Newsroom AI Policy

What This Skill Does

Generates a complete, publication-ready AI usage policy document tailored to the newsroom's specific context — covering permitted uses, prohibited uses, disclosure requirements, quality gates, training, accountability, and review cadence.

When To Use This Skill

• Your newsroom has no AI policy and reporters are already experimenting with AI tools — you need clear rules before ad hoc habits calcify into unwritten norms
• Your existing AI policy is a one-paragraph statement ("use AI responsibly") and you need specific, enforceable guidance that reporters can actually follow
• You're launching a new publication or digital desk and want AI rules baked in from day one
• Your editor-in-chief or legal team has asked for a formal policy document they can circulate to all staff and freelancers
• You've had an incident — a factual error traced to AI-generated text, a source who discovered AI was used to process their interview, a reader who noticed AI-sounding prose in a bylined column — and you need a policy response fast
• You manage a publication that relies on freelancers and contributors, and you need a policy that extends to people who are not on staff but publish under your masthead
• Your newsroom is part of a larger media group and you need a local policy that fits within corporate guidelines but addresses your specific editorial context
• A journalism school, press association, or industry group has asked you to draft a model AI policy that others can adapt

What You Need To Provide

Required: A description of the newsroom — size (number of editorial staff), beats covered, publication type (daily newspaper, weekly magazine, digital-only, broadcast, etc.), and the editorial values or principles the publication already follows (even informally).

Optional inputs that improve the output:

• Specific AI tools already in use or under consideration (Claude, ChatGPT, Gemini, Otter.ai, Descript, Midjourney, etc.)
• Any incidents or concerns that motivated the policy request — the more specific, the more targeted the policy
• The newsroom's stance on transparency with readers (do you have a public trust statement, a reader advisory board, a corrections policy?)
• Existing style guide or ethics code to align with (SPJ Code of Ethics, AP style, internal house rules)
• Whether the policy needs to address freelancers, stringers, contributors, interns, or only full-time staff
• Any regulatory or union constraints (collective bargaining agreements, GDPR, state privacy laws, industry codes)
• The name and title of the person who will own the policy
• Whether the publication uses AI-adjacent tools already (automated CMS tagging, SEO suggestion tools, audience analytics dashboards) and whether those should be in scope
• The publication's stance on AI-generated images, audio, and video — not just text
• Whether the newsroom has a formal editorial workflow (story budgets, assignment trackers, CMS with tagging) or works informally

How the Assistant Approaches This

1. Maps the newsroom profile to risk categories. A 15-person regional daily covering courts and city government has different AI risks than a 200-person national magazine running long-form investigations, a digital-only outlet producing 40 stories a day, or a monthly trade publication covering a single industry. The assistant identifies which AI use cases are relevant to this specific newsroom, ranks them by editorial risk (from low-risk tasks like transcription to high-risk tasks like drafting publishable prose), and excludes use cases that are irrelevant to the publication's work.

2. Builds the permitted/prohibited framework. For each relevant use case, assigns a clear status: permitted (no special review required beyond standard editorial checks), permitted with disclosure (editor must be informed, readers may need to be informed), restricted (requires senior editor or editor-in-chief approval on a case-by-case basis), or prohibited (never, regardless of circumstances or time pressure). The framework is binary enough that a reporter reading it knows immediately whether their intended use is allowed — no "it depends" without specifying what it depends on.

3. Writes the disclosure rules. Defines three layers of disclosure, each with concrete triggers:

   • Internal disclosure: what reporters must tell their editors, in what format, and at what stage of the workflow
   • Reader-facing disclosure: when and how the publication tells its audience that AI assisted in the work, including the exact placement and language of the disclosure
   • Source disclosure: when sources should be informed that AI tools were used in processing their information, especially in sensitive reporting contexts (whistleblowers, legal matters, personnel issues)

4. Designs the quality gates. Specifies who reviews AI-assisted content before publication, what the reviewer checks for (fabricated facts, unverified claims, AI-sounding prose, source integrity), and how AI-assisted content is flagged in the editorial workflow (e.g., a tag in the CMS, a note in the story budget, a field in the assignment tracker). The gates are proportional to the risk level: routine permitted uses get lightweight review; restricted uses get senior sign-off; any use involving sources or data gets verification against original materials.

5. Addresses edge cases. Identifies the grey areas that a vague policy would leave unresolved and writes specific guidance for each. Common edge cases: using AI to prepare for an interview (permitted or not?), using AI to summarise a source's prior public statements before contacting them, using AI to draft internal memos or pitches that are not published, using AI-powered features built into tools the newsroom already uses (spell-check in Google Docs, smart compose in email, transcription in Zoom).

6. Adds training, accountability, and review mechanisms. Includes how new and existing staff learn the policy (with a specific format — briefing session, onboarding packet, reference card), what happens when the policy is violated (graduated response with at least three severity levels), and a fixed review cadence so the policy stays current as AI tools change. Closes with an effective date and a signature line.

7. Writes in policy language, not tech language. The entire document is written for journalists, not for engineers or lawyers. No jargon. No buzzwords. Every section is specific enough to apply and short enough to read in one sitting. The tone is collegial, not corporate — written by editors for reporters, not by compliance for employees.

Output Format

A single structured document of 1,000 to 1,500 words, formatted for internal distribution. Sections are numbered for easy reference in newsroom discussions. The document includes:

• A preamble stating the publication's editorial values and why the policy exists (3-5 sentences, not a mission statement)
• Numbered sections covering: Permitted Uses, Prohibited Uses, Disclosure Requirements, Quality Gates, Training and Onboarding, Accountability, Review and Updates
• A summary table or quick-reference list at the end for daily use — reporters should be able to check this table without reading the full document
• An effective date placeholder and signature line
• Section numbers and sub-numbers (1.1, 1.2, etc.) so staff can reference specific rules in conversation: "That's a 2.3 situation"

Tone: direct, practical, collegial — written by editors for journalists. Not legalistic, not preachy. The policy should read like a well-written staff memo, not a corporate compliance document and not a manifesto about the future of journalism.

Quality Criteria

• Every section contains specific, actionable rules — not general principles or aspirations
• The permitted/prohibited framework covers at least eight distinct use cases relevant to the newsroom described
• Permitted uses include concrete verification requirements — not just "verify the output" but what to verify and how
• Prohibited uses state clear red lines with no ambiguity about what falls on which side
• Disclosure rules define exactly when readers must be told, when editors must be told, and when sources must be told — with concrete triggers for each
• Quality gates specify who reviews, what they check for, and how AI-assisted content is flagged in workflow
• Training section describes a concrete onboarding mechanism — not just "staff should be trained" but the format, timing, and content of the training
• Accountability section includes graduated consequences — at least three severity levels with examples of what falls into each
• The policy includes a fixed review cadence (e.g., every six months) with a named owner or role
• The document is between 1,000 and 1,500 words — long enough to be useful, short enough to be read
• The quick-reference table at the end is complete enough to answer the most common questions without reading the full document
• No AI buzzwords: no "leverage," "harness," "empower," "cutting-edge," "transformative," "game-changer," "unlock," "supercharge," or "best-in-class"
• All examples are generic — not traceable to any real publication, person, or incident
• Output is free of generic AI-template phrasing — no "delve into," "navigate the complexities of," Latinate filler verbs ("utilize," "facilitate," "leverage" where a plainer verb fits), restate-the-question intros, summary-of-the-summary closes, or em-dash overuse
• The policy acknowledges at least three honest limitations of AI in journalism (fabrication, bias, confidentiality risk)
• A reporter reading the policy can determine within 30 seconds whether a specific intended use is allowed
• If the newsroom includes freelancers or contributors, the policy states explicitly how it applies to them
• The policy addresses what happens to existing content produced with AI before the policy was adopted

Example

Input

Newsroom description: A 15-person regional daily newspaper covering city government, courts, education, and local business. Two reporters have started using Claude for research and draft outlines. The editor-in-chief wants a clear policy before use spreads further. The paper has published a community trust statement promising readers that every story is reported and written by its journalists. No freelancers. No existing AI policy. The editor-in-chief, Maria Torres, will own the policy.

Additional context: The paper uses WordPress as its CMS, with a simple story budget spreadsheet. Reporters file stories via email or directly in the CMS. No formal assignment tracker beyond the story budget. One reporter covers courts and has expressed concern about entering case details into AI tools. The paper runs a weekly education column that draws on state test-score data.

Output

---

AI Usage Policy The Riverdale Gazette — Editorial Staff

Preamble

The Gazette's journalism depends on the trust of our readers and our sources. We promised both that every story we publish is reported and written by our journalists. This policy defines where and how AI tools may be used in our newsroom without breaking that promise. It applies to all editorial staff, effective on the date below. It will be reviewed every six months.

This is not a technology policy. It is an editorial standards document. The same principles that govern our reporting — accuracy, fairness, transparency, independence — govern our use of any tool, including AI.

---

1. Permitted Uses

The following uses of AI tools are permitted without prior approval from the editor-in-chief. Staff must still follow the quality gates in Section 4 and the disclosure requirements in Section 3.

1.1 Transcription. Using AI to transcribe recorded interviews, press conferences, public meetings, or court proceedings that are part of the public record. The reporter must verify all direct quotes against the original recording before publication. AI transcription is a first draft, not a final source.

1.2 Background research. Using AI to summarise public records, compile timelines from published sources, explain technical concepts (zoning codes, court procedures, budget line items, education policy terminology), or identify relevant prior coverage. The reporter must verify every factual claim against primary sources. AI summaries are starting points for reporting, not substitutes for it.

1.3 Headline and summary drafts. Using AI to generate draft headlines, deck lines, social media post text, or newsletter teasers. The final published version must be written or substantially rewritten by a staff member. No AI-generated headline or teaser may be published without staff review and revision.

1.4 Data analysis. Using AI to identify patterns in public datasets — election results, school test scores, property records, budget figures, crime statistics. All findings must be verified by the reporter using the original dataset. If AI-assisted analysis produces a finding that drives the story, reader disclosure is required (see Section 3.3).

1.5 Grammar and style checks. Using AI as a proofreading or style-checking tool after the reporter has written the story. This includes grammar correction, spell-checking, and style consistency checks. It does not include asking AI to rewrite passages for "better flow" or to "improve the prose." Proofreading is permitted; rewriting is not.

1.6 Translation assistance. Using AI to produce rough translations of documents, public statements, or interviews conducted in languages the reporter does not speak. A qualified human translator must verify any translated material that will be quoted directly in publication. Rough translations are research tools, not publishable text.

1.7 Internal communications. Using AI to draft internal documents — story pitches, meeting agendas, notes for the story budget — that are not published. This is a convenience, not an editorial function, and does not require disclosure.

1.8 Calendar and scheduling. Using AI-powered tools to manage schedules, set reminders for court dates or meeting agendas, or compile event listings from public sources. This is administrative, not editorial.

---

2. Prohibited Uses

The following uses are prohibited for all staff, without exception. No deadline pressure, staffing shortage, or efficiency argument overrides these rules.

2.1 Bylined content. AI may not write or substantially draft any story, column, editorial, opinion piece, or letter that carries a staff byline. "Substantially draft" means generating connected prose that forms the basis of the published text, even if the reporter edits it afterward. Our byline means a journalist wrote it. If AI wrote the first draft and a journalist polished it, the journalist did not write it.

2.2 Quotes and dialogue. AI may not generate, reconstruct, paraphrase-as-if-direct, or "clean up" direct quotes from sources. If a source's exact words are unclear from your recording or notes, contact the source to confirm the quote, or paraphrase with clear attribution ("Smith said the project was behind schedule" rather than a fabricated direct quote). This rule applies even when the AI-generated version is probably accurate. Probably is not good enough for a direct quote.

2.3 Source impersonation. AI may not be used to simulate a source's voice, generate hypothetical statements attributed to real people, or produce any content presented as coming from a named individual. This includes using AI to draft "what a source might say" as preparation for an interview — the risk of contaminating your reporting with AI-generated expectations is too high.

2.4 Image generation and manipulation. AI-generated images may not be published in news coverage. This includes fully synthetic images, AI-composited images, AI-altered photographs, and AI-generated illustrations on news pages. Feature or opinion sections may request a case-by-case exception from the editor-in-chief; if granted, the image must carry a visible label: "Image generated by AI." Stock photographs and file photos remain governed by existing photo credit policies.

2.5 Confidential material. No confidential source information, unpublished documents, sealed court records, grand jury material, off-the-record statements, juvenile case details, or any material received under an agreement of confidentiality may be entered into any AI tool. AI services may retain, log, or use input data for training. Treat every AI tool as a public channel. If you would not post the information on social media, do not paste it into an AI prompt.

2.6 Editorial judgment. AI may not determine news value, story placement, source selection, angle, or editorial priorities. These decisions belong to editors and reporters. AI may inform your judgment — a data analysis that reveals a pattern is useful input — but the editorial decision about what to cover, how to cover it, and what to emphasise is yours, not a model's.

2.7 Source identification and contact. AI may not be used to identify anonymous or pseudonymous sources, to find personal contact information for private citizens, or to compile dossiers on individuals who are not public figures. Using AI to find a public official's office contact information is fine. Using AI to find a whistleblower's home address is a firing offence.

---

3. Disclosure Requirements

Disclosure has three layers. The first two are always required. The third depends on the situation.

3.1 Internal disclosure (always required). Any reporter who uses AI in the reporting or writing process must note this in the story budget spreadsheet in the "Notes" column. The note should state which tool was used and for what task.

Examples of adequate internal disclosure:

• "Used Claude to transcribe 45-minute city council recording. Verified all quotes against audio."
• "Used AI to summarise 200-page zoning variance application for background. Key facts verified against original document."
• "Used AI to identify trends in five years of school test-score data. All figures checked against state database."

Examples of inadequate internal disclosure:

• "Used AI." (Too vague. For what?)
• "AI-assisted." (Same problem.)

3.2 Editor disclosure (always required). The assigning editor must be informed of AI use before the story is filed — or at the latest, when the story is filed. This is a conversation, not just a note in the spreadsheet. The editor needs to know: What tool was used? For what task? What output did it produce? How was that output verified? The editor may ask follow-up questions and may request additional verification steps before clearing the story for publication.

3.3 Reader disclosure (situational). Reader-facing disclosure is required when AI played a substantive role in producing the published content. Substantive means the AI output shaped the story's findings, structure, or factual claims — not just that a reporter used AI to transcribe a recording.

Triggers for reader disclosure:

• AI-assisted data analysis produced findings that drive the story or its headline
• AI-generated text appears in a non-news context (summaries, teasers, social media posts) under the Gazette's name
• AI was used to translate a document or statement that is quoted in the story

Reader disclosure format: a brief editor's note at the bottom of the story, before the reporter's contact information.

Example: "Data analysis in this report was assisted by AI tools. All findings were independently verified by Gazette reporters using original public records."

Reader disclosure is not required for routine transcription, grammar checks, background research that informed (but did not produce) the story, or internal communications.

3.4 Source disclosure (situational). If AI tools are used to process information provided by a source — transcription, translation, summarisation of interview notes — and the source has a reasonable expectation of confidentiality, the reporter must inform the source that AI tools were used and describe what information was entered.

This is especially important for:

• Sources in sensitive stories (whistleblowers, personnel disputes, legal matters)
• Sources who provided information off the record or on background
• Sources in the courts beat where case details may have legal protections

When in doubt about whether source disclosure is needed, ask the editor-in-chief before entering the material into any AI tool.

---

4. Quality Gates

4.1 Standard review. All AI-assisted content is reviewed by the assigning editor before publication, using the same editorial process as any other story. In addition to the usual editorial checks, the editor confirms:

• All facts are independently verified against primary sources — not against the AI output
• All direct quotes are taken from original recordings or reporter's notes, not from AI transcription without verification
• The story reads as if written by the reporter. If any passage sounds generated rather than written, the editor sends it back for rewriting
• The AI disclosure in the story budget is specific and complete

4.2 Flagging in workflow. When filing a story that involved AI assistance, the reporter adds the tag "AI-assisted" in the WordPress CMS. This tag is visible to editors and is not published. It serves as an internal flag so the copy desk and section editors know to apply the quality checks in 4.1.

4.3 Verification standard for AI-assisted data analysis. When AI is used to analyse a dataset and the analysis produces findings that appear in the story, the reporter must:

• Document the exact prompt or query used
• Run the analysis a second time to check for consistent results
• Verify key findings manually against the original data (spot-check at minimum; full verification for figures that appear in the story)
• Save the original dataset, the AI output, and the verification notes in the story file

4.4 Spot checks. The editor-in-chief will review a random sample of AI-tagged stories each month. The review checks whether disclosure was complete, verification was documented, and the published story meets the standards above. These checks are calibration, not investigation. If patterns emerge — for instance, AI-generated phrasing surviving into published copy, or disclosure notes that are consistently vague — the newsroom addresses them through the training process in Section 5, not through discipline.

---

5. Training and Onboarding

5.1 Initial briefing. Every current staff member will attend a 30-minute briefing on this policy within two weeks of its effective date. The briefing covers:

• A walkthrough of every section, with examples of permitted and prohibited uses
• A live demonstration of the disclosure process (how to tag stories in WordPress, what to write in the story budget, how to have the editor conversation)
• Common grey areas and how to resolve them
• Time for questions

5.2 Quick-reference card. A one-page summary of the quick-reference table (see end of this document) will be printed and posted near each workstation. The card lists permitted and prohibited uses with no further explanation — for full context, staff refer to the complete policy.

5.3 New hires. This policy is included in the onboarding packet for every new editorial hire. New reporters review the policy with their assigning editor during their first week and complete a brief self-assessment: read three scenarios, determine whether each is permitted, prohibited, or requires disclosure, and discuss the answers with the editor.