YouTube Narratives on the Al-Shifa Hospital Raid: Grasping the Debate around the Al-Shifa Hospital Attacks

Yu Hu, Yunuo Wang, Yuxiao Ye, Yuqi Zhang


This research examines narratives that unfold on YouTube regarding the Al-Shifa Hospital raid in Gaza carried out by the Israeli Defence Forces (IDF) on the 15th of November. This operation occurred amid the continuing conflict between Israel and Gaza that intensified on the 7th of October 2023, following a surprise attack by Hamas on Israel and Israeli citizens. After that, “Israel launched a major military campaign in the Gaza Strip with the objective of destroying Hamas - which it classes as a terrorist organization - in response to an unprecedented cross-border attack on 7 October” (Gritten, 2023). According to a news report from The Guardian, “Israeli soldiers were inside Gaza’s largest hospital on Wednesday after an early morning raid that drew fierce condemnation from the head of the World Health Organization, who called it totally unacceptable” (Sinmaz & Burke, 2023). Since then, the Israeli government has aimed to eliminate Hamas, a political militant organization that governs the Gaza Strip.

According to the IDF, Hamas fighters used the hospital for military purposes. Israeli forces carried out a targeted search at Shifa, Gaza's largest hospital, to find weapons belonging to Hamas. The military claimed to have discovered weapons and terrorist infrastructure at the hospital, and they reported the killing of several terrorists. Intense fighting and explosions occurred in the vicinity of the hospital, raising concerns about the well-being of patients and those seeking refuge. According to Omar Zaqout, an emergency room employee at al-Shifa, “More than 180 dead bodies are deteriorating and are still lying in the hospital’s yard,” (Staff, 2023). The Israel Defense Forces delivered medical supplies, incubators, and food. The World Health Organization expressed worries after losing contact with the hospital staff and emphasized concerns about the safety of patients and personnel. Israel has indicated a willingness to potentially suspend operations, process intelligence, and strengthen logistical support.

However, evidence to support this claim is contended by Hamas and humanitarian organizations. “A government spokesperson in Gaza, which is controlled by Hamas, described Israel’s advance into the hospital as a war crime, a moral crime and a crime against humanity” (Financial Times, n.d.). Given that YouTube remains the most widely used social media platform, with a global community actively engaging on the platform, we aim to explore the recent Al-Shifa Hospital incident as a focal point for our investigation due to its significant relevance to this issue. This study attempts to grasp the dominant narratives surrounding the Al-Shifa Hospital raid by analyzing videos of Arabic, Hebrew and English on YouTube.

Research Questions

What are the most reinforced narratives about the Al-Shifa hospital in Gaza that was attacked by Israel on YouTube? Sub-question: What are the most reinforced narratives in the transcripts and the comments of YouTube videos covering the attack on Al-Shifa hospital?


By analyzing both videos and comments, this research captures a more holistic view of public opinion. Videos often reflect the creator's perspective, while comments can provide insights into the broader community's reactions and thoughts. This dual approach helps in understanding both the propagated narratives and the public's reception of these narratives.

Step 1: data collection

We used YouTube Data Tools (Rieder, 2015) to collect all the results of the query “Al Shifa hospital” in English, Arabic and Hebrew each day from November 5 to November 21 (the week before the raid by the IDF on November 15 and the week after the raid.) For each language, we ranked all the videos by their view count and selected 20 most viewed videos per language. For 3 most viewed videos per language, we used YouTube Data Tools to collect all the comments. Thus, our dataset contains 20 most viewed videos and all the comments from 3 most viewed videos of the search queries in English, Arabic and Hebrew.

Step 2: data curation

For the dataset per language, we employed custom prompts on the GPT tool through Prompt Compass (, “a tool designed to leverage Language Learning Models (LLMs) in digital research tasks. It accomplishes this by offering access to diverse LLMs, supplying a library of prompts for digital research, and enabling users to apply these prompts to a series of inputs. The tool offers two categories of LLMs: local LLMs and platform APIs. Local LLMs can be used to provide stable and reproducible results, facilitate in-depth analysis, and support a robust interpretation. [...] The library of prompts comprises prompts extracted from academic literature and other research, each of which is linked to its respective source. [...] The tool accepts user inputs either as text lines or CSV files, to be processed with the chosen LLM and prompt, with each line treated separately” (Erik Borra, 2023). Arabic content was generated using ChatGPT 3.5 Turbo, while English and Hebrew contents were produced using ChatGPT 4 due to the extensive text size, requiring a model with an expanded context window.

  1. extract the transcript of each video;
  2. using gpt to generate the narrative of the transcript of each video and the 300 most liked comments of 3 most viewed videos.

The exact custom prompt for the ‘narratives’ we used is as follows:

Objective: As a scholar in narratology analyzing YouTube video transcripts, your task is to extract narratives and various other entities and labels from video transcriptions. You will be presented with a video transcript which you need to process as follows.
Task Steps:
- Extract Title: Capture the video title to understand the overarching theme.
- Extract Geographical Entities: Identify continent and country names, including acronyms (e.g., "USA," "UK").
- Extract Thematic Terms: Focus on religion, ethnicity, and nationalistic terms (e.g., "Christianity," "Muslim," "nationalism").
- Extract stance: Determine whether the sentiment is pro, against, or neutral regarding a specific entity (e.g., "pro-Israel," "anti-Palestine").
- Identify the 'Other': Highlight who is portrayed as the antagonist or villain in the narrative.
- Identify the 'Hero': Highlight who is portrayed as the protagonist or hero in the narrative.
- Extract Narratives: Identify distinct narratives.
Output Format: Generate a JSON with columns "transcript", "Geographical entities", "Narrative", "Thematic terms", "Identified Other", "Identified Hero", "Stance", "Other", "Hero"
Apply critical analysis to categorize underlying themes and narratives. Do not provide explanations, only labels for stance, identified other, and identified hero. If there is no villain or hero, return an empty string for the respective fields.

Step 3: data analysis

For all the narratives of the transcript or the comment in each language, we grouped them into 3-5 “main narratives”. Raw Graph was used to visualize the main narratives of transcripts and comments.


Figure 1. Narratives of the top 20 viewed Youtube videos (ENG)

Figure 2. Narratives of the top 20 viewed Youtube videos (HEB)


Figure 3. Narratives of the top 20 viewed Youtube videos (ARB)

Figure 4. Narratives of the top 300 most-liked comments under the videos (ENG)