Are Algorithms Value-Free?
Feminist Theoretical Virtues in Machine Learning
Gabbrielle M Johnson
As inductive decision-making procedures, the inferences made by machine learning
programs are subject to underdetermination by evidence and bear inductive risk. One
strategy for overcoming these challenges is guided by a presumption in philosophy
of science that inductive inferences can and should be value-free. Applied to machine
learning programs, the strategy assumes that the influence of values is restricted to data
and decision outcomes, thereby omitting internal value-laden design choice points. In
this paper, I apply arguments from feminist philosophy of science to machine learning programs to make the case that the resources required to respond to these inductive challenges render critical aspects of their design constitutively value-laden. I
demonstrate these points specifically in the case of recidivism algorithms, arguing that
contemporary debates concerning fairness in criminal justice risk-assessment programs
are best understood as iterations of traditional arguments from inductive risk and demarcation, and thereby establish the value-laden nature of automated decision-making
programs. Finally, in light of these points, I address opportunities for relocating the
value-free ideal in machine learning and the limitations that accompany them.
According to a 2018 Pew Research Center survey, 40% of people believe that algorithmic
decision-making can be objective, free from the biases that plague human decision-making.1
Reading this result, one might reasonably ask what it really means for an algorithm to be
objective and free from bias: what does it mean for an algorithm to be value-free? There
are at least three interpretations of this question. On the least sophisticated interpretation,
we are asking whether algorithms operate wholly free of any influence of human values.
1Smith 2018. 58% of respondents believe that computer programs will always reflect human biases.
Provisionally forthcoming in Journal of Moral Philosophy, special issue on “Justice, Power, and the Ethics of Algorithmic Decision-Making”
Are Algorithms Value-Free?
The algorithmsâ€”we might answerâ€”are just math, the data on which they operate are just
facts; at no point in explaining their operation do we need to make reference to human
values whatsoever. This, however, seems obviously false. Problematic social patterns unquestionably exist and are necessarily encoded in the data on which algorithms operate.2
On a second and slightly more sophisticated interpretation, we might recognize the unavoidable encoding of such patterns in the data, and ask instead whether the algorithms
themselves, their designs, are value-free. And here we might answer that even if the data
upon which the algorithms operate are shaped by human values, perhaps the engineers are
still doing the best with what they are given by making value-free design decisions. On
the other hand, we might reason that algorithms often fail to be value-free because their
all-too-human engineers are subject to worldly pressures that in fact command the importation of human values: say, for example, the pressure to produce the highest profits for
oneâ€™s company. On this reasoning, algorithms could be value-free in principle, howeverâ€”as
a descriptive factâ€”they are not. There exists still a third interpretation of the question, on
which we ask whether it is really possible for algorithms to be value-free even in principle.
That is, is it possible for even a superhuman engineerâ€”one impervious to worldly, selfinterested pursuitsâ€”to produce an algorithm that is value-free? It is this third possibility
that occupies me in this paper: when I ask if algorithms are value-free, I am asking whether
values are constitutive of the very operation of algorithmic decision-making, such that on
no idealized conception could they be value-free.3
Debates about at what point (if any) values can and should enter a decision-making
procedure have been popular in various areas of philosophy. For example, the issue of to
what extent epistemic norms in belief formation could be aâ†µected by practical and moral
aims has been widely discussed in literature on both pragmatic and moral encroachment.4
2I discuss the relationship between these problematic social patterns and the operation of algorithmic
bias in more detail in Johnson 2020a. 3Even this question will require some further precisification. In our academic musings, we can no
doubt conjure up imagined algorithms operating on entirely fictitious data sets, whose decisions are totally
divorced from real-world use. Iâ€™m not interested in asking whether those algorithms can be value-free. The
set of algorithms that I am interested in includes paradigmatically algorithms that are used to replace or
supplement human decision-making, that operate on real-world data, and whose decisions come to impact
other human agents. Although possibly failing to include all conceivable algorithms, I still expect this
class to be quite expansive, and certainly to include the algorithms that take center stage in discussions of
algorithmic fairness and bias. I return to this point about the range of conceivable algorithms at the end
of the paper. 4See Stanley 2005, Fantl and Mcgrath 2007, Moss 2018, Basu 2018, 2019, Bolinger 2018, Gardner 2017,
and Munton 2017, 2019a,b, among others.
2 of 34
Are Algorithms Value-Free?
In philosophy of science, prominent debates continue to unfold concerning whether values
can shape not only the research programs scientists choose to pursue, but also practices
internal to scientific inquiry itself, such as evidence gathering, theory confirmation, and
scientific inference.5 Ultimately this is a debate about whether values are a constitutive
feature of scientific inductions. Like scientific inductions, machine learning programs use
evidence (or known data) to form predictions (or generalizations to new phenomena).6
Thus, there exists a natural but underexplored comparison between debates about objectivity in scientific inquiry and machine learning. In this paper, I take up this comparison
by adopting arguments against the value-free ideal in science and extending them to the
domain of machine learning. In so doing, I explore the extent to which machine learning
algorithms are, can be, or should be value-free.
The literature concerning debates about values in science and the literature surveying
philosophical perspectives on the use of machine learning programs are independently extremely vast. Thus, a comprehensive application of the former to the latter is well beyond
the scope of any one paper. Instead, the aim of this paper is to build a bridge between
two domains I anticipate will have much to contribute to one another. I begin this task
by demonstrating how prominent arguments against the notion of scientific objectivity in
the form of the problem of induction, underdetermination by evidence, arguments against
demarcation, and the argument from inductive risk all straightforwardly apply to simple
cases of machine learning use. My hope is that these comparisons will facilitate continued
predictive and explanatory exchange between the algorithmic and scientific domains.
I begin in Â§2 by situating the discussion about values in algorithmic decision-making
against the backdrop of the objectivity of induction more generally. In particular, I describe a notion of objectivity that all inductive procedures fail to meet due to the need to
adopt bias in overcoming the problems of induction and underdetermination. It is against
this backdrop that the so-called â€œvalue-free idealâ€ in science emerges. In Â§3, I present
two prominent arguments against the value-free ideal borrowed from feminist philosophy
of science and argue for their application in the domain of machine learning programs.
5See Rudner 1953, Levi 1960, Douglas 2000, 2009, 2016, Rooney 1992, and Longino 1995, 1996, among
others. 6Here and throughout I use â€˜machine learning programsâ€™, â€˜algorithmic decision-makingâ€™, and â€˜algorithmsâ€™
to pick out a broad class of automated programs that function by capitalizing on or â€œlearningâ€ from patterns
manifest in the data on which they are trained in order to build a predictive model. This includes a
wide range of machine learning programs, including supervised, unsupervised, and reinforcement learning
3 of 34
Are Algorithms Value-Free?
These arguments result in the view that both scientific and algorithmic decision procedures are deeply value-laden. In Â§4, I address a possible response on behalf of value-free
ideal proponents: although many aspects of these decision procedures are value-laden, it
might nonetheless be possible to relocate the value-free ideal by restricting it to just those
processes responsible for updating probabilities in light of a fixed set of evidence. If so,
then the value-free ideal is still arguably a desirable pursuit, albeit highly restricted. I end
by addressing this reply, arguing both that it might nonetheless render algorithms valueladen due to the failure of demarcation and that, although such a restricted role for the
value-free ideal might be salvageable, its application to machine learning programs would
render their use extremely limited.
2 Origins of the Value-Free Ideal
In this section, I want to trace the historical progression of a pursuit of objectivity in
scientific inquiry, and explore how it applies to the domain of machine learning. The unifying feature of the two domains is that both rely on induction. I regard as inductive any
inference that is ampliative, i.e., that goes beyond the information given in the premises.
This includes any inference that is non-deductive, and extends to both enumerative inductions, i.e., those that generalize from known instances to novel instances, and abductions,
i.e., inferences to the best explanation. As weâ€™ll see, inductive inference is critical to both
scientific theorizing and machine learning.
I start by discussing a notion of objectivity in induction that is undoubtably too strong,
but seems surprisingly widespread in folk conceptions of scientific practice. This is a form of
objectivity that Antony (2001, 2006, 2016)â€™s calls â€œDragnet objectivity.â€ The notion comes
from the 1950s-1960s TV cop show Dragnet, in which Los Angeles Police Sgt. Joe Friday
disciplines himself â€œto consider just the factsâ€”the raw, undisputed data of the matter,
unadorned with personal speculation and uncorrupted by emotional interest in the case,â€
a strategy encapsulated in his famous catch-phrase â€œjust the facts, Maâ€™am.â€7 Applying
this idea to scientific inference, the claim is that scientists should aim to favor hypotheses
on the basis of â€œjust the facts,â€ without the influence of personal values. Applying this
idea to machine learning, one can easily see why algorithmic decision-making is thought
to be more objective than human decision-making, since such programs are built to learn
7Antony 2006, 58.
4 of 34
Are Algorithms Value-Free?
from raw data without the interference of personal speculation or emotional interest.
Although it makes for a quaint picture, Dragnet objectivity is an impossible model of
scientific inquiry. The reason is that â€œraw dataâ€ by itself inevitably underdetermines various
conclusions we might draw about some subject matter.8 If ever we considered literally just
the facts, we would never be able to draw inductive conclusions. Indeed, not even Sgt.
Friday used Dragnet objectivity strictly speaking, since no human could. Induction is
useful precisely because it is ampliativeâ€”it allows us to go beyond what is given to learn
informative facts about the world.9 Crucially, the evidence itself is always and in principle
consistent with an indefinite (possibly infinite) number of diâ†µerent conclusions we could
draw. No finite amount of data will ever be able to narrow the hypothesis space to one, since
there will always be more than one hypothesis consistent with the data.10 This is known as
the problem of underdetermination of theory by evidence, and its roots can be traced back
to another famous problem: Humeâ€™s Problem of Induction. To understand fully Humeâ€™s
problem and the implications of it for the discussion of values, it helps to contrast properties
of inductive and deductive arguments. Induction diâ†µers from deduction in two important
ways. The first, which is often mistakenly taken to be the essence of Humeâ€™s problem,
is that induction, unlike deduction, fails to guarantee truth. Because the conclusions of
deductive arguments are always in some sense contained in their premises, if the premises
are true, then the conclusion is guaranteed to be true. Inductions, on the other hand,
are merely taken to provide probable support for some conclusion. If the premises of an
inductive argument provide a confidence of 99% certainty in some conclusion, that still
leaves 1% chance the conclusion could be false. So, in all inductions, thereâ€™s a chance we
could get things wrong.
However, this was not Humeâ€™s problem. Humeâ€™s problem concerns the second way in
which induction diâ†µers from deduction: in justification. The justification of deduction is
a priori and necessary. The justification of induction is not. Arguably, the justification
8This is even granting the idea that data could be â€œrawâ€ in some robust sense, presumably unadulterated
by human collection practices. For a range of criticisms, see Gitelman 2013. 9The notion of Dragnet objectivity is intended to demonstrate an uncontroversial philosophical point:
a list of facts alone will not allow us to form ampliative conclusions. Itâ€™s uncontroversial because one, itâ€™s
definitional on â€˜ampliativeâ€™, and two, even the most stringent and traditional epistemological views will
agree to it. For example, even objective Bayesians will agree that something more is needed beyond merely
a fixed body of evidence, namely some procedure for updating in light of that evidence (for the Bayesian,
Bayesâ€™ Rule). 10To put the point succinctly, for any theory T that is consistent with some body of observed evidence
E, we can imagine another theory Tâ€™ according to which E makes it seem like T is true, but it isnâ€™t. Evil
demons and brains in vats are common resources for philosophers constructing such cases.
5 of 34
Are Algorithms Value-Free?
of induction is contingentâ€”it depends on the world being a certain way.11 Thus, the
problem with induction isnâ€™t that our degree of support always allows some room to be
wrong, itâ€™s that there appears to be nothing to justify why known instances would provide
support to any degree whatsoever to predictions of unknown instances. Thereâ€™s nothing
logically at odds with the world suddenly becoming drastically diâ†µerent. Thus, whether
some premises provide support for some conclusion depends on certain contingent features
of the world, e.g., that the world continue to remain uniform and exhibit the patterns weâ€™ve
seen in the past and that are encoded in the premises. The problems of induction and of
underdetermination generalize to any inductive procedure that attempts to use patterns
present in evidence to make predictions about novel cases. Such inductive procedures are
a critical aspect of scientific theorizing, but theyâ€™re also a fundamental feature of machine
learning. As weâ€™ll see, the feminist arguments in this paper stem ultimately from these
related problems of induction and underdetermination.
Humeâ€™s problem of induction and the problem of underdetermination entail that no
inductive process can be objective to the degree demanded by Dragnet objectivity. However, the success of science and induction more generally is a testament to our regularly
overcoming both. Inevitably, we do this by taking certain facts for granted. By making
non-evidential assumptions, the evidence can guide us toward some conclusions rather than
others. The problems of induction and underdetermination apply to any domain of inquiry
in which we attempt to draw conclusions on the basis of limited data, and each domain
therefore comes with its own set of assumptions on which it relies. Examples of such assumptions include perceptual transformation principles (in the case of visual perception),
cognitive heuristics, biases, and Bayesâ€™ Rule (in the case of belief formation), theoretical
virtues, values, or paradigms (in the case of scientific inference), and parameters, filters,
or constraints (in the case of machine learning programs), to name just a few. Going forward, I call this broad collection of assumptions in diâ†µerent domains â€œcanons of inductive
inference.â€ What unifies them is that each serves as a â€œnon-evidential way of limiting the
hypothesis space to a tractable size.â€12 In other words, canons of inference are necessary
11This is admittedly controversial. Norton (2003, 650, 666-669) makes a compelling case by adopting
what he calls the material theory of induction, according to which â€œall inductions ultimately derive their
licenses from facts pertinent to the matter of the inductionâ€, and demonstrating how it might evade the
problem of induction. Of course, objective Bayesians, inductive rationalists, and logical probability theorists
might all disagree. For notable critiques of these views, see the literature cited in footnote 30. 12Antony 2016, 161. Antony uses the term â€˜biasâ€™ to pick out this general class, though she does not
consider cases of machine-based decision-making. To avoid debates about the normative connotations of
bias, I borrow the notion of â€˜canons of inferenceâ€™ from Douglas (2016), who borrows it originally from
6 of 34
Are Algorithms Value-Free?
means of overcoming underdetermination.
Induction requires the adoption of canons, but crucially, canons are not one-size-fitsall phenomena. There are many possible bridges one might adopt to traverse the gap
between evidence and theory, and there seem to be no a priori grounds for preferring some
bridges over others. Famously, Humeâ€™s own response to the problem of induction was to
claim that it was simply a â€œnatural instinctâ€ of humans to continue to perform inductions
under the assumption that nature is uniform, so unobserved instances will share similarities
with observed instances.13 This canon has become known as â€œthe principle of uniformity of
nature,â€ and it is widely adopted as one of the most fundamental bridges over the inductive
gap. Of course, even saying that the future will be like the past is itself pretty unhelpful,
since there are any number of ways in which one thing can be like another. Thus, the
trouble with adopting this as an overarching solution to the problems of induction and
underdetermination isnâ€™t that the principle is arguably false, but rather that itâ€™s trivially
true, and does little by itself to reduce the hypothesis space to a tractable size.14 So, the
question remains of what other assumptions scientists need to adopt in order to accomplish
the aims of science.
The debate within philosophy of science regarding the value-free ideal has centered
around precisely this question. Itâ€™s not a question of whether scientists must adopt
some assumptionsâ€”reasons weâ€™ve discussed render that point undebatableâ€”rather, itâ€™s
one about which assumptions scientists ought adopt. In other words, a debate about which
canons are acceptable and which are impermissible. A canonical answer to this question
was provided by Thomas Kuhn. Kuhn (1962) famously argued that various scientific research programs come with their own package of assumptions on which to base scientific
inferences, what he called a â€œscientific paradigm.â€ Crucially, he additionally claimed that
the only standards by which to evaluate these assumptions exist within the paradigms
themselves. In other words, the question of whether some method is acceptable can only
Levi (1960). However, I do regularly adopt and endorse the normatively neutral notion of â€˜biasâ€™ (Johnson
2020a,b). John Norton (MS, Ch. 5) adopts the notion of criteria, cautioning against the use of â€˜valuesâ€™ or
â€˜virtuesâ€™, since those connote that the canons could be ends in themselves (rather than means to a greater
end, namely truth) and that the choice among them is entirely unconstrained and amounts to a free choice
among scientific practitioners. 13On a very blunt analysis of machine learning, this single assumption takes up the core explanatory
burden for all cases of algorithmic bias. It is precisely the inductive assumption that the future will be like
the pastâ€”in an environment whose past is shaped by historical injusticeâ€”that produces predictive results
reflective of those historical injustices. Thanks to Ajitha Anand for helpful discussion of this point. 14This point serves as the basis of Goodman (1955)â€™s New Riddle of Induction.
7 of 34
Are Algorithms Value-Free?
be answered from within a particular paradigm, and no one paradigmâ€™s answer to that
question is better or worse than anotherâ€™sâ€”the two are incommensurable. This radical
claim eventually stuck Kuhn with the dreaded label of relativist.
Wanting to avoid the relativist label, Kuhn (1977) eventually walked back his claims,
and put forward a list of â€œtheoretical virtuesâ€ he suggested could serve as paradigmindependent, objective grounds for theory choice. These include accuracy, fruitfulness,
consistency (both internal and external), breadth of scope, and simplicity. These virtues
have now come to be known as â€œepistemic values,â€ and this haloed set was taken to provide
at least a benchmark answer to the question of which canons scientists ought adopt: above
all, they should adopt canons that promote epistemic values, which themselves are taken
to promote truth. Adherence to this list is what historically gave rise to the so-called
â€œvalue-free idealâ€ in science. The answer was benchmark since even Kuhn noticed that
these values will themselves often trade oâ†µ from one another. However, the critical insight
that formed the basis of the value-free ideal was that in deciding what theory or hypothesis
to adopt, scientists should always be guided by these epistemic virtues only, and not by
social, ethical, or political values.
As Douglas (2016, 611) notes, the â€˜value-free idealâ€™ starts to seem like a bit of a misnomer. She suggests a more apt label for the ideal would be â€˜epistemic-values-only-inscientific-inference idealâ€™. There are two important points this new label highlights. First,
as has been argued up to this point, there will always be some role for â€œvaluesâ€ (or canons).
However, those values (or canons), according to the ideal, will be limited to the epistemic.
The second important aspect of this alternative label is that it makes clear that the relevant
focal-point of debates surrounding the value-free ideal is scientific inference. Everyone on
both sides of the debate grants that values can guide some aspects of scientific practice,
e.g., what research projects get taken up, what personal goals of individual scientists motivate pursuing those research programs (fame, fortune, etc.), or which demographics tend to
populate which fields of research. Likewise, I take it that everyone will agree in the debate
about whether algorithms are value-free that values shape some aspects of algorithmic use:
what problems are addressed using machine learning programs, what overarching commercial aims individual companies have in producing algorithms, or which demographics tend
toward or away from the technology industry more generally. These considerations fall
â€œoutsideâ€ of inductive inference itself, namely, the point at which we decide to accept or
15See, for example, Lakatos (1970, 178)â€™s famous claim that Kuhnâ€™s theory rendered scientific theory
choice â€œa matter of mob psychology.â€
8 of 34
Are Algorithms Value-Free?
reject some conclusion, and thus, are irrelevant to the debate about the value-free ideal.16
Thus, in what follows, I set them aside in order to maintain focus on what seems the best
possible candidate for defending the value-free ideal, inference itself.17
It helps to take stock of the dialectic at this point. We began with a notion of objectivity
that was a (surprisingly popular) caricature of science: Dragnet objectivity. According
to this stereotype, science is a lot like deduction. When presented with the evidence
that the world providesâ€”just the factsâ€”we immediately (or a priori and with certainty)
know what conclusion is entailed by those facts. However, as we saw from the problems
of induction and underdetermination, science is not likeâ€”and in principle cannot beâ€”
deduction.18 Evidence alone (irrespective of canons of inference) is insucient to establish
ampliative conclusions. Thus emerged a more substantial and interesting target view of
science: the value-free ideal. According to this view, we accept that both evidence and
canons are essential to scientific inference, but we retain the value-free ideal of science by
restricting the canons to just those values that are properly epistemic and preclude any
canons that are shaped by social, ethical, or political values. Again, this is intended as a
theory of what is possible in principle. That as a matter of descriptive fact scientists often
fall short of this aim wonâ€™t do much to thwart the ideal. Thus, arguments against the valuefree ideal in science strive to show that even in principle, this ideal is unattainable. There
are two standard arguments that aim to establish this point, to be discussed in greater
detail in the remainder of the paper. The first is the argument against demarcation:
that the distinction between the epistemic and non-epistemic canons is untenable, because
demarcating between the two is itself an essentially value-laden enterprise. As weâ€™ll see,
this argument can be interpreted either as a claim about the justification of canons (that
in justifying the demarcation, weâ€™re left with a further choice between epistemic and non16They are, of course, not irrelevant to the discussion about values in science and technology more
generally. For a variety of perspectives on the relationship between values and these wider applications
of technology in society, see canonical work by Friedman (1995, 1997), Johnson and Nissenbaum (1995),
Friedman and Nissenbaum (1996), Nissenbaum (1996), Bowker and Starr (2000), Shmueli (2010), Kroll
et al. (2017), Kroll (2018), and Mulligan et al. (2019). 17I will also be setting aside complications that emerge from the actual production of machine learning
algorithms. For example, I set aside issues arising from inevitable performance errors and the fact that
often algorithmic development is spread out across multiple, modular developmental teams, rarely with just
one individual developer at the helm. The motivation here is the same: to identify the strongest possible
case for the value-free ideal. I take it that if it can be demonstrated that in the best possible case the
value-free ideal is untenable, then it will certainly be untenable in these more complicated scenarios as well. 18I donâ€™t mean that scientists never rely on deduction, only that induction is the canonical form of
9 of 34
Are Algorithms Value-Free?
epistemic virtues, leading quickly to a regress) or as a claim about the constitutive natures
of the canons themselves (that the canons are constitutively tied to value-laden features
of the environment in which theyâ€™re deployed). The second argument is from inductive
risk: that any attempt to form conclusions on the basis of evidence will inevitably run the
risk of getting things wrong, requiring an appeal to value-laden considerations for assessing
this risk. If these arguments succeed, then like the caricature of Dragnet objectivity with
which we started, the more substantial target view of the value-free ideal will need to be
The final critical insight of this section comes from recognizing that as inductive
decision-making procedures, machine learning programs are subject to these same problems of induction and underdetermination. As the discussion of Kuhn brings out, like in
the case of induction more generally, there will be no one solution to these problems in the
domain of machine learning programs. To put the point bluntly, there can be no algorithm
for building algorithms.19 In fact, computer scientists are likely familiar with many of the
points made here already, though perhaps not under the guise of Kuhnian theory.20 Model
builders undoubtably recognize that there are many diâ†µerent, yet acceptable ways to build
predictive models.21 There are of course norms within the professional community that go
a large way toward restricting the domain of acceptable methodsâ€”these norms arguably
comprise a Kuhnian paradigm. However, these norms cannot determinately settle every
question, and some decisions about how these choice points play out ultimately depend on
the goals of the predictive model together with individual aims of model builders. They
19This slogan, like most slogans, favors rhetoric at the expense of precision. It relies on an equivocation:
I mean â€˜algorithmâ€™ in the first use of the word in the way Kuhn (1977, 359) means it when he says that
an â€œalgorithm able to dictate rational, unanimous [theory] choiceâ€ is â€œnot quite [an] attainable ideal.â€ The
second use refers to the machine learning algorithms that are the target of this paper. The idea of the
slogan is that you can (of course) write algorithms for building algorithms for something or other. What
you canâ€™t do is write an algorithm for building algorithms for any arbitrary problem. And the problem of
how machine learning algorithms should respond to data across the boardâ€”like the problem of determining
what theory to adopt given some arbitrary set of evidenceâ€”is one of the latter problems. 20Itâ€™s often thought that Humeâ€™s problem of induction resurfaces in the domain of machine learning in
the form of the No Free Lunch Theorem (see, for example, Giraud-Carrier and Provost 2005, 2, Domingos
2012, 81, and Wolpert 2013, 2). Iâ€™m compelled by arguments made by Lauc (2019) that the NFL is more
closely akin to Goodman (1955)â€™s New Riddle of Induction, though that too famously connected to Humeâ€™s
problem in ways discussed. Either way, both interpretations lead to the result, argued for here, that in the
domain of machine learning, â€œthere is no learning without bias, there is no learning without knowledge.â€
(Lauc 2019, 484, echoing Domingos 2015, 64). For a greater discussion of these points, see Dotan 2020,
where the NFL is used to motivated conclusions similar to those of this paper. Thanks to Kathleen Creel
for helpful discussions about these points. 21The point applies to scientific modeling more generally. See, for example, Weisberg 2007.
10 of 34
Are Algorithms Value-Free?
might have to, say, choose between various types of regression models or make decisions
about what cost functions to adopt. These can be interpreted as computer scientists having to make decisions about what canons to adopt. If program engineers adhere to the
value-free ideal, then theyâ€™re apt to produce programs that draw conclusions from some
dataset in ways that maximize accuracy, fruitfulness, consistency, breadth of scope, and
simplicity. The claim made at the beginning of this paper that algorithms are objective
is, thus, charitably interpreted as the claim that so long as the decision points program
engineers are responsible for are resolved in ways adhering to these canons, the algorithm
itself is regarded as value-free.
The rest of this paper, then, explores the extent to which these decision points can be
resolved by adhering only to the Kuhnian list of epistemic virtues, excluding entirely any
appeals to social or ethical values. In what follows, I present objections to the value-free
ideal in science. I then argue that these same objections apply to the adoption of the
value-free ideal in the production, use, and evaluation of machine learning programs. Once
the objections are in place, I explore how scientists have historically responded to these
objections, and the extent to which these responses are available in the domain of machine
3 Against the Value-Free Ideal
In this section, I present two famous arguments from feminist philosophy of science against
the value-free ideal: the argument against demarcation and the argument from inductive
risk. Following the presentation of each, I discuss how such arguments can be straightforwardly extended to machine learning algorithms.
3.1 The Argument Against Demarcation
The first argument I present against the value-free ideal is borrowed from the work of Helen
Longino (1995, 1996) and disputes the very idea that we could demarcate between so-called
â€œepistemicâ€ and â€œnon-epistemicâ€ values in the first place.22
Longino (1995) makes this point by first curating an alternative list of social and ethical canons of inference drawn from historical work in feminist philosophy of science; this
list includes values of empirical adequacy, novelty, ontological heterogeneity, complexity of
22A similar approach is taken up by Rooney (1992).
11 of 34
Are Algorithms Value-Free?
interaction, applicability to human needs, and diâ†µusion of power. She then goes through
a stepwise comparison of values from the two lists, pitting traditional Kuhnian values and
feminist social and political values against one another and demonstrating various contexts in which the feminist values ought be favored. For example, consider the feminist
theoretical virtue of novelty, which Longino (1995, 385) argues requires new theories diâ†µer
significantly from theories that are currently accepted. This requirement would seem to directly contrast with the Kuhnian virtue of external consistency, which requires new theories
be consistent with theories that are currently accepted. Longino argues that both novelty
and consistency can be viewed as appropriate virtues to shape theory choice; however,
given that the two are at odds with one another, which virtue is adopted in any particular instance of scientific theorizing is a contextual matter, and crucially will be settled in
virtue of the socio-political features of that context. As Longino (1995, 396) states, her
arguments establish that â€œthose traditional values are not purely epistemic (if at all), but
that their use in certain contexts of scientific judgments imports significant socio-political
values into those contexts.â€
Longinoâ€™s arguments are complex, and there seem to me to be subtleties that admit of
multiple interpretations. The most straightforward interpretation of the argument takes it
as a point about socio-political values shaping the meta-decision about whether to choose
values from one list over the other. In the case of novelty, the claim is that feminists
adopt this virtue on the socio-political basis of aiming to depart from theories that have
facilitated gender oppression throughout history. External consistency, on the other hand,
is chosen on the socio-political basis of acceptance of the gender-oppressive status quo.
Thus, in both cases socio-political values guide us (either wittingly or unwittingly) in
accepting the canons that we do. No decision of which canons to adopt can ever be purely
epistemic or value-free. This, she argues, renders a strict demarcation between the two
lists on the grounds that one set is value-free untenable, because the choice between the
two lists is itself a value-laden decision. I regard this as the justification argument against
demarcation: if your justification for choosing an epistemic virtue over a non-epistemic
virtue (or vice versa) depends on social and political values, then a strict demarcation
between the epistemic and the non-epistemic is untenable.
The second, more subtle interpretation of Longinoâ€™s argument takes it as a point about
the natures of the values themselves: that in virtue of being context-dependent, epistemic
virtues necessarily imbibe the value-ladeness of the features of the environment to which
12 of 34
Are Algorithms Value-Free?
they correspond in that context.23 As another example, consider the contrast Longino
(1995, 393-394) presents between simplicity and ontological heterogeneity. Using a more
recent example, we can see this tension arise in cases of clinical drug trials. Consider the
case of Ambien. In 1992, the prescription drug Ambien was approved as a sleep aid by the
FDA. However, clinical trials for Ambien didnâ€™t take into account the average metabolic
diâ†µerences between men and women, resulting in a recommended dosage that was the
same for both. Twenty years later, research on the eâ†µects of sleep aids on impaired driving
found that women were being prescribed nearly twice the amount they should be, resulting
in many women who took Ambien at night to still have enough of the drug left in their
systems the next morning to impair their abilities to operate a motor vehicle.24
Because researchers took male metabolic systems to be the paradigmatic case, an assumption driven by an allegiance to canons that maximized the epistemic value of simplicity, they produced a drug that put the lives of many women taking the drug in danger,
a mistake that took two decades to rectify. Crucially for Longino, scientistsâ€™ decision to
posit the fewest kinds of entities in a context where members of a privileged classâ€”in this
case, malesâ€”are taken to be the primary fundamental entity, lead to theories that both
legitimate and perpetuate the socio-political values on which theyâ€™re built. In a world built
on socio-political values that result in a hierarchy where some members of the population
belong to a privileged class, those same individuals get prioritized in theory, and thus an
allegiance to simplicity will (either wittingly or unwittingly) imbibe the very socio-political
values on which the hierarchal relations are formed. Thus, to demarcate the non-epistemic
from the epistemic in any particular context is untenable. I regard this as the constitutive
argument against demarcation: if the adoption of any seemingly epistemic virtue in a particular context depends constitutively on the socio-political features of the context, then a
strict demarcation between the epistemic and non-epistemic is untenable.
Although never explicitly discussed by Longino, I take the demarcation argument (in
23This interpretation is a potential departure from Longinoâ€™s original intentions for the argument against
demarcation. Minimally, we can say it takes as inspiration Longinoâ€™s original argument. I personally find
it the most compelling case against the value-free ideal. 24Tavernise 2013. A similar story plays out in the diâ†µerential ecacy of HPV vaccine Gardasil in white
women and African American women. Gardasil primarily targets HPV types 16 and 18, while Gardasil 9
targets types 16, 18, 31, 33, 45, 52, and 58. However, African American women are about half as likely
to be aâœicted with types 16 and 18, and neither form of vaccine is eâ†µective against three types of HPV
that most commonly aâœict African American women, 35, 66, and 68 (Vidal et al. 2014). This is arguably
one contributing factor to rates of HPV-associated cervical cancers being higher among African American
women than among white women (Viens et al. 2016, 662).
13 of 34
Are Algorithms Value-Free?
both forms) to be deeply related to Humeâ€™s problem of induction, and in particular to
the second aspect of how induction diâ†µers from deduction: induction, unlike deduction, is
justified contingent on how the world is. Focusing first on the constitutive argument against
demarcation, thereâ€™s no such thing as an induction thatâ€™s justified a priori or irrespective
of facts about the world. Rather, whether any particular induction is justified seems to
depend on features of the world. In this way, features of an induction, e.g., being justified,
seem inseparable from features of the world, e.g., being uniform. Thereâ€™s no such thing as
justification in the abstract. Thereâ€™s only justification that depends constitutively on the
worldâ€™s being a certain way. So too, I like to think of Longino as claiming other features
we take to be independent features of inductions, e.g., being simple or being externally
consistent, are inseparable from features of the world, e.g., promoting the gender oppressive
status quo.25 If induction is justified only by contingent features about the world, and the
world is itself shaped by our values, then our values necessarily influence which canons will
be most likely to generate justified inductive arguments. Crucially, which values are apt,
even narrowly for the epistemic aim of getting us onto truth, will itself be a contingent
matter; it depends ultimately on how the world we live in happens to be.26 Given that how
simple (or externally consistent) some induction is will be a relation between the theory
and the world, the fact that the world has been shaped by various forms of oppression
entails that adherence to these epistemic values will de facto result in adherence to those
social and political values.
Likewise for the similarities between Humeâ€™s problem and the justification argument
against demarcation. As mentioned in the previous section, Humeâ€™s favored response to
25To make this point in another way using a framework familiar to computer scientists, consider the
Proxy Problem: often seemingly innocuous attributes that correlate with socially sensitive attributes, serve
as proxies for the socially-sensitive attributes themselves (I discuss this problem at length in Johnson 2020a,
see also Hellman (MS)). We could think of properties of inductive arguments as necessarily serving as proxies
for socio-political features of the world. Not only do properties of objects unwittingly encode sociallysensitive properties, but also properties of inductions can too. So, just as some property about objects
weâ€™re interested in (zipcode, say) might unintentionally be a proxy for some other feature of that object
(race), so too, I contend, some property about an induction (aiming for simplicity) might unintentionally
be a proxy for some other feature of that induction (perpetuating patterns of injustice). So something like
simplicity might appear value-free, but when you consider it against the backdrop of the subject matter
in which a decision procedure is adopted, it too becomes an unwitting proxy for other properties, like the
perpetuation of patriarchal structures. 26This point is related to the â€œtradeoâ†µâ€ other theorists have argued exists between epistemic reliability
and justice. If our standards for accuracy have been shaped by the social environment, which has itself been
shaped by oppressive structures, then there will arguably exist a tradeoâ†µ between demands of epistemic
reliability and demands of morality. See Gendler 2011, Basu 2018, 2019, Bolinger 2018, Munton 2019a, and
Johnson 2020a, among others.
14 of 34
Are Algorithms Value-Free?
the problem concerned the brute adoption of the assumption that nature will continue to
be uniform, i.e., the Principle of Uniformity of Nature.27 Now let us ask: on what grounds
might we justify the adoption of that assumption? An obvious response would be that
such a premise has held true for us in the past, and so we can expect that it will continue
to hold true in the future. But this is to use induction to justify induction, which would be
to beg the very question under discussion. One could think of Longino as making a similar
point: proponents of the value-free ideal neglect to recognize how their choice to adopt
some canons of inference over others is itself a value-laden judgement, one that comes
equipped with various biases in the form of allegiances to accuracy, simplicity, breadth,
etc. As Longinoâ€™s argument suggests, there might be contexts in which such biases are
apt. However, there might likewise be contexts in which biases in the form of other canons,
such as ontological heterogeneity and applicability to human needs might be more apt. Any
attempt to justify adherence to only epistemic virtues in the meta-decision about which
list to adhere to would itself require justification or else be question-begging. This leads to
an impossible regress in justification.28
Crucially, the arguments against demarcation seem to apply equally well in the case of
machine learning. As mentioned in the previous section, machine learning engineers will
have certain decision points left up to them. Ultimately, the decision to use some data
analysis method over others will depend on the aims of the program and the goals of the
programmer. Here, the question of how to justify some methods over others will likely be
answered by appeal to the value-free ideal: the decisions to, say, use a parametric rather
than a non-parametric model might be guided by the fact that the former is simpler than
the latter.29 Alternatively, one might adopt a non-parametric model due to its flexibility,
opting for a predictive model that more closely aligns with ontological heterogeneity. Crucially, the rationale for choosing one over the other seems itself open to scrutiny and calls
27This is bracketing Goodman (1955)â€™s point that the principle is trivially true and, thus, ultimately
unhelpful in resolving the problem of induction. 28I take it the nail in the con for the value-free ideal on this interpretation of the argument would be to
demonstrate that non-epistemic values alone can end the regress. Iâ€™m told that this sort of argument might
be constructed by drawing on Platoâ€™s argument for the priority of the Good. Iâ€™m not familiar enough with
that argument to make the case here. However, a more flat-footed approach might draw on Humeâ€™s own
response to the problem of induction to point out that justification has got to stop somewhere (see Ward
(MS)); eventually we have to proceed with scientific theorizing. Surely the decision to cut oâ†µ justification
at any particular point will therefore be a pragmatic decision, and thus one that depends on non-epistemic
values. Thanks to Jim Kreines for helpful discussion of these points. 29Of course, what simplicity ultimately amounts to is itself an elusive question in the history of philosophy
15 of 34
Are Algorithms Value-Free?
for justification. It is in providing this further justification that program engineers will
likely have to appeal to facts that go beyond the purely epistemic.30 They often include
considerations about the overall aim of the program and the context in which it is intended
to be used, facts which themselves depend on social and political factors. According to the
justification argument against demarcation, any further justification that involves social or
ethical considerations will render even those first-order decisions value-laden in significant
ways. Moreover, from the constitutive argument against demarcation, even abiding by a
seemingly pure epistemic list of considerations when making design decisions might usher
in socio-political values. A straightforward example of this is in the selection of a loss
function: this selection ideally corresponds to actual expected loss in making an incorrect
prediction; in this way, we want the function to accurately approximate loss. However,
there are many dimensions on which to measure actual loss, corresponding inevitably to
value-laden features of the world. Adoption of any particular loss function so as to approximate real-world loss will, according to the constitutive argument, necessarily imbibe those
loss functions with the very socio-political values of the real-world losses.
Whereas Iâ€™ve described the demarcation argument(s) as related to Humeâ€™s problem of
induction and the fact that inductions (unlike deductions) are contingent, justified only
relative to the worldâ€™s being a certain way, the next argument from feminist philosophy of
science against the value-free ideal focuses on the other way induction diâ†µers from deduction: in the potential for getting things wrong. According to this argument, any inductive
reasoning must always culminate in a decision that involves ethical considerations. This is
because at some point in the chain of justification, one must consider the possibility that
the prediction being made might get things wrong, a risk that inevitably comes with social
and ethical costs. It is the consideration of this risk and the consequences these arguments
have for the value-free ideal that I turn to next.
3.2 The Argument from Inductive Risk
As discussed in Â§2, canons of inference arise from a need to overcome underdetermination;
they are bridges over inductive gaps. Importantly, however, such bridges are inevitably
30This point can be bolstered by literature on probabilistic reasoning more generally that converges on a
similar point. A small sampling of that literature includes Ramsey (1989)â€™s criticism of Keynes that there
are no objective probabilities, Carnapâ€™s failure in constructing a purely formal foundation of inductive
logic (for summary, see Zabell 2011), Titelbaum (2010)â€™s rejection of a purely objective notion of evidential
support, and Fallis and Lewis (2016)â€™s critique of purely objective measures of probabilistic accuracy, among
many others. Thanks to Branden Fitelson for pointing me to this literature.
16 of 34
Are Algorithms Value-Free?
fallible. This is in the nature of biases and heuristics in general: they give us guidance
in cases where weâ€™re unsure, i.e., cases where weâ€™re not guaranteed truth. Such canons
can never guarantee truth, since to do so would be to turn induction into deduction, and
that would be tantamount to omniscient knowledge about the way the world will be in
unobserved cases. So, in all cases where we adopt canons of inference, i.e., in all cases of
induction, we run the risk of getting things wrongâ€”we call this â€œinductive risk,â€ and it
prompts the second argument against the value-free ideal.
The argument from inductive risk can be traced back to Richard Rudner, who states:
[S]ince no scientific hypothesis is ever completely verified, in accepting a hypothesis the scientist must make the decision that the evidence is suciently
strong or that the probability is suciently high to warrant the acceptance of
the hypothesis. Obviously our decision regarding the evidence and respecting
how strong is â€œstrong enoughâ€, is going to be a function of the importance, in
the typically ethical sense, of making a mistake in accepting or rejecting the
hypothesis. … How sure we need to be before we accept a hypothesis will depend
on how serious a mistake would be.31
The idea here is that, contra the value-free ideal, ethical values have a legitimate and
necessary role to play in guiding scientific inference, because they establish confidence
thresholds for ultimately accepting or rejecting a given hypothesis or prediction. This
point is made obvious by comparing two hypothetical scenarios. Imagine that in one
case, engineers are responsible for producing seat belt buckles, while in another scenario,
engineers are responsible for producing pant belt buckles.32 In both cases the engineers
run the risk of getting things wrong in producing defective buckles, but the consequences of
getting things wrong in the former case are much more dire than those of the latter. Clearly
the threshold for confirmation in the two cases should not be the same: we should demand
a much higher degree of confidence in the engineersâ€™ hypotheses in the first scenario than
in the second. Proponents of the argument from inductive risk insist that the threshold for
confidence can only be established by appeal to ethical values, thus rendering the decision
to adopt any particular hypothesis value-laden.
This line was picked up by feminist philosophers of science, most notably in the work
of Heather Douglas (2000, 2003, 2009, 2016). Douglas expands on Rudnerâ€™s initial point
by arguing that not only do scientific inferences involve such risk, but that the risk is
31Rudner 1953, 2, emphasis in original. 32This is an adaption of Rudner (1953, 2)â€™s original case.
17 of 34
Are Algorithms Value-Free?
compounded by the weight given to scientific judgements as expert testimony in various
social and political arenas. As Douglas (2016, 615) states, the â€œbaseline epistemic authority
brings with it general responsibilities to be neither reckless nor negligent in oneâ€™s actions
… scientists must consider in particular the impact of their authoritative statements.â€33
In sum, not only do scientists have to take into consideration the risk of getting things
wrong whenever they perform an inference to some hypothesis, but also they have to keep
in mind the influence that that wrong hypothesis will have in communities in which their
judgement is regarded as expertise. Importantly, all these points apply equally well, if not
more so, in the case of machine learning programs. Consider a minimal pair similar to
Rudnerâ€™s original case, but now in the domain of machine learning. For example, imagine
you?re tasked with building an image recognition program to distinguish human shapes
from non-human shapes. The level of inaccuracy you should tolerate will depend on the
use to which your algorithm will be put. If it is to be implemented in an oce complex
as a trigger to activate the automated lights, 75% accuracy would be inconvenient, but
acceptable. However, if it is to be implemented in a self-driving car to prevent pedestrian
collisions, you should demand near perfection. Algorithmic design decisions about how to
manage error therefore inherently involve values. We also see the analogue of Douglasâ€™s
point in this domain: not only do machine learning programs run the risk of getting things
wrong (a risk the negative consequences of which have been well-documented within the
machine learning activist communities), but also because of the computational prowess,
eciency, and ubiquity of machine learning programs, we can expect the eâ†µects of their
judgements to be wide-reaching and vast, likely more so than any individual scientistâ€™s
judgements.34 Thus, in building machine learning programs, it seems it is not sucient
that a program engineer merely adopt an aim of achieving some traditional canon like
accuracy, since thereâ€™s no such thing as accuracy neat. Instead, how accurate is accurate
enough necessarily involves some determination of the ethical consequences of getting things
33See also Douglas 2003, 2009, chapter 4. 34For a broad discussion of the impacts of algorithmic decision making on discrimination, law, and
policy, see Barocas and Selbst 2016, Kroll et al. 2017, Selbst et al. 2019, and Abebe et al. 2020, as well as
other scholarship produced in association with the Fairness, Accountability, and Transparency in Machine
Learning (FAT* ML) community.
18 of 34
Are Algorithms Value-Free?
Before moving onto possible responses to these arguments against the value-free ideal, it
helps to demonstrate the application of these points to a concrete case in machine learning.
One particular case that is popular in discussions of algorithmic bias and fairness concerns
criminal justice risk-assessment algorithms such as COMPAS. COMPAS, developed and
owned by Northpointe (now Equivant), stands for the Correctional Oâ†µender Management
Profiling for Alternative Sanctions. It is a program used by judges across the United
States to produce recidivism risk scores. The program works by collecting data about
defendants awaiting trial and, on the basis of statistical analysis, produces a risk score
that can then be used by judges to make decisions about, among other things, setting
bail, establishing the need for pretrial detention, sentencing, or parole. However, a 2016
exposÂ´e by ProPublica revealed that in an analysis of over seven thousand COMPAS uses,
the program was almost twice as likely to falsely label black defendants as future criminals
than white defendants, while often mislabeling white defendants as low risk at a higher rate
than black defendants.35 This prompted many advocates for fairness and transparency to
regard the program as problematically biased against African Americans.
However, further analysis proved that this initial assessment was not so simple. Computer scientists researching the so-called â€œimpossibility resultâ€ note that there are (at least)
three candidate notions of fairness between two groups of interest one could adopt in the
production of a risk-assessment program: first, its being well-calibrated, i.e., it has the
same degree of accuracy within each group; second, its having false-positive equality, i.e.,
the same proportion of false positives within each group; and third, its having false-negative
equality, i.e., the same proportion of false negatives within each group.36 The impossibility
result is that these three conditions cannot all be satisfied at once; they necessarily trade
oâ†µ from one another. Since this discovery, a heated debate has emerged regarding how best
to balance these potential notions of fairness and, given that we canâ€™t have all three, which
among the identified criteria of fairnessâ€”accuracy, balancing false positives, or balancing
false negativesâ€”should be given the most weight.
An extended philosophical discussion of the debates surrounding COMPAS and other
35Angwin et al. 2016. This analysis compared risk scores produced by COMPAS with occurrences of
rearrest over the following two years. 36This â€œimpossibility resultâ€ has been formally demonstrated in the work of both Chouldechova (2016)
and Kleinberg et al. 2016. Typically the relevant groups will be two that diâ†µer from one another along some
salient social dimension, e.g., race. Kleinberg et al. label the candidate notions of fairness â€˜well-calibratedâ€™,
â€˜balance the positive classâ€™, and â€˜balance the negative classâ€™, respectively.
19 of 34
Are Algorithms Value-Free?
criminal justice risk-assessment software is unfortunately beyond the scope of this paper.37
Instead, I intend only to apply lessons from the above discussion to demonstrate that two
points can be made to gain traction on this debate. The first is that our preliminary
discussion about the role of canons in inference establish that it would be a mistake to
expect a monolithic set of conditions that all risk-assessment algorithms ought satisfy.
To repeat an important point, there can be no algorithm for building algorithms. This
contention is bolstered by the fact that the three intuitive conditions of fairness identified
are at the same time reasonable and incapable of being jointly satisfied. Moreover, as
weâ€™ve seen from the argument from demarcation, reasonable canons of inference often
pull in opposite directions, and whether we opt for some over others will ultimately be a
The second point concerns which criterion we ought to adopt in the case of COMPAS
and the context in which itâ€™s used, namely the criminal justice system. Notice that those
who advocate for giving greater weight to balancing false positives are in essence reiterating
the traditional argument from inductive risk. Proponents of this view claim that the risk
for the black community of making a mistake and falsely flagging black defendants as likely
to reoâ†µend is high enough that we ought to establish a greater threshold of confidence in
making these assessments (or refrain from making them at all). Moreover, itâ€™s important
to note the point made by Castro (2019, 417-418) that, in the current context of use,
namely, the United States criminal justice system, â€œitâ€™s already disproportionately costly
to be black.â€ Thus, itâ€™s critical that a full assessment of inductive risk take into account
not only the harms of getting it wrong that are shared among black and white defendants,
but also the diâ†µerential harms on the black community and black defendants specifically
when theyâ€™re misidentified as high-risk. Moreover, defenders of the use of criminal justice
risk-assessment tools like COMPAS likewise use arguments about risk. On this side of
the issue, experts argue that to ignore the results of COMPAS and to instead rely on
the judgements made by individual judges (unaided by risk-assessment tools) would likely
37For detailed discussion, see Hellman 2019. 38A cottage industry has since developed within machine learning literature identifying and exploring
novel notions of fairness beyond Chouldechova (2016) and Kleinberg et al. (2016)â€™s initial lists. Mehrabi
et al. (2020, [11-12), for example, count at least ten diâ†µerent notions common to the literature. This trend
echos that in philosophy of science literature of identify and explore novel notions of theoretical virtues
beyond Kuhn (1977)â€™s initial list. Keas (2018, 2762-2763), for example, counts at least twelve diâ†µerent
notions common to the literature. Thanks to Katie Elliott for drawing my attention to these parallel
20 of 34
Are Algorithms Value-Free?
result in decisions that are more problematically biased, not less.39
Again, it is impossible to settle these disputes here. Instead, I use this discussion
to make the more modest point that value-laden debates are already well underway in
the domain of machine learning. That both sides attempt to use considerations of social
fairness and ethical risk in determinations of which notions of fairness to adopt, bolsters
the claim that decisions about what canons of inference to adopt in machine learningâ€”or
even about whether to deploy the program for use in the context of the criminal justice
system at allâ€”will necessarily involve value-laden considerations.40
4 Relocating the Value-Free Ideal
A popular defense of the value-free ideal in science begins by conceding part of the argument
from inductive risk.41 It accepts the point that the act of accepting or rejecting a hypotheses
necessarily incurs inductive risk and, therefore, requires appeal to ethical values. However,
it is for this reason, defenders argue, that scientists should never aim to accept or reject
hypotheses at all, but should instead only ever assign probabilities to hypotheses with
respect to a fixed set of evidence.42 Following this line of thought, one might argue that
in the case of COMPAS, the programâ€™s use is only ever to give a confidence rating that a
39See, for example, Flores et al. 2016, 17 and Corbett-Davies et al. 2017. Itâ€™s important to note also that
the aforementioned impossibility result is a mathematical generalization that applies to any decision-making
procedure, whether human or machine-based. So, the same sorts of disputes ought reoccur in analyses of
decisions made solely by human judges. 40Another interesting project emerges from considerations of adopting traditional feminist theoretical
virtues in the domain of machine learning. For example, looking back at Longinoâ€™s original list, we see as
one possible canon of inference the value of applicability to human needs. Longino (1995, 389) describes
this criterion as a call for scientific inquiry directed at reducing ongoing harms and â€œimproving the material
conditions of human life, or alleviating some of its misery.â€ She gives examples such as pursuing scientific
ends that produce sustainable agriculture, reverse the destruction of the environment, and assist the infirm.
In the context of COMPAS, one can use this value to make the case that software like COMPASâ€”software
that falsely identifies black defendants as high risk, compounds the harm done to the black community, and
perpetuates historical patterns of oppression and injusticeâ€”is antithetical to such aims. Thus, the specific
appeal to feminist theoretical virtues provides a platform for condemning reliance on this software in the
context of the criminal justice system. 41The line of defense that Iâ€™m about to recap here focuses narrowly on the argument from inductive
risk and leaves unaddressed the arguments against demarcation. Iâ€™ll return to this point at the end of the
section. 42Jeâ†µrey 1956. Putting the point in familiar Bayesian terms, one might argue that the role of a
scientist is not to accept some hypothesis H, but merely to assign some conditional probability for
H relative to some body of evidence E, and that this assignment ought accord to Bayesâ€™ theorem:
P(H|E) = P (E|H)P (H)
P (E|H)P (H)+P (E|Â¬H)P (Â¬H)
21 of 34
Are Algorithms Value-Free?
defendant might recidivate, but it should not aim to make a determinate prediction one way
or another. In fact, this arguably accords more closely with the stated use of the program.
For example, in its Practitionerâ€™s Guide, Northpointe (2015, 7) introduces the AIPIE model
of procedures for case management involving assessment tools like COMPAS. The AIPIE
model has five steps: Assessment, Interpretation of results, Plan based on the information
gathered, Implement the plan, and Evaluate the results of the action. This model utilizes
tools like COMPAS at the first stepâ€”in Assessmentâ€”suggesting that the remaining steps
of case management are left up to practitioners. Thus, a defender of the value-free ideal
might argue that the considerations of the risk of getting things wrong come only in the
interpretation, planning, and implementation phases. These are the steps wherein judges
make predictions about whether particular defendants will reoâ†µend and, although they
do this by interpreting the results provided by COMPASâ€™s risk scoring system, the risks
associated with the predictions are not inherent to the algorithm itself. One way to see this
is to notice that there can be several diâ†µerent decisions a judge might rely on COMPAS risk
scores to make. They might use COMPAS risk scores in deciding who should be granted
bail, or alternatively they might use it to decide to whom they should allocate resources
intended to prevent recidivism.43 This suggests that inductive risks associated with the
implementation of risk scores are associated with the decisions that are made by judges
using the program, and not risks inherent in the assessment produced by the algorithm.
Decisions about how and to what extent the risk scores are used in further decisions about
the defendant are beyond the control and, thus, ethical purview of the programâ€™s architects.
This seems like a natural response to the argument from inductive risk. However,
itâ€™s so natural that Rudner anticipates it in his original article. There, Rudner (1953, 4)
preemptively gives the response that even in assigning some probability p, the scientist
is accepting a kind of hypothesis, namely, a hypothesis â€œthat the degree of confidence is
p.â€44 He claims this new hypothesis itself comes with a risk of getting the confidence
wrong and, thus, the argument from inductive risk iterates. The move to confidences or
probabilities only pushes the problem back a level, it does not escape it. Looking back on
Douglasâ€™s arguments that scientists are taken as epistemic authorities, we might look to
evidence of to what extent mere probability estimates in fact determine the judgements
made on the basis of a scientistsâ€™ testimony, regardless of whether it was the intent of the
43As one anonymous referee points out, feminist considerations such as those presented in footnote 40
might be naturally subsumed under this aspect of COMPASâ€™s intended use. 44See also Douglas 2009, 53-54 and Douglas 2016, 610.
22 of 34
Are Algorithms Value-Free?
scientists themselves. This sort of default authority is clearly evident in cases of criminal
justice risk-assessment scores like COMPAS, where analyses indicate that the products of
algorithmic risk assessment programs do in fact sway judgesâ€™ rulings, often resulting in
greater sentencing disparities among individuals from marginalized groups.45 Thus, the
assignment of risk scores will itself come with various inductive risks when the algorithm
is taken as an epistemic authority. Even if we interpret the assignment of risk as a mere
assignment of confidence in or probability of recidivism, that assignment will still arguably
entail inductive risk.
However, defenders of the value-free ideal in science push back: some argue that it is
a conceptual mistake to regard the assignment of probabilities as the adoption of any sort
of hypothesis.46 The assignment of probabilities doesnâ€™t involve any act of acceptance,
but rather merely reflects the degree of confidence we can mathematically hold in some
prediction being true. Thus, thereâ€™s no act to assign risk to, and the mere reflection
of confidence is free from the argument from inductive risk. In the context of actual
scientific practice, these assignments of confidences can be seen as reflecting the actual
confidences scientists have in particular hypotheses, and not an act or decision made on
behalf of the scientist. Opponents of the value-free ideal in turn have a response: they
claim that this interpretation of confidences requires conceptualizing scientists as idealized
Bayesian agents, when in fact they are not. Bayesian models are poor models of the actual
confidences scientists hold, and so in practice scientists do end up doing something akin to
accepting a hypothesis, even when they assign probability estimates.47 However, one might
reasonably argue that this is a point of departure between the domain of scientific inquiry
and machine learning. Although itâ€™s true that individual scientists fall far short of idealized
Bayesian agents, given the computational prowess of machine learning programs, perhaps
their predictive models can more closely approximate such agents. This would suggest that
their assignment of confidences might well be mere reflections of mathematical confidences,
reflections that do not admit of inductive risk.48
Although this is an interesting line of argument, ultimately investigating to what extent
45See, for example, Skeem et al. 2020. See also Logg et al. 2019 and Christin 2017 for relevant discussion. 46Jeâ†µrey 1956, 246. See Steel 2015, 82-83 for discussion. 47Steel 2015, 85-87. 48Although see Parker and Winsberg (2018) for an argument that it is precisely due to this computational
prowess that decisions made earlier in the model-building process quickly compound and exacerbate risks
of error, risks that are eventually inherited in the assignment of confidence measurements. Similarly, Steel
(2015, 85) argues that any complex Bayesian statistical analysis will necessarily involve decisions among
types of probability distribution, and that it is in these decisions that arguments from inductive risk iterate.
23 of 34
Are Algorithms Value-Free?
algorithms can instantiate idealized Bayesian agents without ultimately relying on values
creeping into the models is beyond the scope of this paper.49 I suggest it merely as one
path defenders of the value-free ideal in machine learning might pursue. I instead end by
considering what happens when we grant this point, accepting that we might relocate the
value-free ideal by carving out a minimal role for it in the mere assignment of probabilities.
Notice that, even granting this point for the sake of argument, this still wonâ€™t render most
machine learning programs value-free. This is because most of these programs are used to
automate decisions, not merely in assigning probabilities, but in making predictions about
new cases and applying those predictions in the classification of novel stimuli. Proponents
who adopt this as a potential role for the value-free ideal in machine learning and maintain
that algorithms should remain objective and value-free will not only greatly limit the potential role for algorithms in automated decision-making, but will be forced to accept that
the vast majority of machine learning programs in use today already outstrip the resources
aâ†µorded to strictly value-free processes. Thus, by these standards, not many algorithms
can be value-free, since their very use and function depends on their going beyond what
a purely value-free process could accomplish. Algorithms donâ€™t merely create probabilistic estimates, they use those estimates to make classifications and produce outputs that
make predictions about novel cases. Moreover, note that this whole line of defense for the
value-free ideal leaves unaddressed the arguments against demarcation. If those arguments
are correct, then arguably the design methods that result in the relevant probability assignments will themselves be either justified by or constitutively dependent on value-laden
features of the world.
Where does all this leave us, morally speaking? Admittedly, my goal has not been to
advance any distinctively normative claims, but rather to demonstrate, through the application of feminist insights in philosophy of science to machine learning research, the muchneeded role of moral philosophers in these debates about algorithmic decision-making.
However, some resulting normative prescriptions are clear. Most obviously, if the arguments presented in this paper succeed, then algorithms cannot in principle be value-free.
Since ought implies can, this entails that itâ€™s not the case that algorithms ought be value
free. By foreclosing the normative question of whether algorithms should be value-free, we
are now in a position to usher in the moral evaluation of algorithms.
This endeavor of exploring what normative conclusions can be drawn works best by case
49Cf. footnote 30 for relevant literature in formal decision theory.
24 of 34
Are Algorithms Value-Free?
study. Consider, for example, a concrete attempt made by Johndrow and Lum (2019) to
eliminate unfairness in recidivism risk programs. Their method for eliminating bias relies
on creating statistical parity among the relevant classes, in the context of the recidivism
prediction tool that they analyze, this would mean creating statistical parity among white
and black defendants, producing what they call race-independent recidivism prediction.
What fairness amounts to is â€œdiâ†µerences in the distribution of the modelâ€™s predictions
conditional on the protected variable do not exceed some pre-determined threshold.â€50
Ultimately, the relevant threshold the authors adopt is zero, which entails statistical parity
between the two classes. However, they note that the decision to focus on statistical parity
is motivated by pre-theoretical considerations about the inability to obtain data about
actual crime rates, instead needing to rely on imperfect proxies like rearrest, which are
known to encode systemic biases against black people. Thus, the decision to use statistical
parity in the context of criminal justice is a considered one, which they acknowledge is
open to scrutiny, and that is ultimately up to â€œpolicymakers and ethicistsâ€ to evaluate.51
Regarding import from the discussion of values in this paper, we can see how this decision to leave some vital parameter of the algorithm to be decided by ethical consideration is
licensed by the conclusion of the arguments presented above that algorithms are inherently
value-laden. With little imagination, one can envision an individual who subscribes to the
alternative idea that algorithms can and should be value-free (perhaps one among the 40%
of the population referenced in the first line of this paper) objecting to Johndrow and Lumâ€™s
approach on principle. By building in a placeholder for ethical demands to make a contribution to how the algorithm operates, they might argue, the algorithmic design violates
the normative prescriptions of the value-free ideal. By demonstrating that this ideal is in
principle unattainable, the conclusions of this paper forestall this critique straightforwardly,
and sanction the contributions of explicitly normatively-laden facets of the algorithm itself.
Thus, the general approach of deliberately incorporating explicit value-laden choice points
However, other conclusions of this paper suggest limitations to other aspects of their
general approach. In particular, the basic premise of their strategy can be viewed as
an attempt to equally distribute inductive risk by operationalizing on and eliminating
disparate outcomes for white and black populations. However, itâ€™s not obvious that the
best way to mitigate inductive risk is to adopt a policy of distributing equally across
50Johndrow and Lum 2019, 214. 51Johndrow and Lum 2019, 194, fn. 3.
25 of 34
Are Algorithms Value-Free?
populations. Considering Longinoâ€™s demarcation arguments, one might worry that this
approach taken as a general strategy fails to be sensitive to how the question of which
values one ought adopt will ultimately be a contextual matter.52 One major and recurring
lessons of the paper is that we should stop expecting a universal set of standards that
dictate which values algorithmic decision procedures ought maximize across the board.
Indeed, this is precisely what we see when we reflect on how to expand a commitment
to statistical parity as a strategy for mitigating risks across multiple groups. Inevitably,
such attempts will likely trade oâ†µ from one another, even localizing to a particular context
like criminal justice. Johndrow and Lum (2019)â€™s algorithm creates â€œfairnessâ€ by ensuring statistical parity with respect to race in the outputs. However, even granting that
statistical parity is what we want in the domain of race, this same general strategy does
not straightforwardly apply to other social groups.53 For example, given that overall rates
of criminality are smaller for women than men, we might expect that a commitment to
statistical parity between male and female defendants is misguided.54
Once we recognize that diâ†µerences between social groups matter for the application of
the model, we can anticipate that these complications are only compounded when we consider individual defendantsâ€™ occupying intersectional identities. These diâ†µerent identities
might very well pull in diâ†µerent directions with respect to the need to mitigate inductive
risk, and itâ€™s not clear the intersection of diâ†µerent identities admit of any easy quantitative
analysis. Thus, itâ€™s dicult to predict what eâ†µects Johndrow and Lum (2019)â€™s allegiance
to erasing indicators for race will ultimately have on individuals who occupy multiple
marginalized groups, some of which should aâ†µord them more or less favorable treatment
when calculating inductive risk. More importantly, for these reasons it seems that calculations of inductive risk more generally donâ€™t admit to straightforward operationalization
of any kind, again bolstering the need for experts well-versed in these complex issues at
the intersection of moral philosophy, critical race theory, and gender studies to be able
52To their credit, Johndrow and Lum (2019, 191-194, 214) continually note this limitation of their model. 53I express skepticism about statistical parity in the domain of race equality for two reasons. First,
recall Castro (2019, 417-418)â€™s point that harms from the criminal justice system disproportionately aâ†µect
members of the black community, thus it seems they ought not be given the same weight for the two groups.
Second, I want to allow opportunities for preferential treatment toward historically disadvantaged groups
so as to oâ†µset the accruement of historical injustices. 54COMPAS does not fare much better in this regard. As Angwin et al. (2016)â€™s analysis of COMPAS
notes, â€œSurprisingly, given their lower levels of criminality overall, female defendants were 19.4 percent more
likely to get a higher score than men, controlling for the same factors.â€ For an extensive discussion about
the relationship between gender and compounding wrongs, see Hellman MS.
26 of 34
Are Algorithms Value-Free?
to provide context-specific assessments of algorithmic use. Aspects of these algorithms all
the way down to the very design decisions that produce them are suâ†µuse with normative
implications, and thus, questions of their production, use, and evaluation belong properly
under the purview of ethical theory.
Weâ€™re then left with the question of how best to incorporate these varying perspectives
in ways that allow experts in these intersecting disciplines to contribute to the development
and use of technology. Drawing again on the comparison to scientific practice that has run
throughout the paper, it seems to me that what this situation calls out for is the implementation of institutional mechanisms for ethical oversight for machine learning programs
akin to Institutional Review Boards (IRBs) in science more broadly. As Douglas (2014,
974) notes, â€œan IRB is another mechanism for collectivizing responsibilityâ€ and â€œIRBs were
put in place because scientists as individuals were making bad decisions, and societal anger
was so strong, that IRBs were instituted to provide a check on the practices of scientists.â€
It seems incredible that the explosion of technology has been allowed to grow unchecked
and unfettered to the extent that it now permeates all facets of society without similar
institutional safeguards in eâ†µect. Thus, it seems imperative that some analogous Ethical
Review Board be constructed that can work toward ensuring responsible machine learning and holding practitioners of such technology accountable for the moral consequences
of their machines.55 These are matters of research ethics that moral philosophers and
policy-makers are well-positioned to contribute to.56
Finally, to bring these normative reflections to a close, I want to draw attention to one
final implication of the paper from the constitutive argument for demarcation. According
to that argument, seemingly neutral features of an algorithm imbibe the social and political
values that pervade the environment. The parasitic nature of algorithms on the world entail
that certain morally objectionable qualities of the world will be picked up and perpetuated
in the operation of the machine if left unchecked.
As mentioned at the start of the paper, people are well aware that algorithms can
inherit these moral valences from the world through biases encoded in the data. The
55I would be remiss not to recognize here the tremendous work of grassroots initiatives that are well
underway in attempting to provide such safeguards on the expansion and use of machine learning technology
and big data, programs like the Algorithmic Justice League, the AI Now Institute, Data Feminism, Data for
Black Lives, and FAT*ML. These initiatives make important contributions to the ultimate aim of keeping
developers of social technology in check. However, they are naturally limited due to the lack of institutional
authority that would guarantee compliance. 56See additionally discussion by Shiâ†µrin (2014, 203-206) about the potential threats of IRBs to academic
freedom, again highlighting the importance of moral philosophy to their deployment.
27 of 34
Are Algorithms Value-Free?
arguments of this paper extend these conclusions beyond the material data, to the material
design of the algorithms themselves. Since our current social structures encode problematic
and discriminatory patterns of oppression, programs that attempt to capitalize on those
decision-making procedures that piggyback on the value-laden structure of the world will
inherit problematic features. However, programs that abandon those patterns will fail to
successfully make predictions about our present social environment. Thus, the adoption of
machine learning programs for use in our present social environment necessarily involves an
additional value-laden commitment about whether to inherit or abandon these problematic
social patterns, prompting application of the justification argument against demarcation
Thereâ€™s been a thematic undercurrent running throughout the discussions of this paper:
inductive inferences generally, and machine learning programs specifically, are value-laden
to the extent that they are connected to and dependent on matters that we care about as
human beings. Unlike the detached and abstract logics of deductive inference, inductive
inference depends constitutively for its risk, justification, and very nature on its connections to the world. The risks of getting algorithmic inferences wrong bring real-world
consequences, their justification depends on patterns in the world continuing in familiar
ways into the future, and the principles that govern their ability to traverse inductive gaps
imbibe value structures manifest in society. This connection between value-ladennness as
measured by impact on human endeavors has been noted before. As John Dupre (2007, 40)
puts the point, â€œthe fundamental distinction at work here is that between what matters
to us and what doesnâ€™t.â€ We might, in purely academic abstraction, be able to imagine a
computer algorithm that operated wholly divorced from human endeavors. This algorithm,
whatever it looks like, is far removed from the real-world technology that is the subject of
this paper. So long as machine learning algorithms make important contributions to our
daily lives, and the more we leave certain decision-making that aâ†µects the lives of individuals in the hands of algorithms, we must reject the value-free ideal of algorithmic decision
making. Algorithms function to rank, sort, filter, categorize, assess, label, and draw any
other number of conclusions about real-world phenomenon. They are the useful algorithms
that they are only to the extent that they are undeniably value-laden.
28 of 34
Are Algorithms Value-Free?
Abebe, R., Barocas, S., Kleinberg, J., Levy, K., Raghavan, M., and Robinson, D. G.
(2020). Roles for Computing in Social Change. Conference on Fairness, Accountability,
and Transparency (FAT* â€™20), page 9.
Angwin, J., Larson, J., Mattu, S., and Kirchner, L. (2016). Machine bias: Thereâ€™s software used across the country to predict future criminals. And itâ€™s biased against blacks.
Antony, L. (2001). Quine as Feminist: The Radical Import of Naturalized Epistemology.
In Antony, L. and Witt, C. E., editors, A Mind Of Oneâ€™s Own: Feminist Essays on
Reason and Objectivity, pages 110â€“153. Westview Press.
Antony, L. (2006). The Socialization of Epistemology. In Goodin, R. E. and Tilly, C.,
editors, The Oxford Handbook of Contextual Political Analysis, pages 58â€“77. Oxford
Antony, L. (2016). Bias: Friend or Foe? In Brownstein, M. and Saul, J., editors, Implicit
Bias and Philosophy, Volume 1: Metaphysics and Epistemology, pages 157â€“190. Oxford
Barocas, S. and Selbst, A. D. (2016). Big dataâ€™s disparate impact. California Law Review.
Basu, R. (2018). The Wrongs of Racist Beliefs. Philosophical Studies.
Basu, R. (2019). What we epistemically owe to each other. Philosophical Studies,
Bolinger, R. J. (2018). The rational impermissibility of accepting (some) racial generalizations. Synthese.
Bowker, G. C. and Starr, S. L. (2000). Sorting Things Out: Classification and Its Consequences. MIT Press.
Castro, C. (2019). Whatâ€™s Wrong with Machine Bias. Ergo, an Open Access Journal of
Chouldechova, A. (2016). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. arXiv preprint arXiv:1610.07524.
Christin, A. (2017). Algorithms in practice: Comparing web journalism and criminal
justice. Big Data & Society, 4(2):205395171771885.
Corbett-Davies, S., Goel, S., and Gonzalez-Bailon, S. (2017). Even Imperfect Algorithms
Can Improve the Criminal Justice System. The New York TImes.
29 of 34
Are Algorithms Value-Free?
Domingos, P. (2012). A few useful things to know about machine learning. Communications
of the ACM, 55(10):78.
Domingos, P. (2015). The master algorithm: how the quest for the ultimate learning machine will remake our world. Basic Books, a member of the Perseus Books Group, New
Dotan, R. (2020). Theory choice, non-epistemic values, and machine learning. Synthese.
Douglas, H. (2000). Inductive Risk and Values in Science. Philosophy of Science, 67(4):559â€“
Douglas, H. (2003). The Moral Responsibilities of Scientists (Tensions between Autonomy
and Responsibility). American Philosophical Quarterly, pages 59â€“68.
Douglas, H. (2014). The Moral Terrain of Science. Erkenntnis, 79(S5):961â€“979.
Douglas, H. (2016). Values in Science. In Humphreys, P., editor, Oxford Handbook in the
Philosophy of Science, pages 609â€“630. Oxford University Press.
Douglas, H. E. (2009). Science, policy, and the value-free ideal. University of Pittsburgh
Press, Pittsburgh, Pa. OCLC: ocn297144848.
Dupre, J. (2007). Fact and Value. In Kincaid, H., Dupre, J., and Wylie, A., editors,
Value-Free Science? Ideals and Illusions, pages 27â€“41. Oxford University Press.
Fallis, D. and Lewis, P. J. (2016). The Brier Rule Is not a Good Measure of Epistemic
Utility (and Other Useful Facts about Epistemic Betterness). Australasian Journal of
Fantl, J. and Mcgrath, M. (2007). On Pragmatic Encroachment in Epistemology. Philosophy and Phenomenological Research, 75(3):558â€“589.
Flores, A. W., Bechtel, K., and Lowenkamp, C. T. (2016). False Positives, False Negatives,
and False Analyses: A Rejoinder to Machine Bias: Thereâ€™s Software Used across the
Country to Predict Future Criminals. And Itâ€™s Biased against Blacks. Fed. Probation,
Friedman, B. (1995). Minimizing Bias in Computer Systems. Mosaic of Creativity, page 1.
Friedman, B. (1997). Human Values and The Design of Computer Technology. Cambridge
Friedman, B. and Nissenbaum, H. (1996). Bias in Computer Systems. ACM Transactions
on Information Systems, 14:330â€“347.
30 of 34
Are Algorithms Value-Free?
Gardner, J. (2017). Discrimination: The Good, the Bad, and the Wrongful. Proceedings
of the Aristotelian Society, 118.
Gendler, T. S. (2011). On the epistemic costs of implicit bias. Philosophical Studies,
Giraud-Carrier, C. and Provost, F. (2005). Toward a Justication of Meta-learning: Is the
No Free Lunch Theorem a Show-stopper? Proceedings of the ICML-2005 Workshop on
Meta-learning, page 8.
Gitelman, L., editor (2013). â€œRaw Dataâ€ Is an Oxymoron. MIT Press.
Goodman, N. (1955). Fact, Fiction, and Forecast. Bobbs-Merrill Indianapolis, Indianapolis.
Hellman, D. (2019). Measuring Algorithmic Fairness. Virginia Public Law and Legal Theory
Research Paper, pages 2019â€“39.
Jeâ†µrey, R. C. (1956). Valuation and Acceptance of Scientific Hypotheses. Philosophy of
Johndrow, J. E. and Lum, K. (2019). An algorithm for removing sensitive information:
Application to race-independent recidivism prediction. The Annals of Applied Statistics,
Johnson, D. G. and Nissenbaum, H., editors (1995). Computers, Ethics, and Social Values.
Johnson, G. M. (2020a). Algorithmic bias: on the implicit biases of social technology.
Johnson, G. M. (2020b). The Structure of Bias. Mind, 129(516):1193â€“1236.
Keas, M. N. (2018). Systematizing the theoretical virtues. Synthese, 195(6):2761â€“2793.
Kleinberg, J., Mullainathan, S., and Raghavan, M. (2016). Inherent trade-oâ†µs in the fair
determination of risk scores. arXiv preprint arXiv:1609.05807.
Kroll, J. A. (2018). The fallacy of inscrutability. Philosophical Transactions of the Royal
Society A: Mathematical, Physical and Engineering Sciences, 376(2133):20180084.
Kroll, J. A., Huey, J., Barocas, S., Felten, E. W., Reidenberg, J. R., Robinson, D. G., and
Yu, H. (2017). Accountable Algorithms. University of Pennsylvania Law Review, 165:74.
Kuhn, T. (1962). The Structure of Scientific Revolutions. University of Chicago Press,
31 of 34
Are Algorithms Value-Free?
Kuhn, T. (1977). Objectivity, Value Judgement, and Theory Choice. In The Essential
Tension. University of Chicago Press, Chicago.
Lakatos, I. (1970). Falsification and the Methodology of Scientific Research Programmes.
In Lakatos, I. and Musgrave, A., editors, Criticisms and the Growth of Knowledge, pages
91â€“195. Springer, Dordrecht.
Lauc, D. (2019). How Gruesome are the No-free-lunch Theorems for Machine Learning?
Croation Journal of Philosophy, XVIII(54):8.
Levi, I. (1960). Must the Scientist Make Value Judgments? The Journal of Philosophy,
Logg, J. M., Minson, J. A., and Moore, D. A. (2019). Algorithm appreciation: People
prefer algorithmic to human judgment. Organizational Behavior and Human Decision
Longino, H. E. (1995). Gender, politics, and the theoretical virtues. Synthese, 104(3):383â€“
Longino, H. E. (1996). Cognitive and Non-cognitive Values in Science: Rethinking the
Dichotomy. In Hankinson Nelson, L. and Nelson, J., editors, Feminism, science, and the
philosophy of science, pages 39â€“58. Kluwer, Dordrecht. OCLC: 801321444.
Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., and Galstyan, A. (2020). A Survey
on Bias and Fairness in Machine Learning. arXiv:1908.09635, page 32.
Moss, S. (2018). Probabilistic Knowledge. Oxford University Press, Oxford.
Mulligan, D. K., Kroll, J. A., Kohli, N., and Wong, R. Y. (2019). This Thing Called
Fairness: Disciplinary Confusion Realizing a Value in Technology. Proceedings of the
ACM on Human-Computer Interaction, 3(CSCW):1â€“36. arXiv: 1909.11869.
Munton, J. (2017). The Eyeâ€™s Mind: Perceptual Process and Epistemic Norms. Philosophical Perspectives, 31(1):317â€“347.
Munton, J. (2019a). Beyond accuracy: Epistemic flaws with statistical generalizations.
Philosophical Issues, 29(1):228â€“240.
Munton, J. (2019b). Perceptual Skill And Social Structure. Philosophy and Phenomenological Research.
Nissenbaum, H. (1996). Accountability in a Computerized Society. Science and Engineering
Northpointe, I. (2015). Practitionerâ€™s Guide to COMPAS Core.
32 of 34
Are Algorithms Value-Free?
Norton, J. D. (2003). A Material Theory of Induction. Philosophy of Science, 70(4):647â€“
Parker, W. S. and Winsberg, E. (2018). Values and evidence: how models make a diâ†µerence.
European Journal for Philosophy of Science, 8(1):125â€“142.
Ramsey, F. P. (1989). Mr Keynes on Probability. The British Journal for the Philosophy
of Science, 40(2):219â€“222.
Rooney, P. (1992). On Values in Science: Is the Epistemic/Non-Epistemic Distinction Useful? PSA: Proceedings of the Biennial Meeting of the Philosophy of Science Association,
Rudner, R. (1953). The Scientist Qua Scientist Makes Value Judgments. Philosophy of
Selbst, A. D., Boyd, D., Friedler, S. A., Venkatasubramanian, S., and Vertesi, J. (2019).
Fairness and Abstraction in Sociotechnical Systems. In Proceedings of the Conference
on Fairness, Accountability, and Transparency – FAT* â€™19, pages 59â€“68, Atlanta, GA,
USA. ACM Press.
Shiâ†µrin, S. V. (2014). Speech matters: on lying, morality, and the law. Carl G. Hempel
lecture series. Princeton University Press, Princeton.
Shmueli, G. (2010). To Explain or to Predict? Statistical Science, 25(3):289â€“310.
Skeem, J., Scurich, N., and Monahan, J. (2020). Impact of risk assessment on judges
fairness in sentencing relatively poor defendants. Law and Human Behavior.
Smith, A. (2018). Public Attitudes Toward Compuer Algorithms. Pew Research Center,
Stanley, J. (2005). Knowledge and Practical Interest. Oxford University Press, Oxford.
Steel, D. (2015). Acceptance, values, and probability. Studies in History and Philosophy
of Science Part A, 53:81â€“88.
Tavernise, S. (2013). Drug Agency Recommends Lower Dosage of Sleep Aids for Women.
The New York Times.
Titelbaum, M. G. (2010). Not Enough There There: Evidence, Reasons, and Language
Independence. Philosophical Perspectives, 24(1):477â€“528.
Vidal, A. C., Smith, J. S., Valea, F., Bentley, R., Gradison, M., Yarnall, K. S. H., Ford,
A., Overcash, F., Grant, K., Murphy, S. K., and Hoyo, C. (2014). HPV genotypes
and cervical intraepithelial neoplasia in a multiethnic cohort in the southeastern USA.
Cancer Causes & Control, 25(8):1055â€“1062.
33 of 34
Are Algorithms Value-Free?
Viens, L. J., Henley, S. J., and Watson, M. (2016). Human Papillomavirusâ€”Associated
Cancersâ€”United States, 2008-20112. MMWR Morb Mortal Wkly Rep, (65):661â€“666.
Weisberg, M. (2007). Three Kinds of Idealization:. Journal of Philosophy, 104(12):639â€“659.
Wolpert, D. H. (2013). Ubiquity symposium: Evolutionary computation and the processes
of life: What the no free lunch theorems really mean: How to imporve search algorithms.
Zabell, S. (2011). Carnap and the Logic of Inductive Inference. In Handbook of the History
of Logic, volume 10, pages 265â€“309. Elsevier.
34 of 34
Get Professional Assignment Help Cheaply
Are you busy and do not have time to handle your assignment? Are you scared that your paper will not make the grade? Do you have responsibilities that may hinder you from turning in your assignment on time? Are you tired and can barely handle your assignment? Are your grades inconsistent?
Whichever your reason is, it is valid! You can get professional academic help from our service at affordable rates. We have a team of professional academic writers who can handle all your assignments.
Why Choose Our Academic Writing Service?
- Plagiarism free papers
- Timely delivery
- Any deadline
- Skilled, Experienced Native English Writers
- Subject-relevant academic writer
- Adherence to paper instructions
- Ability to tackle bulk assignments
- Reasonable prices
- 24/7 Customer Support
- Get superb grades consistently
Online Academic Help With Different Subjects
Students barely have time to read. We got you! Have your literature essay or book review written without having the hassle of reading the book. You can get your literature paper custom-written for you by our literature specialists.
Do you struggle with finance? No need to torture yourself if finance is not your cup of tea. You can order your finance paper from our academic writing service and get 100% original work from competent finance experts.
While psychology may be an interesting subject, you may lack sufficient time to handle your assignments. Don’t despair; by using our academic writing service, you can be assured of perfect grades. Moreover, your grades will be consistent.
Engineering is quite a demanding subject. Students face a lot of pressure and barely have enough time to do what they love to do. Our academic writing service got you covered! Our engineering specialists follow the paper instructions and ensure timely delivery of the paper.
In the nursing course, you may have difficulties with literature reviews, annotated bibliographies, critical essays, and other assignments. Our nursing assignment writers will offer you professional nursing paper help at low prices.
Truth be told, sociology papers can be quite exhausting. Our academic writing service relieves you of fatigue, pressure, and stress. You can relax and have peace of mind as our academic writers handle your sociology assignment.
We take pride in having some of the best business writers in the industry. Our business writers have a lot of experience in the field. They are reliable, and you can be assured of a high-grade paper. They are able to handle business papers of any subject, length, deadline, and difficulty!
We boast of having some of the most experienced statistics experts in the industry. Our statistics experts have diverse skills, expertise, and knowledge to handle any kind of assignment. They have access to all kinds of software to get your assignment done.
Writing a law essay may prove to be an insurmountable obstacle, especially when you need to know the peculiarities of the legislative framework. Take advantage of our top-notch law specialists and get superb grades and 100% satisfaction.
What discipline/subjects do you deal in?
We have highlighted some of the most popular subjects we handle above. Those are just a tip of the iceberg. We deal in all academic disciplines since our writers are as diverse. They have been drawn from across all disciplines, and orders are assigned to those writers believed to be the best in the field. In a nutshell, there is no task we cannot handle; all you need to do is place your order with us. As long as your instructions are clear, just trust we shall deliver irrespective of the discipline.
Are your writers competent enough to handle my paper?
Our essay writers are graduates with bachelor's, masters, Ph.D., and doctorate degrees in various subjects. The minimum requirement to be an essay writer with our essay writing service is to have a college degree. All our academic writers have a minimum of two years of academic writing. We have a stringent recruitment process to ensure that we get only the most competent essay writers in the industry. We also ensure that the writers are handsomely compensated for their value. The majority of our writers are native English speakers. As such, the fluency of language and grammar is impeccable.
What if I don’t like the paper?
There is a very low likelihood that you won’t like the paper.
- When assigning your order, we match the paper’s discipline with the writer’s field/specialization. Since all our writers are graduates, we match the paper’s subject with the field the writer studied. For instance, if it’s a nursing paper, only a nursing graduate and writer will handle it. Furthermore, all our writers have academic writing experience and top-notch research skills.
- We have a quality assurance that reviews the paper before it gets to you. As such, we ensure that you get a paper that meets the required standard and will most definitely make the grade.
In the event that you don’t like your paper:
- The writer will revise the paper up to your pleasing. You have unlimited revisions. You simply need to highlight what specifically you don’t like about the paper, and the writer will make the amendments. The paper will be revised until you are satisfied. Revisions are free of charge
- We will have a different writer write the paper from scratch.
- Last resort, if the above does not work, we will refund your money.
Will the professor find out I didn’t write the paper myself?
Not at all. All papers are written from scratch. There is no way your tutor or instructor will realize that you did not write the paper yourself. In fact, we recommend using our assignment help services for consistent results.
What if the paper is plagiarized?
We check all papers for plagiarism before we submit them. We use powerful plagiarism checking software such as SafeAssign, LopesWrite, and Turnitin. We also upload the plagiarism report so that you can review it. We understand that plagiarism is academic suicide. We would not take the risk of submitting plagiarized work and jeopardize your academic journey. Furthermore, we do not sell or use prewritten papers, and each paper is written from scratch.
When will I get my paper?
You determine when you get the paper by setting the deadline when placing the order. All papers are delivered within the deadline. We are well aware that we operate in a time-sensitive industry. As such, we have laid out strategies to ensure that the client receives the paper on time and they never miss the deadline. We understand that papers that are submitted late have some points deducted. We do not want you to miss any points due to late submission. We work on beating deadlines by huge margins in order to ensure that you have ample time to review the paper before you submit it.
Will anyone find out that I used your services?
We have a privacy and confidentiality policy that guides our work. We NEVER share any customer information with third parties. Noone will ever know that you used our assignment help services. It’s only between you and us. We are bound by our policies to protect the customer’s identity and information. All your information, such as your names, phone number, email, order information, and so on, are protected. We have robust security systems that ensure that your data is protected. Hacking our systems is close to impossible, and it has never happened.
How our Assignment Help Service Works
1. Place an order
You fill all the paper instructions in the order form. Make sure you include all the helpful materials so that our academic writers can deliver the perfect paper. It will also help to eliminate unnecessary revisions.
2. Pay for the order
Proceed to pay for the paper so that it can be assigned to one of our expert academic writers. The paper subject is matched with the writer’s area of specialization.
3. Track the progress
You communicate with the writer and know about the progress of the paper. The client can ask the writer for drafts of the paper. The client can upload extra material and include additional instructions from the lecturer. Receive a paper.
4. Download the paper
The paper is sent to your email and uploaded to your personal account. You also get a plagiarism report attached to your paper.