AI use policy

Where we use AI

• Bias framings. For each canonical event we generate up to six perspective-framed summaries (neutral, Western, Iranian, Israeli, Global South, pro-peace). The model is given the underlying source list and asked to write what each perspective would emphasise. The framings are interpretive analysis, not independent verification.
• Daily briefing drafts.The daily executive summary is drafted by an LLM from the day's event records, then reviewed.
• Topic classification. Incoming articles are tagged by event type (strike, diplomacy, economic, humanitarian, etc.) by a lightweight LLM call. A misclassification only affects filtering, not verification status.
• De-duplication. Title similarity for clustering near-duplicate articles uses non-AI string-similarity heuristics. Where embeddings are used, they run server-side on the ingestion pipeline only.

Where we do not use AI

• Verification.A claim's status (Verified, Partially Verified, Unverified, Disputed) is not set by an LLM. It is derived from the source list, independent corroboration, and editorial review.
• Casualty figures. Numbers are taken from named sources (health ministries, OCHA, ACLED). They are not generated, smoothed, or imputed by an AI.
• Source reliability labels. Source profiles (state-affiliated, independent, primary, etc.) are maintained editorially.
• Live user queries. The site does not call a paid LLM in response to a user clicking a page. AI work runs on a schedule, server-side, and the results are cached in the database.
• Defamatory or unsupported claims about named individuals or organisations.

Human review expectations

• High-impact items (verification upgrades, peace proposal status changes, casualty escalations) are reviewed by a human before they appear on the homepage.
• Bias framings are spot-checked. They are not individually approved before publishing because the source material is already public and the framings are clearly labelled as perspective summaries.
• Background ingestion (article fetch, dedup, classification) is reviewed by metrics (provider health, anomaly counts) rather than per-article approval.

Generated perspective rewrites — what to know

• They are notwhat a Western, Iranian, Israeli, Global South, or pro-peace newsroom would actually publish on a given day. They are a model's characterisation of how each perspective tends to frame this type of event.
• They make caricature errors. A "Western" framing is a generalisation; a specific outlet's framing of a specific story will differ.
• They do not change what is verified. If the neutral summary says a claim is disputed, every perspective rewrite respects that.

Data provenance

• The data fed to the model is the canonical event record and its source list. No classified, private, or non-public material is used.
• We do not use AI-generated content to train other AI systems beyond the internal classification models needed to run the site.
• Where a model returns an obvious failure (refusal, hallucinated source, off-topic text) the result is rejected by quality gates and not published.

Error handling

When an AI-generated artefact is found to be wrong (mischaracterised framing, inaccurate summary), the fix is to regenerate from the corrected source data. The correction is logged on the affected event page. If the error reflects a systematic weakness in our prompt or model, we update the prompt and re-run.

What this means for readers

• Treat bias framings as analysis, not reporting.
• Treat verification status as the reliable signal — that is editorial, not generated.
• Read the source list. It is on every event page for a reason.
• Flag obvious mischaracterisations via the corrections page. We'd rather hear about it.