Open dataset · CC BY 4.0

How AI models read the news

For each Satcove News story, several AI models read the same newsroom reports separately. We record what each model covered, what it left out and which claims the reports did not support. This page aggregates those scorecards and publishes the raw data.

What this is not

It measures how a model read the sources, not whether it is accurate in general or whether a claim is true. With 2 stories so far, differences between models are not statistically meaningful. The dataset grows with each published story.

Per model, so far

2 scored stories, 2026-10-15 to 2026-10-15.

ModelStoriesMean coverageCritical omissionsUnsupported claimsContradictions
Claude286%210
GPT259%200
Gemini282%000
Mistral278%210
Perplexity278%000
Grok282%000

Stories in the dataset

Method

  • A story is published only when at least three recognised newsrooms report it. Each model receives the same extracted reports.
  • Models answer independently and do not see each other's output. Models that fail or time out are recorded as skipped, not hidden.
  • Coverage is the share of the panel's points that a model included, within the number of claims it was asked for.
  • A claim counts as unsupported when its supporting quote is not found in the reports. Quotes are matched verbatim by a script.
  • The stories are automated analyses. Treat them as a reading aid and check the linked reports before relying on a fact.

Download and cite

data.json

Licence: CC BY 4.0. Please credit Satcove and link to this page.

Satcove. Satcove News dataset, version 1.0. https://satcove.com/news/data

Back to Satcove News