Explaining how Airwars documents all incidents of alleged civilian harm from U.S. strikes in Yemen
This note is intended to accompany the data and findings following our publication of all alleged US actions in Yemen, beginning under the first Obama administration (January 2009) and running through May 2025 with the end of ‘Operation Rough Rider’ which was carried out under the second Trump administration. All of these assessments are publicly available in our archive.
Airwars published a methodology note in 2023, coinciding with the publication of all alleged incidents carried out by US Forces in Yemen between 2009 and 2013. Since then, Airwars has published all alleged incidents of civilian harm carried out under the second Obama administration, first Trump administration, Biden administration, and – as of June 2025 – the second Trump administration. Airwars has also updated a number of incidents published under previous administrations. The logic behind these updates is outlined below.
In 2002, US Forces carried out its first strike in pursuit of an alleged AQAP target. In 2009, under the first Obama administration, the US began to do so systematically. At the time of writing in June 2025, the US is continuing to carry out such strikes.
Airwars has researched and published all incidents where civilians were allegedly harmed in US strikes on alleged AQAP targets. With the exception of incidents recorded under the second Obama and Trump administrations, Airwars has also published all incidents where only AQAP militants, and no civilians, were allegedly harmed.
On February 2, 2024, under the Biden administration, US Forces carried out the first US strike on an alleged Houthi target in Yemen that resulted in civilian harm. Since then, Airwars has researched and published all incidents where civilians were allegedly harmed in US strikes on alleged Houthi targets.
Airwars continues to monitor all allegations of civilian harm from US actions in Yemen but is not recording and researching incidents where the only recorded casualties were Houthi militants.
Airwars uses an incident-based approach to document alleged US strikes in Yemen: each incident is defined as a moment in time and space where sources alleged US involvement in some form of kinetic violence (ie., airstrike or drone strike) that led to the fatality or injury of civilians.
Where the exact time of an incident is unknown, a general term, like “during the day” or “at night” may be used to mark the time of the incident. Local sources marked as “urgent” may also be used to determine the approximate time of an incident. Where multiple incidents occurred on a given day and in a general area, but sources regarding the exact time and location of the individual incidents are not available, deaths and injuries may be aggregated under one event until more information comes to light.
Each incident is geolocated to the highest possible degree of accuracy by trained geolocation teams. Incidents assessed under the second Trump administration included extensive related imagery, making those incidents more consistently geolocatable than those documented under previous administrations, where geolocation was challenged by a lack of available related imagery.
Where locations cannot be exactly identified, incidents with the same general time and location information may be aggregated until more information is known. All incidents are considered ‘live’ in our archive, and may be updated and changed to account for evolving information. As of May 2025, when an incident was updated, a record of the update was systematically included in the assessment.
Since publishing the original US Forces in Yemen methodology note in 2023, Airwars has included additional coding to account for damage and destruction to infrastructure in cases where civilians were also alleged to have been killed or injured. This damage or destruction is recorded for an incident when described or captured by open sources.
In May 2025, Airwars reviewed all incidents of civilian harm in the US Forces in Yemen archive, and when relevant, updated incidents to include a record of the damage to/destruction of infrastructure.
In 2023, the definition of ‘infrastructure’ accounted for any mention or documentation of the following terms by sources: hospital, school, agriculture marketplaces, gas facility, power station, water station, religious place.
In 2025, the definition of ‘infrastructure’ was updated to include mention or documentation of infrastructure that falls under the broad categories of: agriculture, healthcare, emergency services, humanitarian, education, commerce, legal, financial, bank, warehouse, religious, residential, transportation, vehicle, media, communications, resources. Within each of these categories are subcategories that can also be marked. A complete record of all subcategories is available upon request.
The term ‘strike’ is used throughout this document and in our analysis to mean a kinetic action; each assessment further classifies this action depending on the level of detail provided by sources relating to the incident – for example, a naval bombardment, airstrike or drone strike.
The ‘strike status’ variable includes incidents recorded as ‘contested’, a ‘single source claim’, ‘likely’, and ‘declared’, and assesses the likelihood of the strike having been conducted by a certain named belligerent – in this case – US Forces. The parameters for each categorisation are outlined below.
Declared:
No US strikes conducted between 2002 and 2015 have been officially declared by either CENTCOM or by the CIA.
Due to the fact that US actions in Yemen have oscillated between CIA and US military action, the circumstances surrounding whether an incident has been coded as ‘declared’ depends on the period in which the alleged strike was carried out.
Likely:
The likely strike classification applies to incidents in which all of the sources reported on the incident attributed the strikes to US forces. In cases where all of the sources attribute the strike to the US military, and Yemeni officials have acknowledged to the media that the strike was carried out by the US, the strike is qualified as “likely” as well.
Single source claims:
Only one source was found with an allegation of harm from US forces.
Contested:
The contested strike classification includes incidents in which attribution of the strike to the US military was not agreed upon by all sources reporting on the incident.
This category has also been used to capture incidents using a broad inclusion criteria that reflects on the likelihood of US involvement, even if it was not explicitly mentioned by the sources. For strikes on alleged AQAP targets, this includes cases of precision strikes on a moving target during night time or drone strikes, which fit a wider pattern of US military engagement as well as the documented limitations of the Yemeni military to carry out s types of strikes. We have chosen this broad approach given the high level of secrecy around US actions, as well as findings from investigators and in Wikileaks around the Obama administration explicitly instructing the Yemeni government to take credit for US strikes.
Each case should be treated with caution and read closely in order to understand a ‘contested’ designation, which is outlined clearly within each assessment.
All published incidents in Yemen attributed to US Forces can be found on our public archive. Once in the archive, users can filter by date, strike type, causes of death/injury, strike status, civilian harm status, belligerent assessment, casualty demographics, and infrastructure. Users can also search for key terms included in the written body of the assessment.
Airwars’ analysis of incidents of alleged civilian harm in Yemen are filtered by strike status. View incidents pre-sorted by strike status, below.
Strike status, contested
Strike status, single source claim
Strike status, likely
Strike status, declared
In 2019, Airwars took over from The Bureau of Investigative Journalism (TBIJ) its monitoring of covert and clandestine US actions in Yemen, Somalia and Pakistan, which covered a nearly ten-year period between 2010 and 2019. This involved transitioning the TBIJ’s original assessments to meet the Airwars methodology. As such, many incidents documented between 2010 and 2019 in Airwars’ US Forces in Yemen archive originated from the TBIJ dataset.
TBIJ’s research also includes incidents where no fatalities – civilian or otherwise – were recorded, and focused predominantly on English-language media. Though Airwars has not otherwise investigated strikes without casualty allegations, these incidents have nevertheless been kept within the Airwars archive to preserve the original record, and include the following standard statement: “This incident was identified by the Bureau of Investigative Journalism and has been included in Airwars’ database even though no casualties are mentioned.”
In terms of additional intra-organisation cooperation, all incidents in the US Forces in Yemen archive that occurred prior to 2013 underwent a data mapping exercise in coordination with other civil society and documentation groups. This process allowed us to cross-check our database with existing and similar datasets, and when relevant, update our archive.
Airwars uses a classification and grading system within each incident to account for ambiguities and ensure that each incident best reflects what can be complex information environments.
When presenting our aggregate civilian harm figures in our reporting, we typically take the most conservative estimate of civilian harm, tallying the minimum civilian death toll across the incidents of concern. This means that while each incident reflects a range, defined by a minimum and maximum, unless otherwise stated, our given death toll will be the minimum figure.
Additionally, we filter our incidents when presenting the minimum range to reflect what our researchers estimate to be the most reliable minimum toll. These filters may vary between conflicts due to differences in the information environment, and will always be outlined as such in our reporting.
In some cases, Airwars presents the ‘total number of alleged deaths’. This number is taken across all incidents (including ‘contested’ cases), and also includes the maximum civilian harm figure.
For our Yemen research and reporting, unless otherwise specified we adopt the following filters when identifying a cumulative minimum civilian toll:
Whenever we present figures in a report, these totals should be considered a snapshot in time. Considering all our incidents ‘live’, we update individual cases when new information comes to light. As such there may be minor discrepancies between published results in static reports compared to the evolving tallies over time.
The Airwars report on Operation Rough Rider, the second Trump administration’s 2025 air campaign in Yemen, published in June 2025, compared the civilian harm allegedly committed by US Forces in Yemen under the second Trump administration to that committed by all previous US administrations. The full dataset on which the report is based is composed of all alleged incidents of civilian harm committed by US Forces in Yemen – ranging from the first reported US drone strike in Yemen (December 17, 2009) to the last reported incident of civilian harm occurring under Operation Rough Rider (May 5, 2025).
At the time of publishing the report (June 2025), Airwars had recorded eight incidents of civilian harm in Yemen whose strike status was a ‘single source claim’. All of these incidents, save for one (USYEM250505b), were allegedly carried out by administrations that preceded the second Trump administration.
Additionally, four incidents of civilian harm with a ‘contested’ civilian status were included in the chronological dataset under consideration (USYEMTr034-C, USYEMTr081-C, USYEMTr096-C, USYEMTr122-C).
Of the twelve incidents outlined above, eleven were recorded under US administrations that preceded the second Trump administration. As the key takeaway of this reprort is that harm to civilians carried out by US Forces in less than two months in Yemen was unprecedented, Airwars included these twelve incidents which fall outside of usual analytic parameters as they contribute to a ‘maximal’ understanding of civilian harm committed by previous US administrations – an understanding which is still nearly matched by alleged harm documented under the second Trump administration.
The full dataset on which the report is based can be viewed here.
Sources are identified by our trained team of Yemen researchers, using an incident-based method to develop a continuously evolving list of sources for initial monitoring of allegations of US strikes. This monitoring process is followed by a deep dive into incident-specific information across social media platforms including Facebook, Twitter/X, and Telegram. Looking at the table below, you can see the way in which the sources captured by our research teams evolved across the five administrations.
Originally building off the work of TBIJ, the majority of sources relating to alleged incidents under the first and second Obama administrations were English language sources. Under the following Biden and Trump administrations, our research was conducted primarily in Arabic, complimented by an English language research team that collected statements from the US government, officials, and military bodies (an explanation of such actors are outlined in the Strike status section, above).
Across the five administrations, we included any and all information relevant to a single incident of alleged US strikes that included mention of a casualty, regardless of the affiliation of the source. As such, publishing this methodology note in June 2025, we’ve included a column that outlines the extent to which sources collected under the Biden and second Trump administrations came from Al-Masirah, a Houthi-affiliated media organisation.
| US Administration | # incidents* | # sources | Average # of sources/ incident | Arabic language sources (%) | English language sources (%) | Facebook (%) | Twitter/X (%) | Other (%) | Al-Masirah (Houthi-affiliated) |
| Obama 1 | 131 | 2,606 | 20 | 34.2% | 65.5% | 0.8% | 41.4% | 57.8% | N/A |
| Obama 2 | 19 | 633 | 33 | 46.3% | 53.7% | 3.8% | 45.5% | 50.7% | N/A |
| Trump 1 | 232 | 4,437 | 19 | 60.4% | 39.5% | 2.2% | 54.9% | 42.9% | N/A |
| Biden | 27 | 386 | 14 | 82.0% | 18.4% | 29.5% | 37.6% | 32.9% | 2.1% |
| Trump 2 | 35 | 973 | 28 | 96.8% | 3.1% | 53.1% | 41.8% | 5.1% | 18.1% |
* Includes incidents where the only casualty was a militant
Due to a number of specific complexities in the information environment during Operation Rough Rider, the following section outlines in more detail the source profile as documented by Airwars during this campaign.
Over the course of the campaign, Airwars linked more than 900 sources to 35 civilian harm incidents. All but four incidents included corresponding video and imagery, which Airwars reviewed to corroborate harm claims. As outlined above, the majority of sources were identified through social media channels Facebook and Twitter/X. Al Masirah (Houthi-affiliated media) accounted for a large proportion of the sources, particularly because multiple different reports run by Al Masirah were linked to each incident (on average when Al Masirah covered a harm incident, Airwars identified five different Al Masirah related reports). See more on Houthi-backed media below.
Removing high profile outlying incidents which garnered extensive media coverage, (the Ras Isa incident – which was ‘declared’ by US Forces – and the Remand Detention Center incident), there were on average 23 sources per incident.
As in any conflict, images and videos are often widely recirculated online and social media platforms often strip images of metadata, meaning it is often impossible to verify the original source for visual media. To mitigate the inclusion of false or misleading images and videos, Airwars teams carry out a variety of checks, including reverse image searching, and use a grading system in all incidents to reflect any uncertainties in the information environment. Airwars also considers our archive ‘live’ – should new information come to light that contradicts the original assessment, Airwars will update the incident to reflect this.
All but four incidents included corresponding video and/or imagery to accompany harm claims. Through reverse-image searches, Airwars identified ten images/videos which were connected to other incidents or contexts. In all ten scenarios, the information connected to the irrelevant media was vague, generally referring to an airstrike in an unspecified area.
In Yemen, the Houthis seek to control media narratives and have cracked down on any human rights investigative practice and civil society more generally, which significantly limits independent reporting and the ability to triangulate Houthi-backed claims.
Just days into the US campaign, the Houthi Ministry of Information published a statement on social media (archived link) calling on Yemeni citizens to:
“1- Completely refrain from publishing names, photos, or any personal information about martyrs before the official statement is issued by the competent authorities.
2- Exercise extreme caution in circulating any unverified information, and rely only on official statements.
3- Focus online activity on highlighting the grievances of our people, their resilience, the heroism of our armed and security forces, and our steadfast position in defending the oppressed Palestinian people”.
Following this statement the Ministry of Information launched the “I don’t know” campaign – a widespread information campaign calling on the population not to share the details of strikes online and to stop using Elon Musk’s Starlink internet service.
While telling the population to refrain from spreading details about the strikes, the Houthis released more ‘approved’ content via their own media outlets. Al-Masirah, the most prominent media outlet, reported from strike sites after attacks and presented footage claiming to indicate US actions.
In all but five incidents of civilian harm documented by Airwars, Airwars identified reporting from Al-Masirah in connection to harm claims. This includes reporting recirculated by news aggregators, where information was not directly cited as having originated from Al-Masirah, but in recirculating images or other media the Al-Masirah watermarks were often evident in the upper right hand corner.
As in many contexts in which Airwars works, allegations of harm that originate from media accounts linked to parties of the conflict does not automatically mean that the allegations themselves are false, though as with all open source content they should be treated with caution.
Despite this controlled environment, Airwars both found sources that diverged from Houthi interests (reporting harm from Houthi misfires, for example), and was able to triangulate Houthi-backed media reports with information from third parties. In 13 out of 33 civilian harm cases, this included sources likely directly connected to those harmed – in other words testimony from relatives, friends or neighbours posting about people they knew harmed in the attack – which corroborated other reporting on the incidents to support the original harm claims.
It is also unclear the extent to which the Houthi’s ‘I don’t know’ campaign was effective in limiting the spread of information, with popular Facebook accounts such as “Where was the explosion” sharing details of strike sites throughout the campaign, while other accounts shared throughout Yemeni social media appeared to poke fun at the Houthi messaging (on Facebook, TikTok, X, and Youtube).
Four of the 35 incidents included images of munition remnants, which experts at Airwars’ partner project the Open Source Munitions Portal were able to link to munitions held in the US arsenal. Images for two of the incidents came directly from an official Houthi-backed account and another came from the Houthi-affiliated Al-Masirah media outlet. The fourth image, connected to the strike on Remand Detention Center, was reshared by Al-Masirah but may have originally come from an individual user. However, that user’s post appears to have been taken down.