The Articulation Trap
As soon as an organization turns taste into a set of criteria, it might start optimizing the very thing that the criteria were meant to safeguard.
A preprint recently published by Stanford and some of its collaborating institutions examined 2.8 million texts alongside 317 million judgments of human preference across seven areas related to creativity. The researchers aimed to grasp something that has generally been ignored in the process of training AI: the difference between what people prefer and what they can say they prefer.
They called it the articulability gap. And they found it everywhere.
Subjective judgment is required not only in creative writing or in peer review, but also in mathematics, in software engineering, and in any field where it is assumed that the criteria for quality are completely known and can be expressed; there is still a significant amount of human preference that cannot be reduced to explicit rules. This gap existed across all tested domains, although its magnitude varied and its presence was not denied.
The fact that brings this to light is that the gap widens as more people take part in the judgment. The authors link this to Collins' idea of collective tacit knowledge.
Once judgment becomes an institutional process, the tacit aspect is not averaged out. It is averaged in.
The researchers considered preferences in three ways:
By means of executable programs (those that can be verified)
By using rubric banks (those that can be described)
By employing densely trained models that capture the full spectrum of human responses, including those that resist both description and verification.
The difference between the rubric model and the full model is known as the articulability gap. This gap is not due to noise; rather, it represents preference information that does not survive the codification process.
The analogy the authors make use of is Goodhart's Law: once a measure becomes a target, it is no longer a good measure. What the present study provides is a more exact mechanism. The issue is not that rubrics are gamed. The issue is that stating preferences pulls the model away from the tacit aspect. The rubric doesn't fail to capture what is important; instead, it captures something nearby, thereby excluding what it cannot name.
This is a kind of problem distinct from that of imprecise criteria since imprecise criteria can be made more precise. The articulability gap refers to a type of preference that is structurally resistant to expression — not due to our lack of vocabulary, but because the judgment involves a kind of knowing that operates at a level below that of rules.
Where Organizations Replicate This Mistake
The researchers were conducting studies on AI training, and the problem they mention is nothing new in the field of AI.
Brand standards turn creative judgment into rules that can be reproduced. The criteria included in RFPs turn procurement judgment into weighted scores. Creative approval checklists turn aesthetic discernment into simple yes-or-no decisions. Employee evaluations turn judgment about performance into measurable competencies. All of these tools are legitimate, and they all involve the same structural risk.
The problem isn't that the criteria are incorrect; it is that something can meet all the criteria and yet still be completely wrong. Anyone who is an experienced creative director, anyone who has a good deal of experience as a procurement officer, and anyone who is a hiring manager with the ability to recognize patterns will know this. They have all come across a submission which fulfilled every criterion and yet seemed entirely devoid of substance. The criteria didn't fail; instead, they achieved what they were capable of doing, and the very thing they were unable to do is precisely what the articulability gap refers to.
A piece of work can meet all the criteria and yet be clearly wrong.
The institutional approach is more hazardous than the individual one. If one reviewer has good judgment and uses a rubric, he or she can go against it when it is misleading. But a committee carrying out the same rubric cannot do this. The group process requires that disagreements be resolved through what can be stated clearly. Since anything that cannot be named cannot be defended in the room, such things are discarded. As a result, the work that comes out conforms and is seldom distinguished.
It is not a flaw of the process; rather, it is the effect that the process has on informal judgment when that informal judgment has no legal protection.
A Decision Framework
It isn't the case that all judgments should be protected from being put into code. The issue is which type.
The study offers a helpful distinction:
Verifiable judgment — that is, judgment in which correctness is definite and can be checked — should and can be codified. A brand standard specifying color hex codes is not an articulability issue, since the rule includes everything the judgment demands.
Articulable judgment — where criteria exist and can be stated, even if they are subjective — makes rubrics useful as a starting point, provided the rubric is applied loosely, and experienced reviewers are allowed to override it.
Tacit judgment falls into the third category; it is in this area that the gap exists. When considering any decision which at present goes through a rubric, the question to ask is:
Could this work be wrong in a way that we should be able to recognize even if we couldn't justify it in a meeting?
If the answer is yes, the rubric should be supplemented by a named reviewer whose judgment cannot be reduced to the rubric's. This is not to override the process, but to take on the part of the judgment that the process cannot hold.
The practical categories break down as follows:
Codify fully when the judgment is verifiable: compliance, technical specifications, format requirements, and regulatory thresholds. Rules capture everything.
Use rubrics as scaffolding when the judgment is articulable yet expert-dependent: brand voice, campaign strategy, and vendor selection. The rubric organizes the evaluation; experienced judgment governs edge cases.
When the judgment in question is constitutive, protect it—for example, in deciding whether creative work has the quality that enables it to succeed, whether a candidate possesses the pattern recognition needed for the role, or whether a brand expression makes an impact. Identify the person whose judgment it is. Do not request them to explain it using a rubric.
What This Means for How You Work
The fact that standards have limitations does not justify rejecting them; rather, it underscores the need to understand what standards can and cannot achieve.
Organizations that replace tacit judgment with fully auditable processes do not become more rigorous; instead, they become more consistent at producing work that just passes the bar without actually meeting the standard. The criteria guard against catastrophic failure but do not lead to distinction.
For human-generated text, the full preference model fit human judgment more closely than the articulated rubric model; the rubric model and the full model frequently reached different conclusions, and in those cases, the rubric model was incorrect.
All organizations have individuals who can tell the difference between work that meets the criteria and work that actually works; the issue then becomes whether those people have the authority to make such judgments, or whether the process has been set up to bypass them.
You cannot convert everything an organization knows how to recognize into a rule without losing some information, and the first duty of leadership is to know which decisions need a rule and which need a person.
Spangher, A., Huang, S. S., Haupt, A., Goodman, N. D., Yang, D., Ho, D. E., & Koyejo, S. (2026). Verifiable, articulable, and tacit components of preference. arXiv. https://arxiv.org/abs/2610.03025
Some ideas are worth discussing in the context of your organization.
