What a Word Is Allowed to Mean: Regulatory Fragmentation and the Language of Algorithmic Discrimination 

Most legal disputes, underneath their procedural surface, center on what a word is allowed to mean. A contract turns on whether "reasonable" bears one definition or another, and a constitutional case turns on whether "commerce" extends this far or only that far. This is not a defect in law; it is the condition law operates under, because words carry more than one plausible meaning and someone, eventually, has to decide which one governs. What is new, and urgent, is that AI regulation has inherited this same indeterminacy at a scale no prior body of law has had to contend with. A single contested word no longer settles one case between two parties. It can settle, silently and simultaneously, how a trained model behaves toward millions of people, for as long as the model remains in use. 

The current developments surrounding algorithmic lending are a case in point. 

The United States has no federal legislative framework governing AI comparable to the European Union's AI Act— a single document that every subsequent guidance, standard, and enforcement action can be checked against, the way a constitution checks ordinary legislation. What exists instead is a patchwork: the Consumer Financial Protection Bureau maintaining that fair lending law applies equally to human and algorithmic decisions; the Equal Credit Opportunity Act; a scattering of state statutes, each independently worded. Colorado's AI Act, the first comprehensive state law addressing algorithmic discrimination, requires "reasonable care" from developers and deployers of high-risk systems to protect consumers from discriminatory outcomes in lending and other consequential decisions. It has already been delayed twice, and as of this spring, a state court has ordered enforcement halted until rulemaking is finalized—rulemaking that had not yet begun. New Jersey has taken a different path, codifying disparate impact directly under state law with explicit algorithmic-decisioning guidance. Massachusetts secured a $2.5 million settlement against a student lender whose underwriting algorithm produced disparate outcomes across protected classes. Each jurisdiction is reaching, independently, toward the same underlying concern, in different words, with different enforcement postures, and without shared authority reconciling them. 

This alone would be a story about fragmentation. What makes it a story about language is what happened next. 

A federal executive order issued in December 2025 explicitly targeted state laws—Colorado's chief among them—that require companies to avoid "differential treatment or impact" on protected groups. The order uses the very wording of these requirements that the states presented as consumer protection to demonstrate the opposite: the intent to harm consumers by demanding that AI systems "produce false results" to avoid a differential outcome. What the states portrayed as an attempt to advance social inclusion, the order described as an embedding of "ideological bias." The burden of this entire argument rests on a single word—discrimination. In one framing, algorithmic discrimination is the harm the law exists to stop. In the other, preventing algorithmic discrimination is defined as forcing a system to lie and the law that requires that such discrimination be eliminated is ideological interference with a neutral technology. Nothing about the underlying algorithmic behavior has changed between these two framings. Only the word governing it has been made to carry a different weight. 

This problem is neither uniquely American nor new to law generally. From the linguistic point of inquiry, the question is whether the EU's binding approach actually solves or simply relocates it. The EU AI Act provides something the American patchwork does not: a shared vocabulary that every subsequent rule must remain consistent with, a linguistic standard embedded in the governing law that ordinary legislation is checked against. That structural floor has real value. However, language is a living system, and a floor built from it is only as stable as the contextual meaning of words that constitute it. Consequently, binding legislation is not immune to the same indeterminacy that afflicts everything built from words. Guidance documents interpreting the Act have already been found, on close reading, to conflate related but distinct concepts—for example, treating "assessment" and "oversight" as near-synonyms in places where the distinction has practical regulatory implications. A binding framework reduces the number of independent battlegrounds over meaning. It does not eliminate the battlefield. 

What would it take to build something closer to a genuine floor—not a complete rulebook, which no document could be, but a stable enough shared vocabulary that ambiguity at any single point could be caught and clarified to stop it from propagating, unchecked, through every subsequent layer of guidance, statute, and model behavior built upon it? This is not Ia question one research project, or one researcher, can answer. It resembles the kind of standard-setting work that would need an institutional home with genuine authority to convene the interdisciplinary and inclusive input that meaningful legal standards have always required, with linguists included not as an afterthought but as a discipline with real jurisdiction over the question. What this project is set to investigate is whether such a floor is achievable at all without binding federal legislation or whether, as the current fight over the word "discrimination" suggests, even binding legislation only narrows the argument, but does not resolve it. This is an open question that necessitates dedicated, thoughtful, and rigorous research. 

What is not in question is the cost of leaving the gap where ambiguity governs over definition unexamined. Somewhere in the space between "reasonable care to prevent differential impact" and "producing false results to avoid differential impact," there are real families, applying for real mortgages, with their applications evaluated by AI systems. And regulators, who interpret those systems’ training data and deployment guidelines, do not work from a shared definition of the one word the entire dispute turns on. The stakes of assigning a stable, unambiguous definition to that word are not abstract—they are the difference between a family's home and a lender's algorithm quietly deciding, in a manner no one downstream is positioned to catch, that the family does not qualify. 

Written by Jelena Simmons on July 22, 2026