1. What problem does this paper address?

The problem is that governments make decisions about AI before they can measure its public effects. The authors write that those decisions determine whether a person can learn that AI influenced an outcome, challenge it, and see it corrected[1].

Existing research has documented what AI governance produces: principles, strategies, registers of AI systems, and audits. What the authors find missing is a shared language for understanding AI policy while it is still being made. A full set of principles and strategies does not, by itself, tell you what an individual can now actually do.

The paper asks whether AI policy can be judged not by what it makes visible, but by who can act on what becomes visible. The work discussed here is a preprint posted to arXiv and has not been peer reviewed; what follows should be read as the authors' claims.

2. What does it propose?

The core idea is to read AI policy instruments by whether a person can get notice, challenge a decision, and see it corrected. The authors call this the Society-Centered AI Governance Framework, intended to apply to instruments in any jurisdiction[1].

The reading follows three rules. First, read a jurisdiction's instruments together, by date and binding status: executive orders, strategies, guidance, and statutes are layered in time order rather than read one by one. Second, separate an agency duty from a right a person can invoke. An agency may be obliged to give notice, but if the person has no right to demand it, the person has nothing to act on. Third, trace which later instrument must enforce each commitment.

The third rule is the heart of the framework. A strategy document may promise to set up a challenge mechanism, but without a later instrument that carries it out, the promise stays a promise. By following which commitment is carried out by which document, and under whose responsibility, the authors argue one can measure how far protection reaches even while policy is still forming.

3. What does it show? Only what the abstract states

The authors tested the framework in North Carolina[1]. They used two kinds of material. One was interviews with nine members of the state's AI Leadership Council, conducted before and after the state's 2026 AI roadmap. The other was a provision-level reading of 49 AI governance documents from 19 states, with model-assisted coding reviewed by humans.

For the state's own AI systems, the authors describe what they call visibility-first governance. It comes down to three points.

The roadmap assigns owners and dates to the missing protections, so, in the authors' words, the gaps are documented but not yet closed. They add that recent federal preemption proposals focus on AI development, so recourse for a state's own systems is likely to remain a state decision. They then identify "legitimacy before protection" as a testable risk of visibility-first governance and name the milestones that would test it.

4. A concrete case

Consider a state agency that uses AI in reviewing benefit applications, and an applicant whose application is put on hold. This is a hypothetical for explanation, not a case from the paper.

Visibility only

The state's online register lists AI as being used in benefit reviews, and the agency has a duty to publish that use. But informing the applicant that AI played a part in their own review is written only as an agency duty. The applicant has no right to demand notice and no dedicated procedure for contesting the AI-influenced decision. Reading the register reveals that AI exists; the applicant cannot go any further.

Recourse that reaches the person

The applicant learns through a notice that AI was involved. Receiving notice is defined as the applicant's right. If they disagree, there is a procedure to challenge the decision; if an error is found, the decision is corrected and the result comes back to them. It is also clear which document implements this procedure and which office is responsible.

The difference is not whether AI use is visible. In both cases it is. The difference is whether the person can act on what is visible. The authors' framework tries to pick out this gap from the language of policy documents. If transparency efforts are credited and the policy gains legitimacy while things stay as in the left column, that is how the "legitimacy before protection" risk can be read.

5. What is not new, and where are the limits?

Treating notice, challenge, and correction as pillars of AI governance is not new in itself. The right to know the reasons for an administrative decision, contest it, and have it corrected has long been discussed in administrative law. The paper's contribution is to turn that into a reading procedure for AI policy documents: layer them by date and binding force, separate agency duties from individual rights, and trace the later instruments that implement each commitment.

Several limits are visible from the abstract. First, the framework is evaluated in one state. Documents from other states are read too, but the interviews and the detailed application center on North Carolina, and whether it works the same way elsewhere is not shown. Second, the interviewees are council members, the people who make the policy. The experience of residents who are actually subject to AI-influenced decisions is not part of the material, as far as the abstract shows.

Third, document coding was model-assisted. Humans reviewed it, but the abstract does not describe how the review was done or how well coders agreed. Fourth, part of the conclusion, such as the expectation that recourse will remain a state decision, is a forecast that depends on how federal proposals develop. Fifth, "legitimacy before protection" is a risk the authors name as testable; it has not yet been tested. Within the scope of this paper, what can be said is that in this state's documents and interviews, visibility duties took effect first, while a right to notice and procedures for challenge and correction were not yet in place.

6. Why this topic is clustering now

In the selection record, this paper sits in a cluster with another preprint on a global framework for AI safety and governance; the cluster size is 2. Hugging Face upvotes were 0, and GitHub stars were 0. Attention numbers are small; it was selected because independent papers questioning governance frameworks appeared together. Upvotes and stars indicate topicality only, not correctness or importance.

The other paper in the cluster is a report that, grounded in multilateralism, recommends a global AI governance system centered on the United Nations[2]. It discusses how to design institutions that provide safety, inclusiveness, clear responsibility, and effective coordination.

Both use the word "framework," but they look at opposite scales. One addresses the institutional design of international bodies; the other asks whether a single resident can contest a decision. Keywords such as "governance," "framework," and "safety" appearing alongside "society-centered" and "well-being of humanity" suggest that the question of whom AI rules serve, and at what level they are built, is being raised from above and below at the same time. With only two papers, it is too early to call this a broad trend.

7. What connects to pharmaceutical and regulatory practice?

The connection for pharmaceutical work is the habit of separating an organization's duties from the rights of the people affected. A pharmaceutical company is not a government agency, but it increasingly uses AI in decisions: reviewing internal documents, triaging safety information, drafting answers to inquiries. If internal rules say that AI use will be recorded, but those affected by the resulting decisions have no path to challenge and correct them, then, in this paper's terms, the company has visibility without recourse.

The same question arises on the regulatory side. The US FDA's draft guidance on AI used to support regulatory decision-making for drugs and biological products describes assessing model credibility for each context of use and documenting the results[5]. That framework mainly asks those who use AI to explain its credibility. What affected parties can learn about an AI-influenced decision, and how they can contest it, remains a separate issue.

The paper does not address pharmaceuticals or drug regulation. Its subject is AI systems used by a US state government itself, and transferring its conclusions to companies or drug regulation would need separate examination. Even so, one checking step can be borrowed directly: when writing rules on AI use, do not stop at duties to record and disclose, but write down who implements challenge and correction procedures, in which document, and by when. That mirrors how North Carolina's council[4] used its roadmap[3] to assign owners and dates to missing protections.