Results — Event Type vs Victim Identity (Stage 1)

Navigation

Hub: 00 - Fact-Check MOC · Previous: 03 - Test Design · Next: 05 - Findings Register

98 GDELT queries across seven events. A mixed result, and an explicit statement of what it cannot settle.

98 queries, CNN and MSNBC, seven-day windows, zero errors. Metric is share of that event’s victim-group coverage containing each term, computed as summed percentage-of-airtime for the term query divided by the same for the victim-group baseline.

Conditional rates

Absolute clip counts

Massacre and slaughter terms, seven days, CNN plus MSNBC. Capped at 50 per station, so capped cells are floors.

Reading

The two measures point different directions, and the honest answer is that neither hypothesis wins outright.

The conditional rate does not support a victim-identity effect. Among close-range events, massacre-plus-slaughter share falls in a narrow band regardless of victim group: 1.13 percent for October 7, 0.86 for Nuseirat, 0.79 for the flour convoy, 0.72 for Bucha. Conditional on an event being discussed at all, the vocabulary appears at broadly similar rates. Taken alone this cuts against the book.

The absolute counts show a large gap — roughly 100 and 86 clips for the Israeli and Ukrainian close-range events against 16 and 15 for the Palestinian ones. But this is substantially a coverage-volume effect: the flour convoy and Nuseirat drew base coverage of 979 and 825 against October 7’s 11,875, a difference of more than an order of magnitude. Attention decay is a plausible non-bias explanation, since those events fall five and eight months into the war.

Mariupol is the awkward case for a simple victim-identity story

Mariupol, an aerial attack on Ukrainian civilians, drew the lowest massacre rate in the entire set at 0.05 percent — well below Jabalia and al-Ahli, both aerial attacks on Palestinian civilians. If media vocabulary tracked victim identity in a straightforward way this should not happen. The aerial comparison runs opposite to the close-range one.

The strongest signal is qualitative, not quantitative. Inspecting the actual clips from the Nuseirat window, the massacre usages are attributed rather than spoken in the network’s own voice: coverage reports that Hamas called it a massacre, and that a diplomat called it a massacre. By contrast the October 7 clips include substantial editorial-voice usage — descriptions of a thousand Israelis massacred, of a music festival massacred, of a horrific massacre of almost a hundred Israelis, stated as narration rather than attributed to a source.

This matters because Johnson’s counts explicitly exclude quoted usage and count only editorial language. If that pattern holds across a coded sample, his near-zero figure for Palestinian victims is measuring something real that my unattributed query cannot see — and the conditional-rate similarity above would be an artifact of counting attributed and editorial usage together.

What this settles and what it does not

Settled: the event-type objection is not sufficient on its own. Bucha is close-range killing of civilians by a state military, five weeks into a war rather than on day one, and it drew massacre-language at rates and volumes comparable to October 7. Violence form alone does not explain the gap.

Not settled: whether the residual gap reflects editorial bias, coverage-volume decay, or the attributed-versus-editorial distinction. Stage 1 cannot separate these because the query cannot tell who the term refers to or whether it was spoken in the network’s voice.

Required next step

Manual coding. Export clips for each event with VICTIM_CODE and IS_QUOTE columns, code at least 100 per event, then recompute editorial-voice-only rates. That is the only version of this measurement comparable to Johnson’s, and until it is done neither the conditional-rate finding nor the absolute-count finding should be reported as a result.

Secondary: raise the clip cap above 50 per station via pagination so October 7 and Bucha yield exact counts rather than floors.