AI assists. Humans verify.

No reader-facing AI output ships without a named editor approver. This is a non-negotiable foundational rule. It is not a guideline. It is enforced at the level of our content management system: stories where AI-drafted content has not been signed off by a human editor cannot be published.

AI is used in five concrete operations at CERAP:

• Source clustering. When multiple outlets publish on the same event, AI groups them into a single “cluster” that an editor turns into one Story — enforcing our “one event, one Story” rule operationally.
• Drafting the 15-second summary. AI proposes the structured summary (what happened, why it matters, who’s affected); a named editor verifies and approves before publish.
• Proposing key points. Drafts 3–6 bullets from the source material. Editor verifies.
• Suggesting perspectives. AI may identify candidate perspectives present in the sources. Editor verifies each label and confirms attribution. Perspectives are brand-critical and the AI’s suggestions are treated as drafts to evaluate, not facts.
• Translating into other languages (planned). When CERAP launches Chinese and Malay editions, AI drafts the translation and an editor reviews before that language version goes live for News, Analysis, Perspective, and Fact Check formats. Explainer and Timeline formats may auto-publish for speed, with editor review post-publish.

The following are explicitly off-limits to AI at CERAP. Editors handle these directly:

• Issuing fact-check verdicts. The verdict on a contested claim (True, Misleading, False, Unverified) is always written and approved by a named editor with fact-checking authority. AI may surface the claim and flag inconsistencies; it does not adjudicate.
• Originating quotations. AI never writes a quotation. Quotations come from sources, verified by a reporter.
• Publishing autonomously. AI cannot trigger the publish action. That is reserved to authenticated editorial roles.
• Characterising stories with internal signals. Our AI tracks internal editorial signals like sentiment and misinformation risk, but these are never displayed as public labels on a Story. They are used by our recommendation engine and to prioritise fact-checker review only.

Every Story carries structured metadata that records which sections were AI-drafted, AI-assisted, or human-written, the approving editor, the AI model’s confidence score, and the timestamp of the last AI assist. This record is internal but auditable.

If a correction is later issued on AI-assisted content, the record allows us to trace back exactly which model produced what and which editor reviewed it — so we can learn from the error.

Before any News or Analysis Story is published, our AI runs a balance check that flags one-sided perspective sets. If the AI finds that all the perspectives represent only one viewpoint, publication is blocked until additional perspectives are added or a Senior Editor overrides with a written justification.

This makes our labelled-balance commitment from Editorial Policy not an aspiration but a system constraint.

As of June 2026, CERAP’s drafting and translation pipeline uses Anthropic Claude as the primary large-language model, chosen for its handling of multilingual editorial nuance and its safety behaviour around politically sensitive content.

We may change underlying models as the field evolves. The rules above — human verification, no autonomous publishing, tracked provenance — apply regardless of model.

If you suspect AI-drafted content on CERAP contains an error, or you have a question about our use of AI in a specific Story, email editorial@cerap.ai with the Story URL.