By Javier Surasky
A Spanish version will be published on Thursday
.png)
Introduction
Artificial
intelligence (AI) is often presented as a tool that enables faster
decision-making, reduces errors, and ensures greater consistency, all under the
premise of objectivity rooted in the belief that the system minimizes human
arbitrariness. However, such objectivity is far from automatic.
For it to
exist, a necessary condition would be that the world could be represented
through a set of data that constitutes its perfect representation, which is impossible: As Shin
warns, algorithmic truth is a construction based on
datasets, classifications, heuristics, and statistical models, and then cannot be neutral (Shin, 2025).
This signals an underlying political shift, in which dominant criteria of justification and performance become institutionalized in ways that are not always apparent, and, consequently, the discourse often centers on metric compliance rather than the underlying purposes for which AI is employed, which may
diverge significantly.
We contend
that AI functions as an infrastructure of instrumental reason by converting
normative ends into optimization variables, and this formalization
does tends to
shift the focus from value-based discussions to metric-based evaluations.
Invisible Decisions in AI Models
Let us begin by clarifying the problem of AI and value complexes.
AI cannot “have values of its own,” since it is not a moral subject: In its operation, AI becomes dependent on prior axiological decisions regarding matters such as what to measure, with which indicators, and which losses to tolerate, decisions that appear to be technical when, in fact, they are political, entailing value-based interpretations. In other words, AI is “a tool, a means that lacks its own ends” (Surasky, 2023:269). This relocation of the values at stake is often rendered invisible by the technocratic narrative.
However, if AI is an assemblage of data, computing power, and human skills, then it should be read as a social practice that reorganizes rationality by establishing what counts as a problem, what counts as a solution, and which sacrifices are acceptable in the name of performance.
Lindgren
(2024:17) moves in the same direction when he asks us to understand AI “not as
a narrow technology, but as a ubiquitous apparatus, a complex agglomeration of
different components, which is entangled with human experience,” reinforcing
the view that the conversion of ends into metrics is a way of organizing experiences under conditions of control.
Instrumental Reason and Rationalization: A Critical Theory Framework
Critical theory long ago gave a name to the mechanism by which means and procedures gain primacy over ends. Horkheimer, for example, in distinguishing “subjective” reason, defines it by its orientation toward instrumental adequacy: “It is essentially concerned with means and ends, with the adequacy of procedures for purposes more or less taken for granted” (Horkheimer, 2004:3).
What is crucial here is the reference to “purposes taken for granted,” which are presupposed, inherited, and naturalized, ceding to the analysis of procedural performance the space that should be occupied by debate over values. AI intensifies this pattern by optimizing procedures under an objective function.
The politically problematic effect emerges when that function becomes the dominant language of justification.
What is
justifiable is measured using internal optimization metrics such as accuracy, error minimization, and efficiency, along with
external metrics such as savings, productivity, and time reduction, and all of
this requires defining in advance what counts as “the good” in operational
terms and formalizing it so as to guide all subsequent decisions, thereby
constructing an algorithmic rationality that stabilizes decisions around what
can be measured and optimized.
Dialectic
of Enlightenment offers a useful antecedent: when the world becomes legible as
calculation, narrative, and interpretation lose authority in the face of formal
operation. “The scientific calculation of events annuls the account of them
which thought had once given in myth” (Horkheimer and Adorno, 2002:5). This can
be understood as a structural analogy of how the public explanation of ends can
be displaced when calculation dominates the language of rationality and
optimization presents itself as a sufficient substitute for the justification
of ends.
We are not claiming that all rationalization or formalization constitutes domination, the issue is much more specific: when the rationality of means is imposed as the guiding criterion, consideration of ends is subordinated or externalized, and AI has the potential to drive such operational subordination when integrated into performance evaluation frameworks: it automates and stabilizes the institutional preference for what is calculable.
What can be quantified becomes what is prioritized in decision-making and in the explanation of decisions.
The Axiological Operation of Measurement
Before the model comes measurement, and before measurement comes the definition of what is measurable, what Espeland and Stevens (2008:407) call “the production and communication of numbers,” which fabricates numbers in a social nature through categories, instruments, thresholds, scales, and conventions. For this reason, algorithms and neutrality do not necessarily go hand in hand.
Douglas (2009:14), in her critique, tells us that “the value-free ideal for science, articulated by philosophers in the late 1950s and cemented in the 1960s, should be rejected, not just because it is a difficult ideal to attain, but because it is an undesirable ideal.”
In
response, one might argue that “AI can be neutral if the data are good,” but
that claim overlooks the persistence of a blind spot: data will be “better” or
“worse” depending on what has been defined as relevant within a value framework
that represents reality in a partial and biased way.
Measuring implies ordering and classifying, so categories are not merely names but “gatekeepers” to spaces of visibility: Data that manage to pass through become measurement, capture, and standardization, determining whether experiences are allowed into the decision-making circuit.
Once the measurement
framework is fixed, what is measurable becomes data, and what is not measurable
becomes residue.
Data: An Operation of Capturing the World
AI learns patterns based on what the system can record, giving its “reasoning” the form of an inference conditioned by the representation of the world that feeds it.
Before machine learning, there was an operation of transmuting lived experience
into standardized input. “In the contemporary era, it is human life, extracted
in the form of data, that is being appropriated,” Mejias and Couldry remind us
(2024:18), while O’Neil (2016:21) reaffirms that “A model’s blind spots reflect
the judgments and priorities of its creators.”
There are at least three types of blind spots: absence (no available record), normative decision (what fails to pass a cost-efficiency analysis is excluded as irrelevant), and proxy (a substitute is measured, reordering the phenomenon to make it manageable).
Blind spots are products of
an institutional construction of what is considered relevant. In AI, blind spots become structural because optimization requires a stable and measurable criterion for success.
Some may
argue that biases can be corrected technically, bur our view is that many
corrective proposals are nothing more than translations of values such as
equity, non-discrimination, or justice into metric constraints.
Implementing such formalization is necessary, even indispensable, but it requires that every value introduced as a corrective be explicitly defined.
Rather than resolving
underlying tensions, this produces a convergence of multiple formalizations,
each embodying the issues previously highlighted, and resulting in what might
be termed “valuewashing”: “the principle-based approach seemed toothless,
devoid of normative ends” (Krijger, 2022:1429), a process obscuring normative-axiological conflicts, activating what Mökander and Schroeder
(2024:1359) describe as “the replacement of traditions with instrumental
rationality [as] the most calculable and efficient way of achieving any given
policy objective.” Thus, even when the stated aim is commendable, the instrumental
logic tends to subsume it within what can be calculated.
Social Effects: When the Indicator Rules
When metrics become institutionalized, they return to reality, transformed into criteria. Bowker and Star offer a clear example: “The decision to classify students by their standardized achievement and aptitude tests valorizes some kinds of knowledge skills and renders other kinds invisible” (Bowker and Star, 1999:6).
With AI, that effect is amplified because classification is automated
and deployed as a supposedly objective allocation infrastructure.
Here we meet what sociology and philosophy term the “reactivity” or “performativity” of public measures: the act of recording a state of affairs transforms data into indicators that prompt adaptations, which subsequently alter the phenomena being measured, because once publicized, data reorganize behaviors and priorities, acting as a binary-code version of Schrödinger’s cat.
To see this
mechanism in operation, we can imagine a credit eligibility scoring system in
which a complex judgment crystallizes into a defined threshold that becomes the
practical objective to be achieved: improving the score replaces debate over
its rationale.
This reality is articulated with a broad regime of extracting life in the form of data, which Mejias and Couldry (2024:17) describe as a process of systematic appropriation: “Welcome to data colonialism. It is happening everywhere. It is an appropriation of resources on a truly colonial scale. A data grab that will change the course of history as the original colonial land grab did five centuries ago.”
This is about highlighting how metric infrastructure enables life to be treated as
a standardized, exploitable territory or, as Lindgren (2023:17) puts it, “The
danger of ideology is that once the dominant views and priorities have been
established, they begin to disguise themselves as ‘the common sense’, thereby
becoming naturalized.”
A typical
consequence of indicator naturalization is that improvements in measured
results are celebrated as substantive progress, even as unmeasured aspects, such
as dignity, autonomy, meaning, or justice, may deteriorate.
From Indicator to Governance: Why Measuring Is Not the Same as Deciding
We arrive
at what I call the “algorithmic alibi,” an effect whereby prior normative
decisions are justified by the appearance of technical necessity: “the model
says so,” “the indicator requires it,” “it’s the most efficient option.”
This alibi emerges from the institutionalization of formalization in pursuit of politically determined objectives: as metrics become the accepted language, debates about ends are reduced to mere compliance with indicators.
It looks like an update of an old mechanism: “Under the rule by law, the law is merely a
formal tool that legitimizes executive action” (Smuha, 2024:12), and now the
algorithmic alibi follows that way, presenting contestable decisions as
incontrovertible requirements.
A defender
of quantification might say that human deliberation is also biased and
opaque, and that indicators at least allow for auditing. Point well taken:
formalization facilitates control, but if auditing merely verifies compliance
with metrics, it certifies performance within an already axiologized framework
and helps keep it “under the rug.”
Moreover,
under a rationality of measurement, when objectives prove to be
“non-measurable,” they lose real existence, regardless of how laborious and
valuable the processes that established them may have been: “indicators that
capture globally agreed (but contested) policy goals can silently disappear due
to a lack of data and measurability” (Bexell, 2024:278).
And just as
we concede the auditing advantage that indicators provide, we ask in return
that it be acknowledged that when indicators become institutional performance
criteria linked to audits, reputation, funding, or sanctions (Morgan-Thomas et
al., 2024), they can induce forms of data recording that push the reality they
aim to reflect to its limits in order to improve what is deemed “successful”
(Runhardt, 2025).
Conclusions: What Should Be Done with the Non-Measurable?
The possibility of mediations that sustain non-reducible ends without abandoning all measurement is a critical element for the future of AI governance, calling on us to recover a distinction between rationalities.
Habermas tells us that,
in speaking, we formulate validity claims: “we are constantly making claims,
even if usually only implicitly, concerning the validity of what we are saying”
(Habermas, 1984: p.x). What matters is that these claims can be discussed and
must be publicly justified, opening a deliberative dimension that optimization
seeks to occupy in its entirety.
Someone can say that if we
don’t measure, we can’t act based on evidence. That is true, and far from
demonizing measurement our aim is to show that action requires combining it with responsible judgment, including the definition of where and when
optimization is legitimate and where and when it should be limited by
principles that are not reducible to metrics.
I propose a simple test to decide where optimization should be limited, based on five questions:
1. Does the decision have irreversible or difficult-to-repair impacts?
2. Is there a high asymmetry of power and a low capacity for exit on the part of those affected?
3. Are the harms compensable?
4. Do the harms systematically fall on certain groups?
5. Does the proxy negatively affect, or potentially negatively affect, human rights?
When at
least one answer is affirmative, I suggest inverting the burden of proof as a
prudential criterion: it is not enough for the model to “work”, it must be
publicly justified why its measurement framework is acceptable.
If technology is not neutral, then its design and implementation must be opened to public dispute over what is optimized, by which criteria, and under which readings of the values involved. The selection of metrics must be repoliticized and democratized, wich requires generating basic statistical literacy programs so that all people can participate in these debates.
AI, as an infrastructure of instrumental reason, shifts debates over ends and values toward efficiency metrics.
We invite readers to distinguish between the two and
to problematize the conceptualization of indicators as objective measurements
of reality, because what is valuable is never exhausted by what is measurable.
References
Bexell, M.
(2024). Indicator accountability or policy shrinking? The effects of indicators
on multilateral development banks. Global Policy, (15)2, 276-287. https://onlinelibrary.wiley.com/doi/epdf/10.1111/1758-5899.13348
Bowker, G.C. y Star, S.L. (1999). Sorting things out: Classification and its
consequences. MIT Press.
Douglas,
H.E. (2009). Science, policy, and the value-free ideal. University of
Pittsburgh Press.
Espeland,
W. y Stevens, M. (2008). A sociology of quantification. Archives Européennes de Sociologie /
European Journal of Sociology, 49(3), 401-436. Doi 10.1017/S0003975609000150
Habermas, J.
(1984). The theory
of communicative action (Vol. 1, Reason and the rationalization of society; T. McCarthy,
Trans.). Beacon Press.
Horkheimer,
M. (2004). Eclipse of reason. Continuum.
Horkheimer,
M. y Adorno, T. (2002). Dialectic of enlightenment: Philosophical fragments.
Stanford University Press.
Krijger, J.
(2022). Enter the metrics: Critical theory and organizational
operationalization of AI ethics. AI & Society, (37), 1427-1437. https://doi.org/10.1007/s00146-021-01256-3
Lindgren,
S. (Ed.). (2023). The handbook of critical studies of artificial
intelligence. Edward Elgar Publishing.
Lindgren,
S. (2024). Critical theory of AI. Polity Press.
Mejias, U.
A., & Couldry, N. (2024). Data grab: The new colonialism of big tech and
how to fight back. The University of Chicago Press.
Mökander,
J., & Schroeder, R. (2024). Artificial intelligence, rationalization, and
the limits of control in the public sector: The case of tax policy
optimization. Social Science Computer Review, 42(6), 1359-1378. https://doi.org/10.1177/08944393241235175
Morgan-Thomas,
A.; Tsoukas, S.; Dudau, A. y Gąska, P. (2024). Beyond declarations: Metrics,
rankings and responsible assessment. Research Policy, 53(10), 105093. https://doi.org/10.1016/j.respol.2024.105093
O’Neil, C.
(2016). Weapons of math destruction: How big data increases inequality and
threatens democracy. Crown.
Runhardt,
R. (2025). Reactivity, validity, and measurement in the social sciences. Studies
in History and Philosophy of Science. https://doi.org/10.1016/j.shpsa.2025.05.006
Shin, D.
(2025). Automating epistemology: how AI reconfigures truth, authority, and
verification. AI & Society. https://doi.org/10.1007/s00146-025-02560-y
Smuha, N.
(2024). Algorithmic Rule by Law. Cambridge University Press. https://doi.org/10.1017/9781009427500
Surasky, J. (2023). Gobernanza de la inteligencia artificial
para el desarrollo sostenible. Salimena, G. (Comp.). La política
internacional en el proceso de transición intersistémico. Teseo, 269-292.