AI policy in journalism| pasteup of 107 policies| the opening Aug 2026

AI policy in journalism

Who allows what — and who actually means it

107 policies, two questions

Every mark on this board is one published AI policy: from professional newsrooms (blue circles), student publications (orange squares), journalism schools (green triangles) and industry bodies (gray diamonds).

Each policy is scored twice: how much AI it permits, and whether it can actually be enforced. Keep scrolling; the board composes itself.

How to read the board

Left to right: what the policy allows — from AI text reaching publication with light gates, out to near-total bans on generative AI in published work.

Bottom to top: whether it has a mechanism: named tools, a defined approval chain, a person accountable, a stated consequence. Strong language is not a mechanism, and “be thoughtful with AI” scores low no matter how stern it sounds.

The dashed crosshair splits the board into four postures a newsroom can hold.

Strictness and seriousness are separate decisions

Across all 107 policies the two axes correlate at only r = 0.33, a weak relationship (1.0 would mean strictness always arrived with enforcement). All four quadrants are populated: 31 permissive-and-precise, 29 restrictive-with-teeth, 28 ungoverned, 19 prohibition-on-paper.

Knowing how much AI an organization allows tells you almost nothing about whether anything happens when the rule is broken.

Students set the strictest standards

Student newspapers median 70 on restrictiveness and 72 on enforceability, against 45 and 50 for professional newsrooms. Six of the ten most enforceable documents in the entire dataset are student papers; only 3 of 70 professional newsrooms reach that tier.

The Student Press Law Center, which advises them, explicitly counsels against blanket bans — and they banned anyway.

They are also rare. An August 2026 audit of all 22 CSU campuses found exactly three student newsrooms with a published AI policy: Golden Gate Xpress, Mustang News and Sacramento State’s State Hornet. The rest either keep an ethics code that never mentions AI, or publish no standards document at all.

Unions wrote the enforcement

The three collectively bargained AI provisions here — Politico, VTDigger and the Ziff Davis Creators Guild — median 78 on enforceability against 50 for everything else.

Politico’s was tested in binding arbitration (a private judge whose ruling binds both sides) and management lost. Ziff Davis’s contract bars AI-driven layoffs outright. Editorial standards desks wrote principles; labor wrote consequences.

J-schools are the most permissive sector on the board

Journalism programs median 22 on restrictiveness and 36 on enforceability — the loosest and vaguest group in the study, sitting well to the left of the newsrooms they place students into.

Of roughly 65 schools checked — including every CSU campus — three have journalism-specific AI rules with operative teeth. Columbia has an AI initiative, two research centers and a summit series — and no discoverable student-facing rules. Not one of the 22 CSU campuses’ journalism programs has a published AI policy, even as the system signs a $17M deal to put ChatGPT in front of nearly 470,000 students.

The board is yours

Below: the full instrument. Filter by sector or region, search any organization, click a mark to pin its entry. Every score is anchored to quoted policy language, and every entry links to its source document. Where something could only partly be verified, the entry says so in red.

Work the board

The working board

All 107 policies, live. Hover a mark for the short version; click to pin the full entry and open its source. Hollow marks are stated postures or cautionary cases with no operative rules.

Published operative rules Stated posture only — no operative rules Hover a point for detail · click to pin it below
Click any point to read its policy summary and open the source document.

Every policy, with its source link

Sort by any column. Organization names link to the source document.

OrganizationSectorRegion Restrictive →Enforceable → DatedSource

How the two axes were scored

X — Permissiveness (0 permissive → 100 restrictive)
  1. 0–20 AI use encouraged; AI text can reach publication with light gates
  2. 21–40 Permitted across many workflows including drafting, with review and disclosure
  3. 41–60 Assistive and back-end work; generated content needs editor sign-off
  4. 61–80 Narrow enumerated uses; generative output barred from published work
  5. 81–100 Near-total prohibition on generative AI in published journalism
Y — Specificity and enforceability (0 vague → 100 enforceable)
  1. 0–20 Principles only — nothing operative
  2. 21–40 General do's and don'ts; no named tools, process or owner
  3. 41–60 Enumerated permitted and prohibited uses; stated disclosure requirements
  4. 61–80 Plus named tools, a defined approval chain, or a required labeling format
  5. 81–100 Plus stated consequences, audit or logging duties, or a named accountable role or committee

Why this second axis

The one large content analysis of newsroom AI policies — Becker, Simon and Crum's study of 52 policy documents across 12 countries, published in Digital Journalism — found that roughly 90% require disclosure when AI is used, but only 8% say how the policy would be enforced. That gap is the y-axis. The paper's own framing is institutional convergence rather than "permissive vs. enforceable," but its coded variables — per-use allow/disallow rates, and the presence or absence of an accountability mechanism — are close to a direct operationalization of these two axes.

What is not on this chart, and why that matters

Roughly 120 college publications' policy pages were checked. Eighteen had an AI policy. At least 37 more had a published ethics or editorial-standards document with no AI language in it at all — including the Harvard Crimson, Columbia Daily Spectator, Daily Northwestern, Daily Californian, Cavalier Daily, Daily Tar Heel, Minnesota Daily and GW Hatchet. A further ~25 sites were unreachable to automated fetching and could not be assessed either way.

An August 2026 audit covered every CSU campus — all 22 — plus the Chancellor's Office. Three CSU student newsrooms publish an AI policy: Golden Gate Xpress (SF State), Mustang News (Cal Poly SLO) and The State Hornet (Sacramento State). Eight more keep a real ethics or standards document with no AI language (the Daily Aztec, Daily Titan, Long Beach Current, CSUDH Bulletin, Cougar Chronicle, The Runner, Fresno's Collegian, Chico's Orion stylebook); the remaining papers publish no standards document at all. No journalism program at any of the 22 campuses has one — several campuses have no standalone journalism department, and none of the communication departments that house journalism publish one either — while the system itself signed a $17M OpenAI deal in February 2025 that deliberately leaves AI rules to individual campuses and faculty. Unlike the broader sweep above, every campus publication and journalism-program site in this round was fetched and read directly, including the ones that block automated tools; the one exception is the Chancellor's Office itself, whose own pages block fetching — its quotes come from campus republications and CalMatters.

Roughly 65 journalism schools and programs were checked. Three had a journalism-specific AI policy with operative rules — UNC Hussman, BU's College of Communication, and Minnesota's Hubbard School. Several more have well-funded AI initiatives, labs and summits but no discoverable student-facing rules. The University of Florida's own college page states it has no college-level policy and points students to the university-wide framework.

Where the populated regions sit is informative too. The bottom-right — restrictive but unenforceable — is the thinnest quadrant (19 of 107), which fits the intuition that an organization willing to prohibit AI outright usually also says what happens when someone doesn't listen. But it is not empty, and the four quadrants are otherwise close to evenly loaded (31 / 29 / 28 / 19). That balance is the point: there is no single industry drift, and no reliable way to guess one axis from the other.

Method, and what to distrust

  • Every source URL was fetched and checked against the organization it is attributed to. Of 107 compiled in the July round, 101 verified cleanly, one was wrong and dropped, and four are flagged in-chart as thin; none were dead. The August CSU audit added one entry, read directly on the live site, and re-verified both existing CSU entries.
  • 63 of 107 entries rest on a primary source — the organization's own published policy. The rest rely on trade-press or academic reporting, because several major outlets' own domains (ap.org, nytco.com, wired.com, theguardian.com, bbc.co.uk, cbc.ca, reuters.com) block automated fetching. Filter by source type in the table.
  • Scores are one reader's application of a fixed rubric to the text as published (see "How this study was made," below) — not a measure of what any newsroom actually does. Politico's practice is far more permissive than its contract; Colorado Public Radio's practice is far stricter than anything it has written down.
  • Where two research passes scored the same organization differently, the more recent documented version won. The Guardian is the clearest case: its June 2023 guidance barred generative output intended for publication, while the March 2026 revision permits drafting assistance with senior-editor sign-off.
  • A handful of entries are not policies at all — the Los Angeles Times and Sports Illustrated appear as cautionary cases, plotted where "no rules" actually lands. They are drawn hollow.
  • Union contracts (Politico, VTDigger, Ziff Davis) are included because in several cases they are the only enforceable AI language a newsroom has.

How this study was made

This analysis was directed and orchestrated by Assoc. Prof. Jesse Garnier, who is solely responsible for its content and any errors. The underlying research was compiled, scored, designed and executed by Claude Fable 5; Garnier framed the research questions, set the scoring axes, chose the design direction, and reviewed the findings. Each score was applied against the fixed rubric above and anchored to linked and quoted policy language; sources were machine-fetched and verified as described in the method notes.