus
freeing our conscious brain from the burden of dealing with these and many other
low-level, yet vital, functions’ (Kephart and Chess, 2003: 41). One of the
objectives of autonomic computing is to prevent or resolve the advancing software
complexity crisis, by creating a network that is capable of self-management: self-
configuring; self-healing; self-optimising; self-protecting (CHOP). Visions of
Ambient Intelligence (AmI), pervasive computing or the RFID Internet of Things
depend on extended interconnectivity and we are being warned that without
self-management the design of the integrated network architectures will become
entirely impossible. Another objective is to allow the user of the system to collect
the fruits of ubiquitous computing without being bothered with the flow of minor
and major adjustments that need to be made to keep the system operational.
Kephart and Chess (2003: 42) distinguish different stages in the development of
autonomic systems, starting with automated functions that collect and aggregate
data and ending with automation technologies that can move beyond advice on
decision-making, taking a large amount of low-level and even high-level decisions
out of human hands.
To target the difference between organic, human and machine profiling it is
interesting to discuss automated profiling in terms of autonomic machine behav-
iour. With autonomic machine behaviour I mean the behaviour of machines that are
part of a network of machines that exchange data and make decisions after process-
ing the data. This need not incorporate the entire concept of autonomic computing
with its CHOP attributes, but is based on what is called M2M talk (machine to
machine communication) (Lawton, 2004:12-15). ‘Machine’ can be anything, such
as a RFID-tag (radio frequency identification tag), a PDA (personal digital assist-
ant) or a PC (personal computer). I call the behaviour autonomic in as far as the
network of machines processes data, constructs knowledge and makes decisions
without the intervention of a human consciousness. This autonomic machine
behaviour will be part and parcel of ambient intelligent environments, which moni-
tor subjects and adapt the environment in real time, necessitating autonomic
machine decision making.
The most simple form of automated profiling is when profiles are generated and
applied in the process of data mining, after which human experts sit down to filter
the results before making decisions. In this case we have no autonomic machine
behaviour, because decisions are taken by human intervention. It may, however, be
the case that these decisions routinely follow the machine’s ‘advice’, bringing the
whole process very close to autonomic machine profiling.
Art. 15 of the Directive on Data Protection 95/46/EC attributes a right to ‘every person 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.’. The fact
that usually some form of routine human intervention is involved means that art. 15 is not applicable,
even if such routine decisions may have the same result as entirely automated decision making.
2 Defining Profiling: A New Type of Knowledge? 29
2.4.5 Organic, Human and Machine Profiling:
Autonomic and Autonomous Profiling
After discussing organic, human and machine profiling we can now draw some pru-
dent conclusions. It seems that most organic profiling does not involve conscious
reflection or intentional action. It is important to note that an important part of
human existence itself is sustained by the autonomic nervous system, which continu-
ously profiles its environment inside and outside the body, by means of operations
that we are not aware of. On top of that human profiling is done ‘automatically’ to
a very large extent. This automation or habit-formation is the result of a learning
process that often starts with conscious attention shifting to implicit behaviour as
soon as the habit is inscribed in our way of doing things. The competence to act on
this basis is referred to as implicit or tacit knowledge (Polanyi, 1966). Machine pro-
filing seems similar to organic profiling, in the sense that it does not involve con-
scious reflection, nor intentional action. However, organic profiling presumes an
organic system that constitutes and sustains itself. Maturana and Varela (1991) have
coined the term autopoiesis for this self-constitution.
Even if autonomic computing
– as defined by IBM – can be successfully compared to the autonomous nervous
system, we may have a problem in defining it as self-constituting as long as it needs
an initial software architecture provided by human intervention.
In other words, machine profiling is like organic profiling to the extent that it is
part of autonomic behaviour and like human profiling to the extent that human profiling
is done implicitly. At the same time, machine profiling differs from human profiling in
two salient ways: (1) other than human and organic profiling machine profiling is not
part of an autopeiotic system that constitutes itself, (2) other than human profiling
machine profiling does not integrate conscious reflection or intentional action.
2.5 Conclusions: From Noise to Information,
From Information to Knowledge
As they say, we live in an information society and in a knowledge society. One of
the challenges of the present age is how to deal with the overload of information,
or rather, how to discriminate noise from information. Another challenge is how to
The term has been introduced into sociology by Luhmann and Teubner, who also build on Heinz
von Foerster (cybernetics), implying that not only individual cells or metacellular organisms form
autopoietic systems, but also social systems. However, system theory and other sociological
schools that claim that individuals are determined by the social system or the underlying structure
do not seem to build on Maturana and Varela, who explicitly claim that social systems amplify
the individual creativity of its components, arguing that the social system actually exists for these
components (and not the other way round, as is the case in metacellular organisms, cf. Maturana
and Varela (1998:199).
(re)construct knowledge out of the flows of noise and information, how to deal with
the growing complexities of our scientific knowledge constructs and with the
emerging unpredictability of the complex technological infrastructures built to face
the increasing mobility of human and nonhuman imbroglios.
One of the answers to both questions is the use and further development of pro-
filing technologies. They may incorporate the only way to reduce the overload of
information, to make it ‘manageable’, to make sense out of it and to regain control
of the effects of one’s actions. In other words, they may provide the only way to
adequately anticipate the consequences of alternative courses of action. If freedom
presumes anything, it is precisely this: a reliable anticipation of the results of the
choices we have. This is why legal certainty and scientific experiment create the
freedom to act, allowing citizens to adapt their position in the world to the realities
it contains. At this point in time scientific experiment already makes widespread
use of profiling technologies and it may be the case that legal certainty will need
profiling technologies to interpret the overload of legal cases and decisions
regain some control over the way one’s personal data are put to use.
The biggest challenge however may be how to constrain profiling practices in
order to prevent the coming-of-age of a technological infrastructure that is entirely
geared for dataveillance, normalisation and customisation – practically destroying
the effectiveness of our rights to privacy, fairness and due process (Leenes and
Koops, 2005: 329-340). It would be unwise to wait for such an infrastructure to be
in place, before establishing constraints, as this may render effective restraint an
illusion. In chapter 15 these issues will be discussed, in order to assess the implica-
tions of profiling for the identity of the European citizen.
2.6 Reply: Further Implications?
In her essay, Mireille Hildebrandt raises the important issue of considering profiles as
knowledge itself, and not as mere information. This implies a new way of considering pro-
files and profiling, as knowledge is subject to interpretation and meaning and inseparable
from the rich social contexts in which it is embedded.
This perspective has some profound implications in the way the information society is
going to extract, manipulate and exploit data of human beings, or shall we say knowledge
and apply it to the design of new categories of applications and services. In the new infor-
mation society, applications “know” the people and not the other way around.
*Institut Européen D’Administration Des Affaires (INSEAD)
The sheer volume of case law that is published (online) would in the end destroy legal certainty,
because no human individual would be able to find her way in the proliferating decisions.
For example, by the use of private Identity Management Devices that enable tracking of one’s
data and can be used to restrict the leaking of personal data. To regain control written law in itself
will not suffice; the right to hide certain data must be inscribed into the technologies that would
otherwise threaten one’s personal autonomy.
30 T. Nabeth
2 Defining Profiling: A New Type of Knowledge? 31
In our reply to Mireille Hildebrandt, we will further explore the implications of this shifting
of conceptualisation of profiles from data to knowledge and the ensuing consequences of
almost intimate understanding of people and groups. We will in particular try to understand
in which cases profile-informed applications will be used to better serve people and groups,
or will - on the contrary - be used to alienate them.
The essay of Mireille Hildebrandt on the subject of “defining profiling” comes very
much as a surprise but a pleasant one. One would initially expect a formal definition,
some descriptions of algorithms, some indication of security issues and a series of
illustrative examples, providing finally more of a description than an explanation of the
concept of profiling. What she provides though is a much more profound attempt to
understand the concept of “profiling” that borrows ideas from many different fields
and areas such as philosophy, complexity, anthropology and cognition (theory of
action). This perspective is particularly useful in providing readers with the conceptual
tools that will help them to articulate the different parts of this volume: the description
of profiling, including an in-depth discussion of algorithms and a first indication of
risks in part I; a series of illustrative examples (applications) that were presented in part
II and the wider implications for democracy and rule of law in part III.
2.6.2 Profiling as Knowledge
In the first part of her chapter, Mireille Hildebrandt engages the discussion on the
nature of the knowledge generated by the profiling process. This “profile” knowl-
edge originates from the automatic extraction from an important amount of infor-
mation aiming at discovering patterns that will have some predictive capabilities.
Indeed, the underlying assumption is that the function of profiling is to help to
reveal some hidden “order of things” and therefore to provide an oracle that will
predict how people will behave in the future. Indeed, if they have behaved in a cer-
tain way in the past, they will most probably behave the very same way in the
future. Mireille Hildebrandt also very rightly points to one of the main limitations
of this form of knowledge: profiling “knowledge” does not explain things, as it is
of a more inductive nature. It is then suggested that profiling should therefore be
complemented by well thought out profiling practices.
Mireille Hildebrandt then indicates what the different stakeholders of profiling
are: the subject (human or nonhuman, group or category) to which a profile refers
(referred to as the data subject), the entity of which data are used to generate pro-
files and to which a profile is applied (referred to as the subject) and the actor that
is the initiator of the profiling and the exploiter of the profiling data (referred to as
the data controller). This distinction is very important, since it raises the question
of different actors that may have conflicting objectives.
If Mireille Hildebrandt presents profiling at the individual level via the concept
of personalised profiling, we have to admit that profiling at the group level receives
a much higher level of attention in this chapter. Indeed, even if she acknowledges
the importance of personalised profiling, she does it principally for an application
in an Ambient Intelligent (AmI) context. We personally believe that it would have
been useful to generalise the reflection to a much broader context, such as the
domain of e-learning or e-commerce to cite a few.
At the group level, one should
distinguish between characteristics that belong to all the members of a group
(referred to as distributive attributes) and from characteristics that only statistically
belong to this group (referred to as non distributive attributes). It is of particular
importance to identify the non-distributive nature of the knowledge (also known
as non-monotonic logic
in artificial intelligence), since it can be at the origin of
errors in segregating people due to the merely probabilistic nature of some
Moreover, we should point out the danger of the segregating function of profil-
ing. Even in the case where the characteristic is distributive and no error is made,
how should we deal with using profiling, which usage can be directly associated to
segregation? The answer to this question follows an interesting angle: profiling did
not have to wait for the advent of the computer to appear, since society, and for
instance the social process, can be considered as a big profiling machine. Societies
use categorisation and generalisation to function better, since it allows anticipation.
To answer the question of whether generalisation is a good or bad thing, we will
follow the reasoning of Mireille Hildebrandt by excluding generalisations that do
not have a statistical or a factual basis. We also agree with the idea that some ethical
issues may apply, depending on the context and for instance, taking the example of
the “presumption of innocence”, the role of society should help to erase the ine-
quality that originates from circumstance. Even if it is proved that someone who
only has one parent is more likely to become a criminal, such knowledge should
not be used as a tool to segregate this category of persons; for instance by reducing
the level of protection provided by the presumption of innocence or by removing
some of their rights.
2.6.3 A Knowledge Ecology Perspective
The second part of this chapter is very interesting since it situates profiling accord-
ing to a systemic and knowledge ecology perspective (the term organic profiling
is used in this chapter). This part in particular relates to all the theories of complexity
See on this chapter 10.
See http://plato.stanford.edu/entries/logic-nonmonotonic/ for a description of the non-monitonic
32 T. Nabeth
2 Defining Profiling: A New Type of Knowledge? 33
and collective intelligence that have emerged during the last decades, e.g., with the
work of Varela and others (those involved in the Santa Fe movement). Applied
to the context of ambient intelligence, it draws a vision that is not far from
the ‘Universe’ of the great Sci-Fi author Philip K. Dick, for which the separation
between the real word and the virtual word tends to blur. In particular, with the
advent of RFID and other similar devices, this vision proposes to dissolve
the frontier between human and machine and in our belief, introduces the concept
of the trans-humanity that will merge the human and the machine. Indeed, RFID
represents the typical device helping to create the bridge between the physical and
the digital world (with RFID, the virtual world has access to “sensors” relating
what happens in the physical world).
The consequences of this vision of seeing the world as a system closely inte-
grated in society rather than as a well identified “machine” are many. First, and as
previously indicated, the distinction between the physical world and the virtual
world becomes artificial and should no longer be made, since we are talking about
the same world. Second, profiling the environment does not necessary “involve a
conscious mind” that is controlled by a central body (such as a government) but can
also happen quasi spontaneously in society by a variety of actors. Let us also add
that the existence of technology, even if it is not a mandatory condition for profiling
to happen, can have tremendous consequences. For instance, the combination of
autonomy and profiling can lead to the concept of autonomic profiling, for which
the profiling processes do not need to have the “man in the loop” and as a conse-
quence risk the loss of control by humanity. Even if we do not believe in the taking
over of society by machines – a doom scenario popular in Sci-Fi literature - a real
risk exists that people will lose the ability to control what is happening because of
the complexity (in particular if machines gain the capabilities to learn and adapt).
As a consequence, humanity may very well become dependant on profiling proc-
esses (typically in ambient intelligence) as is the case with an addiction: being
aware of the dangers but having no capabilities to act.
However, we do not believe that the consequences need to be apocalyptic.
The systemic vision may also mean a shift from the idea of very “controlled”
profiling systems counceived by engineers, to systems that are more “self-regulated”
and for which the designers also include people from the social sciences or law
field who know how to deal with less mechanical and less deterministic
approaches. In this later case, engineering systems involving profiling would
mean working on the level of the different feedback loops helping to regulate the
systems that continuously evolve (and deciding which ones are acceptable),
rather than very supervised systems in which profiling represents one of the
critical parts of an effective mechanism. To conclude, it would be the responsi-
bility of this new category of designers, able to reason in a more holistic way, to
ensure that the profiling mechanisms are put in place to service the good of
society and individuals (for instance profiling can enable better personalisation,
or can help to reduce inequalities by exposing them), rather than a tool of which
the role is only to enforce social control (and typically used to maintain people
in their initial condition).
2.6.4 What to Conclude About This Chapter
From Mireille Hildebrandt?
We feel there exists a risk, every time we enter into an epistemological, philosophical or
complex discussion, to detach too much from reality. Many discourses about the nature of
knowledge and complexity easily become very abstract and tend to lose their readers in
abstractions with little possibility to apply in reality.However, in this case Mireille
Hildebrandt was able to avoid this trap by providing an illustration of what may be the
concrete consequences for reality, for instance when applying it to ambient intelligence
environments. Her chapter is therefore successful in proposing a global picture of how to
conceptualise profiling at a higher level without losing the ground of reality.
If we could add something to this chapter, it would probably consist first in
incorporating research linking cognition and profiling and second in investigating
the consequences and impact of technology on the evolution of “profiled” environ-
ments. In the first case we will refer to the work that we will call instant cognition,
consisting in the unconscious perception / classification / generalisation that people
perform in their everyday life, which leads to very effective results but is also sub-
ject to bias and is at the origin of many dysfunctions in society, such as racism
without real intention (the reader is invited to read the book from Malcolm Gladwell
for information on this subject).
In particular, it could be interesting to explore
how the new profiling approaches can be used to counterbalance the biases we have
indicated (making them more visible). In the second case, it would be interesting to
investigate very futuristic scenarios exploring the limits of extreme profiling. For
instance, what would be the consequences of very “efficient” profiling done by the
society? The movie industry has already given us some food for the thought with
movies such as Gattaca
or the Minority report
that explore the less positive con-
sequences of profiling but we do not doubt that similar work can be conducted
exploring the more positive side, such as improving the effectiveness of education
or work via a better personalisation that profiling would authorise.
2.7 Reply: Direct and Indirect Profiling in the Light
of Virtual Persons
In our reply, we elaborate the difference between individual and group profiling in a
slightly different manner, by distinguishing and defining direct and indirect profiling. We
study these two types of profiling in the light of virtual persons.
*VIP, Berne University of Applied Sciences and ESC, University of Lausanne, Switzerland
Gladwell, M. 2005.
Gattaca is a 1997 science fiction drama film that describes the vision of a society driven by lib-
Minority report describes a society able to predict the crimes before they happen and imprison
people for crimes that they have not yet committed but intended to.
34 D.-O. Jaquet-Chiffelle
2 Defining Profiling: A New Type of Knowledge? 35
In direct profiling, data is typically collected for one single subject or a small group of
subjects. Knowledge built on this data then only applies to this specific subject or this small
group of subjects.
Direct profiling can be used to uniquely characterise a person within a population or to
infer, for example, future behaviour, needs or habits of a specific target.
In indirect profiling, data is collected from a large population. Groups and categories of
subjects with similar properties emerge from the collected data. Each group has its own iden-
tity defined through a small amount of information. The typical member of one group can be
modelled using the concept of virtual persons. It is then sufficient to identify a subject as a
member of the group, i.e., with the corresponding virtual person to be able to infer, for this
subject, knowledge inherited from the group itself: probable behaviour, attributes, risks, etc.
In this chapter, Mireille Hildebrandt presents three key concepts related to profil-
ing, namely the data subject, the subject and the data user. First, we want to
enlighten these concepts using the concept of virtual persons. Then we will refine
the concepts of individual and group profiling by distinguishing direct and indirect
profiling. The different types of profiling will be illustrated using the generic model
based on the concept of virtual persons.
2.7.2 Individual and Group Profiling
Individual profiling is used either to identify an individual within a community or
just to infer its habits, behaviour, preferences, knowledge, risks, potential or other
social and economic characteristics. Forensic individual profiling, for example,
covers both aspects. Commercial individual profiling on the other hand is more
interested in the latter, the inference of knowledge or rules about the individual.
Group profiling is used either to find shared features between members of a pre-
defined community or to define categories of individuals sharing some properties.
Forensic group profiling could, for example, find common characteristics in the
community of convicted murderers or define risk categories of individuals. More
generally, group profiling often raises ethical issues as it can lead very quickly, for
example, to discrimination.
Several techniques can be used together or separately to define a direct profile:
Information collected about an individual may directly give some important
attributes of his profile (age, gender, etc).
Data mining techniques applied to the data collected about an individual may
help to induce his habits, his preferences, etc.
Data mining techniques also help to find correlations between large sets of data
collected about groups of people. These correlations might allow in turn the cre-
ation of categories: for example individuals sharing some attributes, living
downtown, earning more than €100,000 a year, etc. Profiles are defined by asso-
ciating knowledge with each category.
Subsets are defined as elements sharing some properties. With each subset found in
this process is associated its profile: attributes, rules, preferences, etc.
A category results from a process of generalisation. Each defined subset can be
virtualised in a generalised subset or category, defined by the properties identifying
the original subset; it inherits the profile of the original subset. The generalised
subset may then exist independently of the original data subject.
2.7.3 Virtual Persons
We have elaborated the concept of virtual persons within the second work package
of FIDIS Identity of Identity.
Virtual persons create an abstract layer allowing a
more faithful description of many real-life scenarios appearing in our modern soci-
ety. We want to apply this generic model to data mining and profiling, in particular
to the data subject, the end user and the categories.
Virtual persons traditionally refer to characters in a MUD (Multi User
Dungeon), MMORPG (Massively Multiplayer Online Role Playing Games) or
other computer games.
These characters interact in a game; some of them rely
on human beings (players) for their actions and/or behaviour, while others
data mining / data analysis
Fig. 2.1 From information to knowledge
Jaquet–Chiffelle D.-O., Benoist, E., Anrig, B., Chapter 3 of Nabeth, T. et al. (eds), 2006a and
Jaquet–Chiffelle D.-O., Benoist, E., Anrig B, in Jaquet – Chiffelle D.-O., Benoist, E., Anrig B.
(eds.), 2006b: 6-7.
http://dud.inf.tu-dresden.de/Anon_Terminology.html (Version 0.28; May 29, 2006).
36 D.-O. Jaquet-Chiffelle
2 Defining Profiling: A New Type of Knowledge? 37
might be directed by the game itself. For an external observer, it may be impos-
sible to decide whether the subject behind a specific virtual person is a real
player or just a computer programme. We see these virtual persons (characters)
as masks used by subjects (human players, computer programmes) to act and/or
interact within the game.
Laws also create a virtual world by associating rights, duties and/or responsibili-
ties with virtual persons.
The one who is older than 18, the one who is married,
the one who is president of a company, the person legally responsible… are typical
examples of virtual persons living in the virtual legal world. These virtual persons
are not linked to any physical or legal entity until the given conditions described in
the law are fulfilled. Several physical or legal entities can be linked to these virtual
persons as actions and/or transactions take place. Moreover, a single physical or
legal entity may be linked to several virtual persons. These links are very often time
As an example, we consider the person legally responsible in a given transac-
tion. The subject, i.e., the physical or legal entity behind this virtual person, could
be the person executing the transaction himself; but it could be someone else, not
necessarily visible: a tutor or the parents of a child – it could even be a company.
Such analogies between multiplayer games and real-life scenarios extend the
field of application of virtual persons.
The concept of an abstract subject used by some authors is very close to our
concept of virtual person. However, in using virtual person, we take advantage of
the similarity between characters appearing in computer games and characters cre-
ated in our daily life scenarios. Moreover, etymologically speaking, person comes
from personae which means mask. Instead of adding a new theoretical term, we
extend a well-known concept that is easy to imagine, even for non-specialists. Last
but not least, in using two very distinct terms (subject and virtual persons), we avoid
a possible confusion between both concepts and emphasise their differences: the
virtual person is like a mask, the subject is the entity behind this mask.
In order to better understand the concept of virtual persons, we need a few core
definitions. A subject is any physical or legal entity having – in a given context –
some analogy with a physical person. Here subject is not opposed to object. Indeed,
physical objects may satisfy our definition of a subject. In our definition, subjects
typically act or play a role. Our subjects are, they have, they do or they know some-
thing just like physical persons.
Typical subjects are persons or groups of persons
but can also be animals or computer programmes for example. A subject can be
alive or not, can exist or not.
In her article, Danièle Bourcier (2001: 847-871) introduces the concept of virtual persons in the
context of artificial intelligence (intelligent programmes, software-agents, etc). She also refers to
previous uses of this term in similar contexts. Our concept of virtual persons covers this approach
while being more general and more adaptable to a wide variety of real-life scenarios.
Our subjects look like the grammatical «subject» in a sentence as pointed out by Sarah Thatcher,
London School of Economics, during the FIDIS WP2 workshop in Fontainebleau (December 2004).
A virtual person is a mask for a subject. In the context of individual authentica-
tion and/or identification, a virtual person is usually defined by what it is and/or
what it has and/or what it does and/or what it knows.
The one who knows your credit card PIN code is a virtual person defined by
what it knows. The subject behind this mask should be yourself and yourself only.
However, the subject can be a group of persons (e.g., you gave the PIN code to
other members of your family) or there might be no subject at all (e.g., you do not
remember the PIN code).
More generally, a virtual person can also be defined by its attribute(s) and/or its
role(s) and/or its ability(-ies) and/or its acquisition(s) and/or its preference(s) and/
or its habit(s), etc.
The President of the United States is a virtual person defined through its role;
the subject linked to this virtual person might change after each Presidential election.
However, rights, duties and responsibilities described in the law and associated
with this role, i.e., with this virtual person, do not depend on who is elected.
Actually, a virtual person acquires its own existence, which is not necessarily
correlated with the existence of any real subject behind it.
Fig. 2.2 Virtual persons
38 D.-O. Jaquet-Chiffelle
2 Defining Profiling: A New Type of Knowledge? 39
18.104.22.168 Virtual Persons Applied to Profiling
In her contribution, Mireille Hildebrandt gives the following definition for the data
subject: “subject (human or non-human, individual or group) that data refer to.” Her
concept of subject is covered by our own definition of this term and is therefore
compatible with our approach.
Using data mining techniques, subsets of elements sharing some properties can
be defined. Virtual persons allow representation of the corresponding categories.
With each category, i.e., with each virtual person, is associated the inherited
Indeed, with each virtual person is associated attributes, rules, preferences, etc.
deduced from the correlations found via the data mining techniques: its profile. For
example, people living downtown and earning more than €100,000 a year are likely
to be more than 30 years old and not retired.
The one who lives
The one who earns
more than 100 K
The one who lives
downtown and earns
more than 100 K
Fig. 2.3 Subsets of data subjects with their corresponding virtual persons (categories)
Fig. 2.4 Profile associated with a virtual person
At a later point, this knowledge may be used to infer probabilistic characteristics
about what Mireille Hildebrandt calls the “end user” or the “profiled data subject.”
Virtual persons acquire their own existence, which no longer depends on any
specific, original subset of data subjects. Different data subjects can lead to equiva-
lent subsets that define the same category, i.e., the same virtual person
2.7.4 Direct and Indirect Profiling
Profiling an end user consists in finding his profile by linking the end user to vir-
Information gathered about the end user enables the data controller
to find virtual persons linkable to this end user and to use the corresponding pro-
files for this end user. We consider that the classical distinction between individ-
ual and group profiling is not precise enough. We want to refine these concepts
using direct and indirect profiling. Direct profiling occurs when the end user and
the original data subject used to define the virtual person with its profile are the
same. Indirect profiling aims at applying profiles deduced from other data sub-
jects to an end user.
22.214.171.124 Direct Group Profiling
In the first chapter, Mireille Hildebrandt gives two examples of group profiling. In
the case of a pre-existing community (members of a local church, students living in
a certain dormitory) data are “collected, aggregated, stored and processed, in order
Fig. 2.5 Equivalent subsets defining the same category
In case of a direct profiling, it is essentially a direct construction of the profile. In case of an
indirect profiling, it is the construction of the profile by applying a typical profile of a category.
40 D.-O. Jaquet-Chiffelle
2 Defining Profiling: A New Type of Knowledge? 41
to find shared features”. Knowledge about this community (data subject) is estab-
lished as the profile of the virtual person defined by this group.
When the end user is later the community itself, we have a typical example of a
direct group profiling.
126.96.36.199 Indirect Group Profiling
Another example of group profiling given by Mireille Hildebrandt explains how
data mining techniques find subsets of individuals in the group, who share certain
attributes. This case illustrates the analogy between group profiling and the natural
process of categorisation and generalisation of Schauer also described in the first
Each category defines a corresponding virtual person. The profile of this virtual
person can then be applied (successfully or not) to any group (end user) linked to
this virtual person. This gives a typical example for an indirect group profiling.
188.8.131.52 Direct Individual Profiling
In the case of individual profiling, the data subject contains one single element, the
individual himself. Information is gathered about this individual and processed
using data mining techniques, for example, in order to define his profile.
Knowledge in this profile derives directly and exclusively from the information
about this individual. Such a profile will typically describe his habits and prefer-
ences, directly deduced from the observation of him.
This profile is then used for the individual himself in order to anticipate, for
example, his actions, his behaviour or his preferences. This is what we call direct
The one who lives
The one who earns
more than 100 K
Fig. 2.6 Direct group profiling
184.108.40.206 Indirect Individual Profiling
Direct individual profiling produces knowledge. This knowledge can in turn be
mapped to pre-existing compatible virtual persons in order to infer probable pro-
files for this individual. These probable profiles come from pre-existing group pro-
files. This is what we call indirect individual profiling.
As an example, an insurance company might use group profiles in order to esti-
mate risks associated with a potential client. If the person smokes, the group profile
associated with the virtual person the one who smokes is used to infer probable
characteristics of the potential client.
In a recent paper,
the authors explain how the knowledge of what a consumer
watches on television (direct individual profiling) allows us to infer demographic char-
acteristics about this consumer, such as his age or gender (indirect individual profiling).
The one who lives
The one who earns
more than 100 K
Fig. 2.7 Indirect group profiling
Spangler, W.E., Hartzel: K.S. and Gal-Or, M., 2006: 119-123.
Data subject = End user (one entity)
Fig. 2.8 Direct individual profiling
42 D.-O. Jaquet-Chiffelle
2 Defining Profiling: A New Type of Knowledge? 43
The online bookstore Amazon gives personalised advice such as “people who
have bought this book have also bought these others”. Furthermore, it proposes per-
sonalised offers when the client is recognised through a cookie or when he enters his
personal account. Those are typical examples of indirect individual profiling.
Real-time adaptive indirect individual profiling is part of the vision of the future
AmI space, where the environment interacts with the individual in order, for exam-
ple, to anticipate his needs.
Individual and group profiling have been refined using the new concept of direct and
indirect profiling. While direct profiling is expected to be more reliable, indirect profil-
ing uses the full potential of knowledge based on categorisation and generalisation.
We have shown how the generic model of virtual persons helps to describe pro-
filing types. The four combined types of profiling have been illustrated using this
model: direct group profiling, indirect group profiling, direct individual profiling
and indirect group profiling.
Aarts, E. and Marzano, S., (eds.), The New Everyday. Views on Ambient Intelligence. 010
Publishers, Rotterdam, 2003.
Bourcier, D., ‘De l’intelligence artificielle à la personne virtuelle: émergence d’une entité juridique?’,
Droit et Société, Vol. 49, l’Association française Droit et Société, Paris, 2001, pp. 847-871.
Custers, B., The Power of Knowledge. Ethical, Legal, and Technological Aspects of Data Mining
and Group Profiling in Epidemiology, Wolf Legal Publishers, Nijmegen, 2004.
de Waal, F.B.M., The Ape and the Sushi Master: Cultural Reflections of a Primatologist, Basis
Books, New York, 2001.
Fig. 2.9 Indirect individual profiling
Edens, J.R., ‘Misuses of the Hare Psychopathy Checklist-Revised in Court’, Journal of
Interpersonal Violence, Vol. 16, No. 10, Sage publications, London, 2001, pp. 1082-1094.
Gladwell, M., Blink: The Power of Thinking Without Thinking, Little, Brown and Company,
International Telecommunications Union (ITU), The Internet of Things, 7
Internet ITU report,
2005. Available (for purchase) at: http://www.itu.int/osg/ spu/publications/internetofthings/.
ISTAG (Information Society Technology Advisory Group), Scenarios for Ambient Intelligence in
2010, IST – IPTS report, EC, 2001. Available at: http://www.cordis.lu/ist/istag-reports.htm.
Jaquet – Chiffelle D.-O., Benoist, E., Anrig, B., ‘Virtual? Identity’, Chapter 3 of Nabeth, T. et al.
(eds), Set of use cases and scenarios, FIDIS Deliverable 2.2, European Union IST FIDIS
Project, 2006a. Available at: http://www.fidis.net/fileadmin/fidis/deliverables/fidis-wp2-
Jaquet – Chiffelle D.-O., Benoist, E., Anrig B,, ‘Virtual persons applied to authorization, individ-
ual authenticatinon and identification’, Jaquet – Chiffelle D.-O., Benoist, E., Anrig B. (eds.)
FIDIS brochure Deliverable 2.6, ‘Identity in a Networked World: Use Cases and Scenarios’,
2006b, pp. 6 and 7. Avalable at www.Vip.Ch
Kephart, J.O., Chess, D.M., ‘The Vision of Autonomic Computing’, Computer, Vol. 36, No. 1,
IEEE Computer Society, Washington DC, 2003, pp. 41 - 50.
Lawton, G., ‘Machine-to-Machine Technology Gears Up for Growth’, Computer, Vol. 37, No. 9,
IEEE Computer Society, Washington DC, Sept. 2004, pp. 12-15.
Leenes, R. and Koops, B.J., ‘ ‘Code’: Privacy’s Death or Saviour?’, International Review of Law
Computers & Technology, Vol. 19, No. 3, Routledge, Oxford, 2005, pp. 329-340.
Maturana, H.R. and Varela, F.J., Autopoiesis and Cognition: The Realization of the Living, Reidel,
Maturana, H. R., Varela, F.J., The Tree of Knowledge. The Biological Roots of Human
Understanding, Shambala, Boston and London, 1998 (revised edition).
Peirce, Ch.S., Pragmatism as a Principle and Method of Right Thinking. The 1903 Harvard
Lectures on Pragmatism, edited and introduced with a commentary by Patricia Ann Turrisi,
State University of New York Press, Albany,1997.
Peppers, D., Rogers, M., The One to One Future, Currency, New York, 1993.
Plessner, H, Die Stufen des Organischen under der Mensch. Einleitung in die philosophische
Anthropologie, Suhrkamp, Frankfurt, 1975.
Polanyi, M., The Tacit Dimension, Anchor Books, Garden City, New York, 1966.
Rafter, R. and Smyth, B., ‘Passive Profiling from Server Logs in an Online Recruitment
Environment’, paper presented at the IJCAI’s Workshop on Intelligent Techniques for Web
Personalization, Seattle, Washington 4-6 August 2001. Available at http://maya.cs.depaul.
Rasch, G., Probabilistic Models for Some Intelligence and Attainment Tests, University of
Chicago Press, Chicago, Illinois, 1980.
Schauer, F., Profiles Probabilities and Stereotypes, Belknap Press of Harvard University Press,
Cambridge, Mass. London, England, 2003.
Spangler, W. E., Hartzel K. S. and Gal-Or M., ‘Exploring the Privacy Implications of Addressable
Advertising and Viewer Profiling’, Communications of the ACM, Vol. 49, No. 5, ACM Press,
New York, 2006, pp. 119-123.
Stergiou, C. and Siganos, D., ‘Neural Networks and their Users’, Surprise 96 Journal, Vol. 4,
Department of Science Technology and Medicin, Imperial College London, London,
1996. Available at: http://www.doc.ic.ac.uk/~nd/surprise_96/journal/vol4/cs11/report.html #
Thorndike, R.L. (ed.), Educational Measurement, American Council on Education, Washington,
D. C., 2nd edition, 1971.
Turvey, B., Criminal profiling: An introduction to behavioral evidence analysis, Academic Press,
New York, 1999.
44 D.-O. Jaquet-Chiffelle
2 Defining Profiling: A New Type of Knowledge? 45
Van Brakel, J., ‘Telematic Life Forms’, Techné: Journal of the Society for Philosophy and
Technology, Vol. 4, No. 3, DLA, Blacksburg, 1999. Available at http://scholar.lib.vt.edu/ejournals/
Vedder, A., “KDD: The challenge to individualism.” Ethics and Information Technology, Volume 1,
Number 4, Springer, 1999, pp. 275-281.
Weiser, M., ‘The Computer for the Twenty-First Century’, Scientific American, 265, 3, Scientific
American Inc, New York, 1991, pp. 94-104.
Wilkie, William L. and Moore, E.S., ‘Scholarly Research in Marketing: Exploring the “4 Eras” of
Thought Development’, Journal of Public Policy & Marketing, Vol. 22, No. 2, AMA
Publications, Chicago, 2003, pp. 116-146.
Zarsky, Tal Z., ‘ “Mine Your Own Business!”: Making the Case for the Implications of the Data
Mining of Personal Information in the Forum of Public Opinion’, Yale Journal of Law &
Technology, Vol. 5, 2002-2003. Available at: http://research.yale.edu/lawmeme/yjolt/files/