Equations
getty
Mathematical models are often treated as if they were neutral instruments. They are presented as disciplined, objective, accurate and insulated from human subjectivity. In management, finance, public policy, and now artificial intelligence, mathematics is frequently invoked as the language that removes ambiguity and replaces opinion with fact.
But this is misleading.
Mathematics can formalize a worldview with extraordinary precision. It can make decisions consistent, scalable, and defensible. What it cannot do is decide, on its own, what the world is for, what should matter most, what kind of trade-offs are acceptable, or what counts as a good outcome. Those decisions are made before the equation is written.
This is why world modeling is not a path toward more intelligence, but toward the codification of a particular way of seeing, selecting, and valuing the world—toward the formalization of judgment itself, whether human, algorithmically mediated, or algorithmic. It is a reminder of where mathematics begins. A model of the world is never simply discovered or approximated. It is designed. It emerges from prior judgments about purpose, relevance, value, and acceptable sacrifice.
To see this clearly, it helps to leave abstraction behind and work through one concrete case.
Imagine a bank that has enough capital to approve three out of five small-business loans. The applicants are the following:
Fictional lending profiles
Hypothetical dataset
At first glance, this looks like a straightforward analytical problem.
The applicants can be scored. The strongest can be selected. The bank can defend its choices with data. But there is no such thing as the correct mathematical model of this situation. There are many possible models, each mathematically coherent, each internally rational, and each based on the same data, yet each modeling a different world. What changes is not the arithmetic. What changes is the value system being encoded.
Suppose the bank’s primary aim is to maximize expected profit. It builds a score that weights creditworthiness, income stability, and business potential as follows:
Profit Score = 0.45(Credit) + 0.35(Income Stability) + 0.20(Business Potential)
To simplify the calculation, let us normalize the credit score to a 0–10 scale: A = 7.8, B = 7.2, C = 6.8, D = 6.4 and E = 6.1.
The resulting scores are: A = 0.45(7.8) + 0.35(9) + 0.20(6) = 7.86; B = 0.45(7.2) + 0.35(7) + 0.20(8) = 7.29; C = 0.45(6.8) + 0.35(6) + 0.20(9) = 6.96; D = 0.45(6.4) + 0.35(5) + 0.20(8) = 6.23; E = 0.45(6.1) + 0.35(4) + 0.20(7) = 5.55.
Then the maths approves A, B, and C.
The decision appears objective. But the objectivity is narrower than it seems. The equation is already telling a story about what matters. Why is credit worth 45% of the score? Why is business potential worth only 20%? Why is neighborhood wealth excluded but prior financial performance privileged? The answer is simple: because the institution has decided that financial return is the highest good. What looks like neutral mathematics is already a moral ordering of the world.
Let’s now consider that the institution adopts a more entrepreneurial philosophy. Rather than rewarding present stability, it decides to reward future promise. It changes the scoring formula: Growth Score = 0.20(Credit) + 0.20(Income Stability) + 0.60(Business Potential). The new scores are: A = 0.20(7.8) + 0.20(9) + 0.60(6) = 6.96; B = 0.20(7.2) + 0.20(7) + 0.60(8) = 7.64; C = 0.20(6.8) + 0.20(6) + 0.60(9) = 7.96; D = 0.20(6.4) + 0.20(5) + 0.60(8) = 7.08; E = 0.20(6.1) + 0.20(4) + 0.60(7) = 6.22.
Now the bank approves C, B, and D.
Applicant A, who seemed strongest under the earlier models, is rejected.
Nothing in the data changed. Nothing in the mathematics became less rigorous. The only difference is that the model now encodes a different answer to a different question. It is no longer asking, “Who looks safest?” It is asking, “Who seems most capable of building the future?”
That shift reflects a belief that latent potential should matter more than established advantage. That’s another different worldview.
Finally, suppose now the bank recognizes that traditional indicators—credit history, income stability, location—often reflect accumulated social privilege as much as individual merit. It decides that a fair decision process should not merely predict safety. It should also correct for structural exclusion. It constructs the following score: Equity Score = 0.30(Business Potential) + 0.20(Income Stability) + 0.15(Credit) + 0.35(Social Vulnerability). The results are: A = 0.30(6) + 0.20(9) + 0.15(7.8) + 0.35(1) = 5.12; B = 0.30(8) + 0.20(7) + 0.15(7.2) + 0.35(3) = 5.93; C = 0.30(9) + 0.20(6) + 0.15(6.8) + 0.35(6) = 7.02; D = 0.30(8) + 0.20(5) + 0.15(6.4) + 0.35(8) = 7.16; E = 0.30(7) + 0.20(4) + 0.15(6.1) + 0.35(9) = 6.97.
This time, the bank approves D, C, and E.
Under classic financial logic, E was the least attractive candidate. Under an equity-oriented model, E becomes fundable. Again, mathematics has not ceased to function. On the contrary, mathematics is functioning exactly as designed. It is translating an institutional commitment into a decision rule. The commitment here is that fairness is not external to the model; it is part of what the model is trying to achieve.
This does not make the model less mathematical. It makes explicit what all models quietly contain: a theory of what deserves to count. A theory of what kind of world matter.
Using the same people, the same variables, and the same formal discipline, we obtained five different rational outcomes: Bank 1 approves A, B, C ; Bank 2 approves C, B, D ; Bank 3 approves D, C, E.
This is not a failure of mathematics. It is the proper functioning of mathematics inside different normative frames. The maths does not tell us what the world is. It tells us what the world looks like once we have decided what matters in it.
That decision enters at every stage: what problem is being solved, what outcome is worth optimizing, which variables count as relevant, how much weight each variable receives, what trade-offs are acceptable, whether inequality is noise or a moral signal, whether the future should be judged by past patterns or imagined otherwise, etc.
These are not mathematical decisions. They are human decisions. Mathematics just operationalizes them.
Actually, I have made one of these kinds of decisions in writing this article: the act of 'normalizing’ the data—translating a 640 credit score into a 6.4—is a hidden act of governance. By choosing a linear scale over a curve, we decide that every point of credit is created equal. We choose where the 'bottom' is. In doing so, we might mathematically erase the struggle of those at the margins or exaggerate the excellence of those at the top. The bias is not just in the weights we give the numbers, but in the shape we give the numbers before the weights are even applied.
The belief that AI models are neutral is not harmless. It gives companies and institutions permission to disguise judgment as inevitability. It allows priorities to be presented as facts, and trade-offs as technical necessities. It shifts responsibility from decision-makers to systems, as though the equation itself had spoken. As algorithmic systems are used to allocate credit, rank job candidates, predict risk, assign resources, or filter information, the temptation is to treat mathematical formalization as if it were moral accuracy. Yet the opposite is often true: The more mathematically sophisticated a model becomes, the more easily the worldview embedded in its design disappears behind the authority of technical complexity, especially when the underlying modeling choices that determine what counts, what is ignored, and what is optimized are no longer visible. This opacity operates on at least three levels: structural, epistemic and institutional.
As models become more sophisticated, the normative choices that shape them are spread across many technical components rather than appearing in one visible formula. In a simple model, one can often identify the variables, the weights, the thresholds, and the objective directly. In a more complex system, those same choices are distributed across data collection, feature selection, proxy construction, architecture design, objective functions, hyperparameters, filtering rules, and post-processing mechanisms. It creates a structural opacity.
The result is that the worldview of the model does not disappear. It becomes harder to locate. What matters is still being decided, but those decisions are now embedded in layers of design that are difficult to inspect as a whole.
A second layer of opacity comes from the limits of what observers can actually know about the model’s internal logic. Even when an AI system performs well, it may remain unclear why it reaches a particular output, which variables are truly influential, how correlations are being used, or which trade-offs the AI system has learned to privilege.
This matters because opacity is not only a problem of secrecy. It is also a problem of intelligibility. A model may be fully available in technical terms and still remain inaccessible in conceptual terms. That’s an epistemic opacity. We may be able to see the code without being able to reconstruct the reasoning in a way that makes the embedded judgments understandable or contestable.
The third layer comes from the social setting in which models are deployed. Most people affected by a model do not design it, cannot audit it, and often do not even know which assumptions govern it. In practice, access to the relevant modeling choices is usually unevenly distributed across institutions, vendors, regulators, technical teams, and end users. This means that the authority of the model is often accepted without meaningful visibility into the value choices that shaped it. What appears as neutral technical output may in fact reflect organizational priorities, regulatory constraints, commercial incentives, or historical biases that remain hidden from those subject to the decision. Here comes the institutional opacity.
The common assumption is that mathematics reveals reality by stripping away subjectivity. In practice, mathematics often does something more consequential: it stabilizes a chosen interpretation of the world we wish to bring about and makes it actionable. This is why the most important question to ask about an AI model is not only, “Is it accurate?” It is also, “Accurate for what?” Not only, “Does it predict well?” but, “In service of which objective?” Not only, “Is it optimized?” but, “Optimized according to whose values?”
These are not secondary questions to be added after the technical work is complete. They are the precondition for integrity-led technical work. Mathematics is powerful precisely because it can give form, consistency, and force to human judgment. But that is also why humility is required. When we forget that models are built out of choices, we begin to mistake our design for neutrality. And that is the central point: Mathematics is not defining the world. It is the world that we define with mathematics.
That is precisely why Artificial Integrity matters more than Artificial Intelligence.
Artificial Integrity is important because it seeks to restore a forgotten layer of discernment, one that has become inaccessible as we have normalized the misalignment between what we shape and what is.
Without Artificial Integrity, AI is reinforcing a path to turn partial objectives into total algorithmic systems, and contingent assumptions into invisible norms.
It reminds us that the challenge is not simply to build more powerful AI systems, but to ensure that the logics they scale deepen our discernment, so that we can see and acknowledge the neutrality gap we have normalized while mistaking it for neutrality and keep alive the sense of the integrity of the world we inhabit.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。