Counts words that foreground opportunity, growth, improvement or creation.
Cognitive Bias Observatory
Measure the information environment, not the person.
The Observatory tracks visible cognitive-pressure signals in headlines: gain and loss framing, certainty, salience, authority cues, social proof, numerical emphasis and repeated wording. It does not label an outlet or author as biased.
What we measure
Small signals with explicit rules.
Counts words that foreground losses, cuts, decline, threat or damage.
Counts modal or questioning language that keeps a claim open rather than presenting it as settled.
Counts language that presents an outcome or relationship with stronger certainty.
Counts vivid or high-intensity words that can make a headline more attention-grabbing.
Counts references that foreground institutions, experts, studies, reports or named authority statements.
Counts language emphasizing popularity, majorities, widespread behavior or crowd adoption.
Counts explicit numerals, percentages, currencies and number-like tokens in the headline.
Marks headlines ending in a question mark. A question can signal uncertainty, curiosity or rhetorical framing, but the observable fact is only the form.
Groups headlines that share a high proportion of content words after stop-word removal.
Snapshots
Public observations with provenance.
Each snapshot records the query, provider, time window and limitations. The launch snapshot is a curated method demonstration; scheduled GDELT snapshots can follow without changing the analysis contract.
AI and work · 2026-08-19
AI and work: framing pilot
This six-item pilot was deliberately selected to demonstrate contrasting headline frames. It is not a representative sample and must not be used to estimate media prevalence.
6 records · 5 domains
Latest snapshot
AI and work: framing pilot
This six-item pilot was deliberately selected to demonstrate contrasting headline frames. It is not a representative sample and must not be used to estimate media prevalence.
Open data
Built for people, researchers and agents.
Download the normalized observations as JSON or NDJSON. Each row keeps source provenance and the raw cue counts so another system can reproduce or challenge the derived labels.
Longitudinal layer
From snapshots to comparable trends.
Follow the same measurement rules over repeated samples from the same provider and sampling mode. The Trends layer normalizes signals per headline and refuses to publish a direction before enough comparable history exists.
Trend dashboard
4 tracked topics with minimum-history and comparability rules.
Source comparisons
Descriptive source profiles appear only after 5 unique headlines.
Research briefs
When repeated snapshots become useful evidence.
The research layer waits for six comparable snapshots and enough source breadth before publishing a longitudinal summary.