The Cowardice of Inference

“The three gibbering, fumbling creatures, with their enlarged heads and wastedbodies, were contemplating the future.” “The Minority Report” byPhilip K. Dick

People who build predictive systems often describe their work as analysis: discovering patterns, improving accuracy, reducingerror. The language suggests distance from the consequences—as though the model is simply revealing how the world works.

But predictive systems don’t just find patterns. They establish them.(note 1)

And that is where the responsibilitylies.

• When a model denies someone a loan, it’s not uncovering a natural truth about risk. It’s enforcing arule that treats certain people as risky.
• When a fraud detector flags someone, it’s not identifying an objectivecategory. It’s operationalizing institutional judgments about which behaviors deserve scrutiny.
• When anengagement model shapes what people see, it’s not exposing their preferences. It’s steering their attention.

Every model you ship draws boundaries: who gets access, who gets delayed; who is trusted, who is watched; whose errors wetolerate, whose we punish.

Those choices are not technical. They are decisions about how people will be governed.

The Comfortof Indirection

Most of the profession is organized around avoiding that fact.(note 2) Fairness metrics, explainability tools,model cards, differential privacy—all valuable, but all operating at the same safe distance. They assume the categoriesare legitimate, the objectives are appropriate, and the model’s place in the institution is a given.

But thesetools don’t exist despite the field’s need for plausible deniability. They exist because of it. The entire apparatus ofresponsible AI lets you feel ethical while sidestepping the foundational question: do you have the authority to govern people’slives through automated categorization?(note 3)

They help you optimize a system without forcing you to ask whether the system shouldexist. They let you document choices without defending them. They allow institutions to claim oversight while maintaining thatdecisions are technical rather than political.

And when something goes wrong, it’s easy to fall back on the standardline: “The system decided.”

Harm becomes a technical anomaly instead of a consequence of design decisions.

Patterns AreNot Neutral

Data is not neutral history. Labels are not natural categories.

Disability studies and transstudies have extensively documented how algorithmic categorization systems rely on rigid, institutional definitions that erase livedexperience and enforce normative assumptions about bodies and identities.(note 4)

Features reflect

institutional priorities.(note 5) Objective functions encode value judgments.

The model is not discovering what the world “reallyis.” It is producing a particular vision of how the world should be organized.(note 6)

Even deciding which errors mattermore—false positives or false negatives—is a political choice. It decides whose harm is preferable.(note 7)

Thesedecisions require justification, not just documentation.

Why Accountability Feels Uncomfortable

Imagine explaining tosomeone denied a loan why this threshold is legitimate.

Or defending to someone flagged as fraudulent why these featurescapture wrongdoing.

Or telling someone whose content was removed why that boundary is appropriate for speech.

Not with“the model said so” or “the data demanded it,” but with reasons you are willing to stand behind.

Ifthat feels uncomfortable, it’s because the system was never designed for you to occupy that role. Technical vocabulary allowsyou to stay adjacent to responsibility while avoiding its weight.

But predictive systems are now deeply embedded in decisionsthat shape access, opportunity, and dignity. If you build these systems, you are participating in governance—whether you nameit or not.

The Path That Responsibility Requires

A responsible practice would begin with acknowledging, plainly, that:

• Modeling is rule-making.
• Feature selection and labeling are political acts.
• Thresholdsand objectives embed values.
• Deployment choices determine who will bear the burden of error.

Andacting on that acknowledgment would require:

• making the system’s value judgments explicit;
• creating real avenues for people to challenge how they’ve been categorized;
• accepting responsibility forharms rather than attributing them to “complexity”;
• treating model behavior as something institutions areaccountable for, not something they can hide behind.

This isn’t a call to abandon predictive systems. Many canreduce error or mitigate bias when carefully designed. But they can only be legitimate if the humans behind them are willing to standbehind the choices they encode.

Note 1.

On how prediction actively organizes social behavior andexpectations, see Jenna Burrell & Marion Fourcade, “The Society of Algorithms,”Annual Review of Sociology (2021), which describes prediction as a form of anticipatory governance rather than neutralinference.

Note 2.

Ben Green’s 2021 essay“Data Science as Political Action,” arguespractitioners must recognize themselves as political actors engaged in normative constructions of society rather than neutralresearchers.

Note 3.

On how AI ethics frameworksfunction as legitimation rather than accountability mechanisms, see Brent Mittelstadt, ‘Principles Alone Cannot Guarantee Ethical AI,’ Nature MachineIntelligence (2019); Jacob Metcalf et al., ‘Owning Ethics: Corporate Logics, Silicon Valley, and the Institutionalization of Ethics,’ SocialResearch (2019).”

Note 4.

Scully, Jackie Leach & Gemma van Toorn, “Automating Misrecognition: The Case of Disability,” Big Data & Society (2025) examines how algorithmiccategorization systems fail to recognize the diversity and context-dependency of disability experience.

5

Note 6.

On how ML systems instantiate normative worldviews through seemingly technicaldesign decisions, see Madeleine Clare Elish & danah boyd, “Situating Methods in the Magic of Big Data and AI,” Communication Monographs (2018), which examines how data practices encode institutional perspectives into systemoutputs.

Note 7.

White, Jason J.G., “Fairness of AI for People with Disabilities: Problem Analysis andInterdisciplinary Collaboration,” ACM SIGACCESS (2023) argues error trade-offs in algorithmic systemsdisproportionately disadvantage disabled people and cannot be resolved through simple harm reduction frameworks.