News Headline Edits and Bias Annotations from Five Media Outlets

78,910 headline pairs document original and edited versions from five English-language news outlets, annotated with 13 categories of media bias. The companion MediaSpin-in-the-Wild dataset links revised headlines to 180,786 tweets from 819 consenting users. This resource was created by Preetika Verma and Kokil Jaidka for research presented at AAAI-ICWSM 2026.

Use Cases

• Train classifiers to detect specific bias categories like sensationalism or opinion-as-fact within the 13 annotated types.
• Analyze patterns in headline edits across the five outlets to study editorial framing and spin.
• Model the relationship between headline revisions and downstream engagement using the 180,786 tweets in the MediaSpin-in-the-Wild companion data.
• Investigate correlations between bias annotations, such as omission and word choice, and social media sharing behavior.

Strengths

• Contains 78,910 headline pairs, providing a substantial corpus for analysis of editorial changes.
• Features fine-grained annotations across 13 distinct bias categories, including spin, omission, and sensationalism.
• Includes a companion dataset of 180,786 tweets from 819 users, linking headline edits to real-world engagement.

Limitations

• Limited to five English-language outlets, which may not represent the global media landscape.
• The tweet data is sourced from 819 consenting users, which is a relatively small and potentially non-representative sample of social media activity.
• The specific columns, file formats, and data structure are not detailed in the provided input.

Provenance

Source

Harvard Dataverse, accompanying the research of Verma & Jaidka (2026).

Collection Method

Collection of headline pairs from news outlets and annotation with bias categories; companion tweet data gathered from consenting X (Twitter) users.

Freshness

Data was last updated in March 2026.

Geography

Data from English-language news outlets; specific geographic scope not specified.