Where policy stands, what breaks, what holds.

1 · Where policies stand, and how they were made

All six newsrooms were actively drafting or overhauling AI governance during the programme, but only one had published: a fact-checking outlet released its policy in September 2025, explicitly as “a starting point”, and its advice to peers is blunt — “don’t wait anymore”; publish and iterate rather than waiting for a perfect document.

The other five span the realistic spectrum: an advanced draft intended for publication in the outlet’s own pages; a from-scratch redraft after a 2021 policy went dormant (“naive”, its authors now say — written for the predictive-text era, and it stopped circulating); a document deliberately called a “guide” rather than a policy, because a guide is easier to keep adjusting; a draft already operating as the team’s day-to-day reference — “our Bible” — while formal adoption waits on Tanzania’s regulator finishing its own media-AI policy; and guidelines drafted with a law firm, now being rebuilt through a structured mentor process.

The processes converge more than the statuses. All six worked with mentors. At least four opened drafting beyond a small committee — a whole-team brainstorm at one station; organisation-wide comment rights, including fundraising and HR, at another; a deliberately mixed working group including a sceptic at a third. Three ran staff surveys as drafting instruments, and two paired them with audience surveys — one finding that staff and audience alike span the full spectrum, from daily users to people who know AI only as a phone-camera feature, a corrective to drafting policy for an imagined uniformly savvy readership.

Three anchored their documents to national frameworks — a media council’s AI guide in Kenya, a forthcoming regulator policy in Tanzania — a distinctly East African pattern in this set, with no Latin American newsroom reporting regulatory pressure. And two newsrooms independently describe the process itself as the real product: organised internal meetings surfaced unexpected AI adopters and “installed the conversation” in a newsroom with no shared physical office; another calls the whole exercise “a quiet revolution, a slow revolution… don’t rush it.”

2 · Risks the newsrooms identified

Audience trust is the risk named by all six, and it is concrete, not abstract. A subscription-funded paper frames every AI decision against “the trust of the people that are trusting us with the money by month.” A fact-checker concluded it was “impossible to be transparent and not be transparent on how we use AI.” Community stations report audiences detecting foreign-sounding output — “this is not you guys, this English is totally foreign” — and one editor’s rule of thumb is that “anytime they think this is AI, there’s that negative connotation.” Two newsrooms learned that even disclosed, guideline-compliant AI use can draw criticism — one experiment prompted the objection “you are wasting water” — so compliance does not guarantee acceptance.

Errors reaching the audience is the second near-universal risk (five of six describe incidents or near-misses): a hallucinating agent asserting Barack Obama is the current US president; AI-supplied statistics with no provenance; a script rearranged incorrectly without the journalist noticing; a published caption that audibly read as AI; and, at one Latin American newsroom, an AI-generated image with a politically explosive map error published without supervision — the cohort’s sharpest cautionary tale. Notably, every error that stayed out of print was caught the same way: a human reviewing before publication.

Three further clusters recur. Uncontrolled individual use (four of six): “if you have each person doing it their own way, you’re likely to leave room for more mistakes,” plus quiet, uncoordinated adoption eroding trust between colleagues. Over-delegation (four of six): “once people understand AI can do it, they will just be watching AI do it,” against the counter-principle that the machine must not decide what is broadcast. And external contributors (three of six — all Latin American): outlets publishing many outside authors cannot currently know whether a submission used AI; one identified and declined to publish an AI-written piece, caught only by an editor’s hunch — “the policy is for the people to trust us, but for us to trust us too.” Two newsrooms also name governance failures as risks in themselves: policies that go dormant, tools nobody maintains, and — after one failed project — the absence of any channel through which mid-level staff could raise the alarm.

3 · Mitigations and guardrails

Human review before publication is universal — six of six — and it is the cohort’s one non-negotiable. The instructive variation is in how it is made real rather than aspirational: routing every output to a named approver (an art director; an editor at the end of a Slack submission chain); a layered sequence in which the journalist reviews first, then a role-specific checker — editor for on-air content, social media manager for platforms; a coordinator who deliberately “looks behind and beyond” the work, with a stated deterrent effect (“they know very well there’s someone waiting to check that”); and risk-calibrated depth, with internal-facing outputs checked lightly and audience-facing ones strictly.

Beyond review, four families of guardrail recur. Scoping and red lines (four of six): AI for promotional artwork but never article imagery; no AI in the subscriber newsletter, “like a letter to our subscribers” — a collective decision the interviewee personally disagrees with, and observes anyway; centralising sensitive functions such as fact-checking in one trained agent under two editors, with lower-risk uses left open to staff; permission-controlled AI features in the publishing system.

Disclosure to the audience (six of six in principle) — yet in five of six the mechanics were unresolved at interview time: whether to disclose every time or only sometimes, generically or naming the tool, on air or per post. One newsroom separates a stable public policy from internal, frequently updated tool-level protocols so the public document doesn’t go stale. Verification rules: any AI-supplied fact or figure must be traced to a source before use; every number in an auto-drafted article checked by a journalist. Culture-building: internal trainings; language “clinics” that feed a community’s approved registers back into the tools; physical banners in the newsroom stating that AI is allowed “but moderately”, with checklists; a log of past moderation decisions to build in-house precedent; and talking resistant staff through why the sanctioned, house-trained tool beats their personal one.