Anyone following the Gaza war through the news noticed a pattern: headlines about strikes and casualties named Israel as the actor far more often than Hamas. The plain question is whether that reflects something about coverage convention, or simply tracks who was conducting the depicted operations. We coded 1,976 Gaza-war headlines from 11 major outlets for who, if anyone, the headline names as the grammatical agent of a violent act. Only 19.3% of those headlines depict a violent act at all; among the 292 that name any actor, Israel is the named agent 79.8% of the time, roughly 7.8 times as often as Hamas or Palestinian militants, a share that substantially tracks the real-world base rate of who was conducting the depicted operations and that this study cannot separate from coverage convention.
Key Findings
- Only 19.3% of the 1,976 headlines depict a violent act (381 headlines); the remaining 80.7% report the war without naming a depicted act of violence (Hanover Institute headline corpus, 2026, first-party study).
- Among the 292 headlines that name any actor, Israel is the named agent in 79.8%, against 10.3% for Hamas or Palestinian militants, about 7.8 times as often. The period’s military phase was predominantly Israeli operations in Gaza, so this share substantially tracks the real-world base rate and is not a measure of bias.
- 23.4% of violent-act headlines (89 of 381) name no actor at all, and roughly half of those agentless headlines, 50.6%, describe a death or casualties with the actor removed from the sentence.
- The words differ by actor: Israel-named headlines take active violence verbs (strike or bomb in 26.2%, an active form of kill in 21.9%), while Hamas-named headlines lean on actor nouns (militant or fighter in 13.3%), a descriptive framing contrast measured on small counts.
- The pattern holds across all 11 outlets: every outlet names Israel as the majority agent and every outlet has a meaningful agentless share, a consistency finding, not a ranking of outlets.
What do the coded headlines look like?
Each headline was assigned one of six labels from its own text, not from outside knowledge of who acted. The label records whom the sentence names as the agent of a violent act, or marks the headline as naming no actor (agentless) or as not depicting a violent act at all (none). One representative headline from the corpus for each label:
| Label | What it marks | Example headline from the corpus |
|---|---|---|
| Israel | Israel named as the actor | “Israeli airstrikes kill at least 72 across Gaza” (The Guardian) |
| Hamas or Palestinian | Hamas or Palestinian militants named as the actor | “Hamas Attack Raises Questions Over an Israeli Intelligence Failure” (The New York Times) |
| Other adversary | Another named state or armed group as the actor | “Iran-backed Iraqi militia attacks Haifa, Israel with drones” (Fox News) |
| Both sides | Each side named as the actor of its own act | “Hezbollah and Israel exchange fire and warnings of a widened war” (The Associated Press) |
| Agentless | A violent act or casualties with no actor named | “In Gaza, Authorities Lose Count of the Dead” (The Wall Street Journal) |
| None | Not a depicted violent act | “Gaza Cease-Fire Talks Enter Precarious Moment” (The Wall Street Journal) |
The none category is the largest because it captures everything that is not a depicted act: ceasefire talks, diplomacy, aid, accusations and court cases, and analysis. An accusation is coded none rather than as an act by the accused, since a headline that an actor faces a charge reports a claim, not a depicted act. Headlines are quoted verbatim, with the outlet in parentheses.
How often do Gaza-war headlines depict an act of violence?
Most do not. Across 1,976 headlines, 19.3% (381 headlines) were coded as depicting a violent act, and 80.7% (1,595) were not (Hanover Institute headline corpus, 2026). The 80.7% is not a count of peaceful news: it includes diplomacy, ceasefire talks, politics, aid, displacement, hostage negotiations, and analysis, anything the headline frames without naming a specific violent act. The figures below all describe the smaller depicting-violence subset, so every share that follows is computed within those 381 headlines, not the full corpus.
When a headline depicts violence, who is named as the actor?
When a violent-act headline names an actor, that actor is Israel far more often than any other. Of the 292 headlines that name any agent, 79.8% name Israel, against 10.3% for Hamas or Palestinian militants, 6.2% for another adversary (Iran, Hezbollah, or the Houthis), and 3.8% for both sides, a ratio of about 7.8 to 1 between Israel and Hamas. This share substantially tracks the real-world base rate: the period’s sustained military phase was predominantly Israeli operations inside Gaza, so the language of attribution largely follows who was conducting the depicted operations. The study measures the attribution language, not whether it matches ground truth, which it cannot assess; this is not a finding of media bias in any direction.
How often is the actor removed from the sentence?
Often enough that it is the second-largest category. Of the 381 headlines depicting a violent act, 23.4% (89 headlines) name no actor at all, rendering the violence as an event without a grammatical agent, a death that occurs, a building that is hit, a strike that lands. A corpus example is “Gaza university blown up in massive explosion” (BBC), where a building is destroyed but no actor performs the action. Among those 89 agentless headlines, 50.6% describe a death or casualties, the most common kind of agentless construction, followed by bombing or shelling (12.4%), attack (10.1%), strike (7.9%), and hostage or captive (4.5%). About half of the headlines that depict violence without an actor describe a death or casualties with the actor removed; the terms are not exclusive, so a single headline can contain more than one.
Which words attach to each named actor?
Different parts of speech attach to each side. In the 233 headlines naming Israel as the agent, the dominant words are active violence verbs: a strike or bomb term appears in 26.2% and an active form of kill in 21.9%, while actor nouns are nearly absent (militant or fighter in 1.7%, terrorist in 0%). In the 30 headlines naming Hamas or Palestinian militants, the pattern inverts: the words are mostly actor nouns (militant or fighter in 13.3%, terrorist in 3.3%), with active violence verbs rare (strike or bomb in 3.3%, active kill in 0%). The specimen headlines above show the contrast directly: the Israel-named example leads with an active verb (airstrikes kill), while the Hamas-named example carries an actor noun (attack). This is a descriptive contrast in grammar, verbs for one named actor and nouns for the other, measured on a small Hamas-named base of 30 headlines, and much of that base sits in the October 7 attack frame at the start of the window. The terms are not exclusive and the counts are small, so the contrast describes a pattern in the data, not a stable rate.
Did the framing shift as the war went on?
It shifted alongside the events of the period. Among headlines naming an actor, Israel’s share rose from 60.7% in the fourth quarter of 2023 to a range of 76% to 92.2% across 2024, while Hamas’s share fell from 24.6% to as low as 1.3% before recovering to 16.7% in the first quarter of 2025. In October and November 2023 Hamas was 26.2% of named agents, the October 7 attack frame; from December 2023 on its share fell to 7.6%. As the war moved from the October 7 attack into sustained Israeli operations in Gaza, the named agent shifted toward Israel. This tracks the events of the period, not a coverage verdict, and the study cannot separate the two; the trend describes how the agent language moved alongside the phase of the war, not why any outlet wrote it that way.
Does the pattern hold across outlets?
It holds across every outlet in the sample. All 11 outlets name Israel as the majority of their named agents, from a low of 50% to a high of 94.4%, and all 11 carry a meaningful agentless share, from 11.8% to 32.6% of their violent-act headlines. The outlets cluster in the same region of the plot rather than splitting into distinct camps, which makes this a consistency finding: the broad structure (Israel as the most-named agent, a substantial share of headlines with no agent) recurs regardless of outlet. This is descriptive placement, not a ranking; the figures sit within small per-outlet subsamples of roughly 180 headlines each, the differences between outlets are not tested for significance, and nothing here rates any outlet against another.
Methodology and limitations
The corpus is 1,976 newspaper headlines drawn from 11 outlets (The New York Times, The Washington Post, The Wall Street Journal, the Associated Press, the BBC, CNN, Fox News, The Guardian, Al Jazeera, Reuters, and NPR), roughly 180 per outlet, sampled across 18 monthly windows from October 2023 to March 2025. Each headline was coded to exactly one label from a six-label codebook for the agent of violence: none (not a depicted violent act), israel, hamas_palestinian, other_adversary (Iran, Hezbollah, or the Houthis), agentless (a violent act with no actor named), and both. The governing rule is a surface-grammar rule: the label comes from the headline text alone, from who is named as the grammatical agent of a violence verb, not from real-world knowledge of who carried out a given act. A headline that reports a death without a subject is coded agentless even when the wider event is well known.
The labels are a careful coding of every headline by a frontier model against the codebook. Reliability was measured directly: an independent second coding of a stratified 401-headline sample, spanning all 11 outlets with the rare labels and all six codebook traps deliberately oversampled, agreed with the original at a Cohen’s kappa of 0.855 (90.0% of headlines). That figure is a conservative floor, because the sample is weighted toward the hardest cases; agreement met or exceeded it inside every codebook trap, and the residual disagreement falls almost entirely on a single boundary, whether a casualty headline with no named actor reads as not a depicted act (none) or as a depicted act with the agent grammatically absent (agentless). The counts for some categories are small (30 Hamas-named headlines, 18 other-adversary, 11 both), so shares within them are indicative rather than stable.
One limit governs the central finding and is stated plainly: the study cannot separate a coverage convention from the event base rate. The period’s sustained military phase was predominantly Israeli operations inside Gaza, so the high share of headlines naming Israel substantially tracks the real-world base rate of who was conducting the depicted operations. The measurement is of the attribution language, not of whether that language matches ground truth, which it cannot assess. For the same reason the quarterly trend describes how the agent language moved alongside the phase of the war, not why it moved. The report measures the language, not the events.
This is a corpus-linguistics measurement of coverage language. It does not adjudicate the conflict, it does not rate or rank the outlets, and it makes no claim about bias in any direction. Antisemitism is out of scope for this measurement: the report measures the surface grammar of violence attribution in headlines and does not measure antisemitism or make any claim about it. It sits within the Institute’s broader study of the information environment around the conflict, where related work has measured the volume and structure of online conflict discourse (Chen and colleagues, 2024, preprint; Guerra and colleagues, 2024, preprint), a different object than the headline grammar measured here.
Conclusion
So does the grammar of these headlines say something about the coverage, or does it simply track who was doing the fighting? The corpus answers plainly: the pattern is real and it is substantially the base rate. Across the period’s depicted violence, Israel is the named actor, and the data cannot pull that apart from the fact that the period’s operations were predominantly Israeli.
A bounded portrait rises out of the 1,976 headlines all the same. Fewer than one in five depict a violent act at all, just 19.3%, so most of this coverage names no act of violence. Inside that smaller set, the attribution is lopsided: among the headlines that name anyone, Israel is the agent 79.8% of the time against 10.3% for Hamas or Palestinian militants, about 7.8 times as often. Nearly a quarter, 23.4%, name no actor at all, and about half of those describe a death or casualties with the agent removed from the sentence. The words sort by side, active violence verbs for one named actor (strike or bomb in 26.2% of Israel-named headlines) and actor nouns for the other (militant or fighter in 13.3% of the small Hamas-named set), and the Israel share rose from 60.7% in late 2023 toward the high 70s and 90s across 2024 as the war moved into sustained operations in Gaza. Every one of the 11 outlets names Israel as the majority of its agents, from 50% to 94.4%, so the structure is the corpus, not any single newsroom.
That regularity does not stay on the page. The image of Israel as the standing aggressor, and of Jewish actors more broadly as the party who acts upon others, travels well past any single headline, into the everyday talk about who does what to whom and into the contested ground where attitudes toward Israel and attitudes toward Jews are hard to separate. The Institute studies the inputs that shape how a group is seen, which is why a regularity this consistent, repeated across 11 outlets and 18 months, is worth measuring even when its cause is the base rate. So the open question is what a habit of grammar that names one side as the agent of violence four times out of five does to perception, when it recurs across 11 outlets for 18 months. This study holds antisemitism out of scope and measures only the surface grammar, the share, the verbs, the drift over time. What it does to perception plays out somewhere else: in the three seconds someone thumbs past a feed at a stoplight and reads that Israel struck, Israel killed, Israel hit, before the light changes and the phone goes back in a pocket.
Frequently Asked Questions
How were the headlines labeled?
Each of the 1,976 headlines was coded by a careful frontier-model pass against the codebook, from the headline text alone rather than outside knowledge of the event. The procedure’s reliability was checked by independently re-coding a stratified 401-headline sample drawn from all 11 outlets with the hard cases oversampled: the two passes agreed at a Cohen’s kappa of 0.855 (90% of headlines), with the residual disagreement concentrated on the boundary between none and agentless.
Does naming Israel more often mean coverage is biased?
No. The share of headlines naming Israel substantially tracks the real-world base rate: the period’s sustained military phase was predominantly Israeli operations inside Gaza, so the attribution language largely follows who was conducting the depicted operations. This study measures the language, not whether it matches ground truth, and it makes no claim of bias in any direction.
What does agentless mean?
Agentless means the headline depicts a violent act but names no one as the grammatical agent, for example a headline reporting that people were killed or a building was hit without a subject performing the action. In this corpus 23.4% of violent-act headlines are agentless, and about half of those describe a death or casualties.
Why are only 19% of the headlines about violence?
Because the coding counts only headlines that name a depicted violent act, and most war coverage does not. The other 80.7% report diplomacy, politics, aid, displacement, hostage negotiations, ceasefire talks, and analysis, real war news that simply does not render a specific violent act in the headline.
Did the framing change over the war?
Yes, alongside the events of the period. In October and November 2023 Hamas was 26.2% of named agents, the October 7 attack frame, and from December 2023 on that share fell to 7.6% as the named agent shifted toward Israel. The shift tracks the move from the October 7 attack into sustained Israeli operations in Gaza, and the study cannot separate that base-rate change from any change in coverage convention.
Why do the words differ between Israel and Hamas?
In this corpus, Israel-named headlines mostly take active violence verbs (strike, bomb, kill) while Hamas-named headlines mostly take actor nouns (militant, fighter, terrorist). The Institute reports this as a descriptive grammatical contrast, measured on a small Hamas-named base of 30 headlines that sits largely in the October 7 frame; it is a pattern in the data, not a stable rate or a judgment.
Does this report measure antisemitism?
No. This is a corpus-linguistics measurement of how headlines name the agent of violence, and antisemitism is out of scope for it. The report sits within the Institute’s broader study of the conflict’s information environment, but it makes no claim about antisemitism and does not measure it.
Can I tell which outlet is most balanced from this?
No, and the study is built not to support that read. Every outlet names Israel as the majority of its named agents and every outlet carries a meaningful agentless share, so the finding is consistency across outlets, not a ranking. The per-outlet samples are small (about 180 headlines each), the differences are not tested for significance, and the report rates no outlet against another.
Sources
- Chen, He, Burghardt, Zhang, Lerman, 2024. IsamasRed: A Public Dataset Tracking Reddit Discussions on the Israel-Hamas Conflict. arXiv:2401.08202. Preprint.
- Guerra, Lepre, Karakuş, 2024. Quantifying Extreme Opinions on Reddit Amidst the 2023 Israeli-Palestinian Conflict. arXiv:2412.10913. Preprint.
- Hanover Institute, 2026. Agent-of-Violence Coding of Gaza-War Headlines (headline corpus, 1,976 headlines, 11 outlets, October 2023 to March 2025). First-party study.