An algorithm embeds choices about what counts, what is normal, and what deserves to be optimized. Guillaume Chevillon draws out the consequence of this observation: if algorithmic design produces norms, then changing the design changes the norms. Queer Algorithms, published on April 3, 2026 by B42 Editions, proposes making this the program for a different kind of engineering. The essay appears at a moment when the European AI regulation imposes transparency on high-risk automated systems, without yet guaranteeing the right to contest the categories they use.

The Essentials

  • Algorithms construct a version of reality that conforms to the categories of their design and marginalize what does not fit into them.
  • Chevillon proposes treating error, minority experiences, and unforeseen uses as design criteria rather than as anomalies to be eliminated.
  • The thesis aligns with Adam Becker’s critique of Silicon Valley techno-utopianism: algorithmic objectivity conceals normative choices that concentrate the power to define which lives are desirable.
  • The European AI regulation (2024) opens a right to explanation, but does not yet guarantee the right to contest the categories themselves.
  • The essay then examines what institutions enable the correction of automated systems without abandoning their useful applications.

The Author

Guillaume Chevillon is a researcher in statistics and quantitative methods. His academic work focuses on forecasting models, their robustness, and their limits. He writes from within the discipline, understands the technical constraints of algorithmic design, and examines its choices without denouncing technology itself.

The Thesis: Normalization as a Product of Design

Chevillon’s central argument can be stated in one sentence: an algorithm always optimizes toward something, and that something is something someone has defined. When a credit system evaluates solvency, when a recruitment tool ranks applications, when a recommendation algorithm orders content, it does not measure a pre-existing reality. It produces a version of reality filtered through design criteria. These criteria carry assumptions about what is normal, desirable, acceptable.

The word “queer” in the title deserves particular attention. Chevillon does not use it to confine his argument to sexualities or gender identities. He employs it in its broader etymological and theoretical sense: that which deviates, that which resists categorization, that which disrupts taxonomies. A disabled person whose life trajectory does not follow the linear path expected by an insurance algorithm is “queer” in Chevillon’s sense. An unforeseen use of an application, an atypical career path, a non-standard family situation: all these real cases become anomalies to correct or errors to eliminate in current systems.

The essay’s proposal inverts this logic. Making error a design criterion means designing systems capable of treating the exception as useful information rather than as noise to be suppressed. Making unforeseen uses a design signal means recognizing that users know things designers do not. This inversion is less straightforward than it appears: it requires renouncing part of measurable efficiency to gain real robustness.

Objectivity as Ideology

The sharpest part of the essay attacks what Chevillon calls the myth of algorithmic objectivity. Automated systems are regularly presented, by their designers and users, as neutral tools that process data without prejudice. This presentation is either an error or a strategy.

Adam Becker, in his work on Silicon Valley techno-utopianism, has shown that the ideology of universal technological progress regularly serves to mask power choices. Algorithmic systems embody this dynamic at an operational scale. They decide who gets a loan, who passes a job interview, who receives social benefits. These decisions are not neutral. They reflect the statistical correlations present in training data, and this data is produced by societies that have their own inequalities, their own norms, their own blind spots.

Chevillon goes beyond denouncing bias. The problem of algorithmic bias is well documented: studies have shown, for example, that facial recognition systems display significantly higher error rates on the faces of Black women than on the faces of white men. Correcting a bias amounts to working within the same optimization logic: one improves the system’s accuracy without questioning what it optimizes or for whom. Queer Algorithms goes upstream to ask who decides the categories.

This point intersects a tension found in debates over AI regulation. The European AI regulation (2024) subjects high-risk systems in recruitment, credit, social benefits, and justice to transparency requirements and provides a conditional right to explanation for certain unfavorable individual decisions. This is real progress. But the right to explanation is not the right to contest the category itself. One can explain why an algorithm refused a loan without ever questioning the relevance of the variables it uses to define solvency.

Error as a Source of Learning

The second pillar of the essay is a rehabilitation of error. In classical algorithmic design, error is what one wants to get rid of. Models are trained by minimizing a loss function. Their performance is measured by their error rate. They are deployed when this rate falls below an acceptable threshold.

Error is the problem.

Chevillon reverses the argument. An algorithm’s error is often the moment when reality resists the model: the model then exposes its limits, its blind spots, and the categories it cannot handle. Eliminating error amounts to eliminating this signal. Preserving error as design information provides the means to understand what the model does not see.

This position has precedents in the history of science. According to Thomas Kuhn, anomalies announce paradigm shifts. A model confronted with its own limits learns faster than a model that produces only confirmations of its hypotheses. Chevillon applies this intuition to automated decision systems.

It does so all the more because empirical counter-examples are beginning to emerge. Studies conducted on alternative lending systems, particularly fintechs that use unconventional data to evaluate solvency in countries without robust credit histories, suggest that treating atypical profiles as sources of information rather than as anomalies to discard can improve risk prediction in certain contexts. The “default normalization” of large credit models may leave aside creditworthy borrowers who do not resemble the historical average.

There is a real tension, which the essay does not entirely conceal. Making error a design criterion has a cost. Systems that tolerate more uncertainty are often slower, more complex, harder to audit. For low-stakes applications, this trade-off is easy to accept. For real-time automated decision systems at scale, emergency resource allocation, fraud detection, infrastructure security, the trade-off is less obvious.

Chevillon responds by distinguishing contexts, but the response remains partial.

Current Limitations of Regulation

The tension between Chevillon’s critique and the tools available to address it is perhaps the most interesting place in the essay, and where it calls most directly for collective reflection.

The European AI regulation classifies systems according to their level of risk. High-risk systems are subject to documentation, event logging, and transparency requirements toward deployers; individual explanation is provided in limited cases. This is a serious framework, among the most ambitious in the world. But the regulation addresses existing systems, with their existing categories. It improves transparency about how an algorithm decides.

It does not require revisiting the categories from which it decides.

On this point, Jean Tirole’s position—regulation as market design, not administrative control—offers a useful complementary reading. If regulation sets expected outcomes (which populations should be able to access credit, employment, benefits) rather than only processes (documentation, explainability), it creates pressure toward the different design Chevillon calls for. Defining an expected outcome forces designers to find the categories that make it possible.

This approach is not without risks. It presumes the ability to measure outcomes reliably, and to have institutions capable of evaluating them independently. Two conditions that large-scale algorithmic systems make precisely difficult. As analysis of collective bargaining and workers’ right of oversight on AI at work demonstrates, Europe is progressing on the procedural side, but the substantive side, that of the categories themselves, remains largely open.

Institutions Capable of Correcting Categories

Chevillon raises the question of institutions without resolving it. It remains to be determined who decides that an algorithm’s categories are unjust, who has the legitimacy and technical means to contest them, and who arbitrates between measurable efficiency and robustness to minority situations.

Looking toward 2030-2040, three configurations seem plausible.

In the first, regulatory obligations deepen progressively. The right to explanation evolves into a right to contest categories, supported by supervisory authorities endowed with independent technical expertise. Organizations like the European Securities and Markets Authority (ESMA) or the French Data Protection Authority (CNIL) could develop substantive algorithmic audit capacities. This scenario assumes sustained political will and specialized human resources that Europe still struggles to build.

In the second, pressure comes more from the market. Actors who design systems more robust to atypical profiles capture markets that normalizing systems leave aside, exactly as alternative credit fintechs do. Competition on design quality, rather than on average efficiency alone, becomes a competitive advantage. This scenario is more realistic in sectors where poorly served populations represent significant economic volume. It is less likely for populations whose statistical marginality does not constitute an attractive market.

In the third, organized users themselves apply the pressure. Collectives, associations, unions document the effects of algorithmic categories on their members and bring legal or media challenges. This bottom-up mode of regulation has historically produced results in other domains—consumer rights, public building accessibility, food labeling requirements. It assumes minimal technical capacity to understand and document what happens inside systems, capacity that the transparency imposed by European regulation is beginning to make possible. The parallel with the issue of local content visibility on platforms is instructive: regulation alone is not enough when technical architectures are designed to optimize toward other objectives.

These three configurations are not mutually exclusive. Signals to watch to understand which advances: the evolution of European supervisory authorities’ capacity for technical audit, the first cases of precedent on the right to contest an algorithmic category, and the emergence or non-emergence of significant markets for systems designed according to Chevillon’s logic.

The Blind Spots

The essay has its limits, and they must be named. The first is the question of scalability. Chevillon demonstrates that systems designed to handle exception are possible, and that they would be more equitable. He shows less clearly how to deploy them at scale without losing the properties that make them desirable.

The second is political. The logic of algorithmic optimization toward normality is not a naive design error. It responds to real interests: reducing the cost of handling complex cases, minimizing responsibility by taking refuge behind automated procedure, concentrating services where profitability is highest. Contesting categories means contesting business models. The essay describes what a different engineering would make possible; it says less about why actors in a position to design these systems would have an interest in doing so.

The third is an angle of analysis the essay does not explore: the possibility that certain algorithms produce useful normalization. Standardization is not always violence. Standardized medical protocols have saved lives precisely by resisting the individual preferences of practitioners. Distinguishing domains where algorithmic normalization is acceptable from those where it is problematic is work that remains to be done, and which Chevillon leaves open.

The Essay’s Significance

Queer Algorithms addresses three distinct audiences. To algorithm designers, it offers a framework for thinking through the normative externalities of their technical choices. To public decision-makers, it proposes language for formulating design requirements that go beyond procedural transparency. To citizens and organizations living with the effects of these systems, it provides conceptual tools to name what is happening to them.

What one finds in this essay that one does not often read elsewhere: an argument that does not merely denounce algorithmic bias, but proposes an alternative design logic, from within the statistical discipline itself. It is precisely this positioning that gives it its strength. Chevillon does not say algorithms are bad. He says algorithms can be designed differently, and he explains how.

At a time when AI enters lecture halls and companies before institutions have had time to adapt, this kind of argument—technical without being inaccessible, critical without being technophobic—deserves to be widely read.

Bibliographic Information Queer Algorithms: Perturbing Technologies, Imagining Our Futures Guillaume Chevillon B42 Editions, April 3, 2026

Sources

  1. Bibliothèque nationale de France, Nouveautés Éditeurs, Algorithmes queers by Guillaume Chevillon: https://nouveautes-editeurs.bnf.fr/accueil?id_declaration=10000001304211&titre_livre=Algorithmes_Queers
  2. Adam Becker, More Everything Forever: AI Overlords, Space Empires, and Silicon Valley’s Crusade to Control the Fate of Humanity, Scientific American author page: https://www.scientificamerican.com/author/adam-becker/
  3. European AI Regulation (AI Act), European Parliament and Council of the European Union, 2024, Official Journal of the EU