Whitepaper

AI and Editorial Integrity: A Framework for Newsrooms

Governance, oversight, and the limits of automation

Specification-stage draft

A practical framework for newsrooms adopting AI, with roles, risk tiers, review requirements, and incident response.

Draft framework

This whitepaper is a working draft. It is not legal advice and has not been reviewed by counsel. It is published so newsrooms can adapt it before adopting tools.

Executive summary

AI in the newsroom is usually discussed as a capability question: what can the tool do? The more important question is a governance question: who is accountable when it is wrong?

This framework proposes five principles, a risk-tiering method, explicit review requirements, and an incident response process. It is designed to be adopted before a pilot, not after an incident.

1. Five principles

1.1 Human accountability

Every AI-assisted output that reaches publication has a named human accountable for it. “The model decided” is not an accountability statement.

1.2 Transparency to readers

Readers should be able to tell when AI materially shaped content, especially synthetic media. Disclosure should be specific: what was generated, what was edited, and what was checked.

1.3 Evidence over confidence

A confident output is not evidence. Newsrooms should require the same evidentiary standard for AI-assisted claims as for any other claim.

1.4 Proportionality

The level of review should match the stakes. A headline suggestion and a synthetic-media verdict do not need the same process.

1.5 Contestability

Anyone affected by an AI-assisted finding should have a route to challenge it. Appeals must be reviewed by someone who did not make the original decision.

2. Roles

  • Editorial owner: accountable for standards and final publication decisions.
  • AI product owner: accountable for tool selection, configuration, and vendor management.
  • Reviewer: examines high-stakes outputs and records decisions.
  • Security and privacy lead: reviews data flows, retention, and access.
  • Oversight group: reviews incidents, appeals, and policy changes.

In small newsrooms, one person may hold several roles. The important thing is that the roles are named, not that they are separate.

3. Risk tiers

Tier 1: Low

Assistive tools with no publication effect, such as transcription or headline suggestions. Requirements: basic disclosure to staff, vendor review, and a feedback route.

Tier 2: Moderate

Tools that shape content but are reviewed by an editor, such as translation, summarisation, or assessment. Requirements: documented review, evidence visibility, and override logging.

Tier 3: High

Tools whose output can reach publication with limited review, or that make claims about people. Requirements: named reviewer, pre-publication evidence, appeal route, and incident plan.

Tier 4: Prohibited

Uses the newsroom will not support: automated publication without review, individual profiling, synthetic impersonation, or covert generation of quotes.

4. Review requirements

A review is only meaningful if the reviewer has four things:

  1. Information: the output, the evidence, the model version, and known limitations.
  2. Time: a queue that does not force rubber-stamping.
  3. Authority: the ability to override without escalation.
  4. Record: a logged decision with a reason.

Newsrooms should measure review time and override rates. A review process with a 99.9% acceptance rate and a two-second median review time is not a review.

5. Transparency and disclosure

Disclosure should answer three questions: Was AI used? For what? Was the output verified by a person?

A generic “AI was used in the production of this article” is better than nothing and worse than specificity. Where synthetic media is published, provenance metadata and a visible label are both recommended.

6. Vendor due diligence

  • What is the intended purpose, and what is explicitly out of scope?
  • What data is processed, where, and for how long?
  • Is customer content used for training? Under what terms?
  • What human oversight does the interface support?
  • What is logged, and can it be exported?
  • What happens on model change, and how is the customer notified?
  • What certifications exist, and what is merely planned?

7. Incident response

  1. Detect: a wrong output, a data leak, or a harmful publication is identified.
  2. Contain: pause the tool or the workflow; preserve logs.
  3. Assess: determine scope, affected content, and harm.
  4. Correct: publish a correction where needed; notify affected parties.
  5. Learn: write a post-incident review and update policy.

Incidents should be reviewed by the oversight group, not only by the team that owns the tool.

8. Metrics that matter

  • Override rate, by tool and content type.
  • Time to review, and queue depth.
  • Correction rate for AI-assisted content versus baseline.
  • Appeal volume and outcomes.
  • Incident count and time to resolution.

9. What not to do

  • Do not adopt a tool because a competitor did.
  • Do not treat a vendor’s compliance claim as evidence.
  • Do not use AI to replace a role you have not defined.
  • Do not publish synthetic media without a label and a provenance check.
  • Do not measure individual journalists with a noisy metric.

10. Conclusion

Editorial integrity is not preserved by refusing AI, nor by adopting it quickly. It is preserved by making accountability explicit at every step. The framework is deliberately unglamorous: names, records, reviews, and appeals. That is what makes it usable.

Download

Get the full report

The download will be available when the study is published. Leave your email and we will send it to you, with the method and limitations included.