Methodology

Newspectives takes a single news event and produces multiple regional and ideological readings of it side by side. This page documents how that analysis is generated, what the AI does and does not do, and how a reader can verify any claim it makes.

1. Topic intake

A topic enters the system one of three ways: a user-submitted query (a phrase or a URL pasted into the search bar), an admin-submitted topic, or the daily auto-topic job, which scans Google News at 09:00 UTC for the most significant globally-reported event of the day.

If the input is a URL, the system reads that source article first to establish a factual baseline. If the input is a phrase, the system uses Google Search grounding to gather context before any regional analysis runs.

2. Multi-agent dispatch

The engine dispatches one parallel AI agent per perspective on the topic — typically the default lineup, or a reader-selected subset. The default lineup includes Common Ground (neutral framing), USA, United Kingdom, Germany, Russia, China, India, Israel, Arab World, South Africa, Latin America, Humanitarian (focused on civilian impact), and The Jester / Exospective (explicitly satirical commentary). Additional regional lenses (France, Denmark, Turkey, Iran, Japan, North and South Korea, Canada, Australia, Taiwan, the Philippines) are configured and can be activated, supporting up to roughly twenty perspectives per topic. Each agent receives:

• The topic and its baseline context
• A system instruction describing the framing priorities of its assigned lens — what that perspective tends to emphasize, what kinds of outlets typify it, what historical framings it draws from
• Live access to Google Search to ground the analysis in real reporting from the relevant timeframe

Model and provider. Every perspective is generated by Google Gemini 3 Flash via the Google Generative AI API, with Google Search grounding enabled to anchor each claim in real-time reporting. Topic illustrations are generated by Google Imagen 4. Audio summaries and radio broadcasts use Google Gemini Native Audio. Music tracks use the Suno API. The underlying LLM may be updated to newer Gemini releases as they ship; significant model changes are logged on the changelog.

3. Source grounding

Each regional agent retrieves verified articles matching the timeframe of the event. The model is prompted to ground every claim in actual published reporting from that period — not training data alone. The list of source URLs is preserved alongside each perspective.

The sources we sample tend toward each region's mainstream and state-influenced outlets — for example, BBC and The Guardian for the UK lens; Global Times for the China lens; RT and TASS for the Russia lens; Al Jazeera for the Arab World lens; The New York Times, CNN, and the Wall Street Journal for the USA lens. We deliberately include state-influenced sources where they shape the narrative in a given region, because the goal of the platform is to surface how that region is framing the story, not to filter for outlets that share our values.

4. Synthesis & tone classification

The AI extracts a representative regional headline, writes a 50-70 word summary, identifies key points of agreement or contradiction across perspectives, and assigns a tone classification (Optimistic, Critical, Analytical, Diplomatic, Cautious, and others) to help readers see the framing choices at a glance.

5. Common Ground

The Common Ground perspective is generated separately and intentionally last. Its job is to identify what is uncontested across the regional readings — the dates, names, numbers, and events that everyone agrees on, regardless of framing. It is meant as a factual anchor, not as the "right" perspective.

6. What the AI does NOT do

• No investigative reporting. Newspectives synthesizes existing coverage. It does not break new facts, conduct interviews, or publish original journalism.
• No real-time fact-checking. The AI grounds claims in published reporting, but it does not independently verify the accuracy of those reports. If multiple outlets are wrong, the synthesis will be too.
• No predictions or recommendations. Perspectives describe how a region is framing an event. They are not forecasts and they are not advice.
• No editing of the underlying source material. The AI summarizes; it does not rewrite the articles it links to.

7. Known limitations

Large language models occasionally produce inaccuracies, hallucinated quotes, or misattributions. The system retries on detected formatting errors, but it cannot detect every subtle factual problem. The Common Ground summary is the most likely to be reliable; the regional perspectives are intentionally partisan in framing and should be read as "this is how the region is reporting it" rather than "this is what happened."

Every claim in every perspective links back to the source URLs the AI consulted. Readers are expected to verify against those sources when accuracy matters.

8. Human oversight

The platform is operated by Hein Kleinveld (founder, Amsterdam). Human review focuses on system-level prompts, the list of sampled outlets per region, brand-voice rules, and corrections to specific topics when factual errors are reported. Individual perspective summaries are generated and published by the AI without per-topic human editing.

9. Corrections

If you spot a factual error, report it to info@newspectives.com. Corrections will be made and logged on the corrections page. See our editorial standards for the principles that guide every analysis.

Frequently asked