The risks they named, the guardrails they built.
How all six newsrooms are governing AI.
01 · Key risks the newsrooms identified
Declining audience trust tops the newsrooms’ list of risks. 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”). 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. Compliance does not guarantee acceptance.
Misinformation reaching the audience is the second near-universal risk. 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. Every error that stayed out of print was caught the same way, with a human reviewing before publication.
Uncontrolled individual use is another risk. “If you have each person doing it their own way, you’re likely to leave room for more mistakes,” noted an editor, adding that uncoordinated adoption erodes trust between colleagues. One newsroom names the pattern shadow AI, with everyone using it without transparency, and staff declining even the newsroom’s own sanctioned tools rather than admitting they use AI at all.
Over-delegation to AI is another risk. “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.
External contributors submitting AI-generated pieces. More editors are experiencing an increase in AI-generated submissions, further jeopardising audience trust if uncaught before publication.
Tools going dormant after launch. Two of the prototypes documented here are no longer in regular use: a weather-alert generator whose champion changed roles, and a title assistant now used manually by its creator when needed. The causes differ, but the pattern remains the same. Without a named owner and a routine check on whether tools still work, even a successful tool build stops working as intended.
Governance failures as risks in themselves. Policies that go dormant, the absence of feedback channels through which staff could raise the alarm, and low editorial buy-in, with end users brought in late or the newsroom involved at the wrong moments.
02 · Mitigations and guardrails
Human review before publication is non-negotiable. What varies is how it is applied: routing every output to a named approver (an art director, or 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 and social media manager for platforms. Internal-facing outputs are checked lightly and audience-facing ones strictly.
Scoping and red lines. AI can be used for promotional artwork but never to generate editorial imagery. Sensitive functions such as fact-checking are centralised in one trained agent under two editors, with lower-risk uses left open to staff. Permission-controlled AI features in the publishing system. One written policy adds an approved-tools list, which an editor approves before journalists use them, reviews quarterly, and bars unapproved platforms from sensitive editorial material. Another pairs its whitelist with a lifecycle of proposal, risk assessment, pilot testing, conditional approval and ongoing review, under which a tool can be suspended or withdrawn. It also tiers uses by risk, from routine tasks needing no extra approval to prohibited uses touching source confidentiality or unpublished material, and for major stories journalists document which AI tools were used, for what purpose, how outputs were verified, and what disclosure was decided.
Disclosure to the audience is agreed in principle. Whether to disclose every time or some of the time, generically or naming the tool, on air or per post remains a challenge. Chequeado publishes a visible disclosure on any product where AI played a relevant role, with a ready-made formula stating that AI tools assisted the production under the outlet’s editorial standards and that the team reviewed and verified all content, plus credits or watermarks on audiovisual material. Orkonerei requires disclosure when AI generated or significantly rewrote a script, summarised a bulletin, translated in a way that affects meaning, or produced synthetic voices or images; not for spelling checks, formatting or internal transcription. Radio Domus opts for disclosures such as a short tagline, a caption or credit on social media, or an editor’s note voiced on air. This is extended when appropriate to publishing the prompts fed to the model.
Document architecture varies. Chequeado separates a stable public policy from internal, frequently updated protocols so the public document doesn’t go stale. La Diaria tiers its documents by audience, with a public declaration of its philosophy, an internal summary, and the full policy held by an AI-use committee, alongside a misuse track covering protocols, sanctions and responsibility.
Verification rules. Any AI-supplied fact or figure must be traced to a source before use, and every number in an auto-drafted article checked by a journalist. Chequeado’s published policy states that AI can never be a primary source of information. One tool carries the rule into its own interface, with a standing footer telling users to always verify the original sources.
Culture-building and internal training. Language “clinics” that feed a community’s approved registers back into the tools. Physical banners stating that AI is allowed in moderation, with checklists. A log of past moderation decisions to build in-house precedent. Talking resistant staff through why the sanctioned, house-trained tool beats their personal one. Chequeado pairs its policy with a protocol of operational criteria, and staff can ask an internal AI assistant, DobleCheq, when unsure what the policy allows.
Data protection. Radio Domus collects only the data that moderation and feedback require. They restrict who can view, edit or export audience data, keeping sensitive data out of open AI tools and reviewing false positives and negatives for bias. Unpublished investigations, embargoed stories and internal editorial deliberations stay out of external AI systems. Chequeado prohibits entering sensitive information, such as anonymous sources, personal data or confidential documents, into external AI tools, and is the only policy to cover intellectual property, preferring tools with authorisation criteria and renouncing no rights over its own original content. Orkonerei goes furthest into a community newsroom’s particulars. Confidential sources, whistleblower recordings, interviews with survivors of gender-based violence, sensitive land-rights and legal cases, and community members’ personal data must not enter third-party AI platforms without a risk assessment and editorial approval. Journalists first ask where the data is stored, whether the platform trains on it, and whether a source could be exposed.
03 · Shared principles for governance
AI assists, humans decide
“Journalists must be journalists, even if the AI is there.”
Expand capability, don’t replace people
Automation exists to free small teams for higher-value work.
Transparency as identity, not compliance
Disclosure framed as who they are rather than a rule. Honesty about AI use as a duty to a paying or watching audience, including the disarming formulation that anyone claiming in 2026 not to use AI “is lying to you.” The real obligation is to say how you check it and own the mistakes. Chequeado makes owning them concrete: errors derived from AI use are corrected with the same visibility the original had.
Sound like yourself
Identity preservation as governance. Growing “without losing the soul,” output in the community’s own language and register, and house format as the thing a tool must learn before it is trusted.
Policies are living documents designed for revision
They are published as a starting point or guide under continuous, ongoing review, and a draft following national regulation.
A duty to educate and protect the audience about AI
Newsrooms should help their audience understand AI, treating AI literacy as part of news literacy. Their disclosures are written to teach as well as to declare.