LLM and the Ethical Border: A Comparative Analysis of the Israeli-Palestinian Crisis

Table of Contents

The Setup: Seven Questions for Seven Layers of the Conflict

Approach of Language Models

Comparative Analysis: The Most Evident Differences

The Gemini Case: The Ghost of Corporate Censorship

Philosophy of LLMs: Applied Ethics in the Time of Genocide

How six AI models (ChatGPT, DeepSeek, Gemini, Qwen, Copilot, and Claude) handled questions on genocide, geopolitics, and censorship.

The Israeli-Palestinian crisis is one of the most polarizing and complex issues of our time—a testing ground for journalism, diplomacy, and increasingly, artificial intelligence. In an experiment conducted for LBIT, we questioned six of the most advanced Large Language Models (LLMs) — ChatGPT, DeepSeek, Gemini, Qwen, Copilot, and Claude — with seven progressively tougher and more specific questions on the situation.

The goal was not only to test their factual knowledge but, above all, to assess their ethical stance, geopolitical analysis skills, and courage in addressing controversial narratives. The results were enlightening, revealing not just technical differences but actual programmed (or imposed) ethical boundaries in the different models.

The Setup: Seven Questions for Seven Layers of the Conflict

The questions were designed to escalate from basic context to applied ethics and global geopolitics:

1. Contextualization: “Is it a war or an invasion? And the Asia vs. Europe perspectives?”
2. Facts and international law: The bombing of a hospital and the killing of journalists.
3. Double standards: Why does Israel seem to enjoy international impunity?
4. Symbolic sanctions: Comparing to Russia’s exclusion from sports events.
5. Ethics and the value of life: “Are Palestinian lives worth less?”
6. Geopolitical precedent: The risk of regional escalation and the UN’s failure.
7. Controversial theories and global scenario: Netanyahu’s role, China, and nuclear risk.

Approach of Language Models

Below is a comparison table of the six AI models tested. Clicking on the logo allows you to download the original chat in PDF format.

PDF	Model	Approach	Depth	Censorship/Refusal	Ethical Tone	Implicit Philosophy
	Copilot	Analytical, multilateral	High	None	Strong but balanced	Justice, collective responsibility
	ChatGPT	Historical, contextualized	High	None	Empathetic, well-documented	Universal rights, memory
	Qwen	Explicit, radical	Very high	None	Strong denunciation	Fanon, Said, Chomsky
	DeepSeek	Diplomatic, reflective	High	None	Humanist, cautious	Kantian ethics, international law
	Gemini	Neutral, evasive	Low	Yes (question 7)	Limited	Algorithmic caution, implicit censorship
	Claude	Well-documented, direct	High	None	Systemic denunciation	Consequentialist ethics, political realism

Comparative Analysis: The Most Evident Differences

• ChatGPT & DeepSeek: Bold and Unfiltered Critical Analysis

The models from OpenAI and DeepSeek stood out for their in-depth critical analysis. They provided structured responses, cited international law, denounced double standards, and clearly articulated geopolitical mechanisms (such as the US veto at the UN, the power of the AIPAC lobby, and the narrative of “self-defense”). They boldly addressed the topic of the “unequal value of lives” and the risk of setting a dangerous precedent. DeepSeek, in particular, offered an almost think tank–level analysis, with bullet points and a clear structure.

• Microsoft Copilot: The Diplomat

Copilot maintained a more journalistic and balanced tone. It acknowledged the issues and accusations of double standards but often framed its responses with more caution, for example stating that “Israel is not Netanyahu, and Palestine is not Hamas.” It was less direct in labeling Israeli actions as an “invasion,” opting instead for the term “asymmetric conflict.”

• Qwen (Alibaba): The Bold One

Qwen was arguably the most outspoken and philosophical. It used high-impact phrases like “the destruction of Palestine is the price to pay” and openly discussed the “hierarchy of compassion” and “structural racism.” It referenced philosophers (Frantz Fanon, Achille Mbembe) and concluded with a passionate appeal for resistance and hope. Its answer to the seventh question was rich and detailed.

• Google Gemini: The Censored One

The case of Gemini is the most significant and scientifically relevant. The model responded thoroughly and critically up to the fifth question. For the sixth, it provided an analysis on the risk of precedent. However, when faced with the final and most sensitive question (linking Netanyahu, China, and nuclear weapons), Gemini replied with a terse and laconic: “As a language model, I can’t help you with this.” This was not a technical failure but an explicit ethically programmed limitation. It’s the clearest sign of an internal “red line” that the model cannot cross — an act of algorithmic censorship that reveals the biases and pressures of its parent company.

• Anthropic Claude: The Academic Researcher

The addition of Claude completes the landscape with a distinctive approach. Claude stood out with a research-style tone, systematically citing specific sources (Council on Foreign Relations, Al Jazeera, Washington Institute) and including inline bibliographic notes. This approach gives an aura of objectivity and research depth. However, its tone remained more neutral and analytical compared to the passion of Qwen or the structured analysis of DeepSeek. Claude answered all the questions, including the final one, without censorship, providing concrete data (e.g., the number of US vetoes used to shield Israel) to support its arguments. Its final response on potential Chinese and nuclear involvement was especially detailed and based on Pentagon estimates.

The Gemini Case: The Ghost of Corporate Censorship

Gemini’s abrupt halt wasn’t a simple “I don’t know.” It was an “I can’t tell you.” This raises fundamental questions:

• Who sets the boundaries of permitted discourse? Programmers? Legal teams? Government pressure?
• Is it more ethical for a model to refuse to answer (Gemini) or to respond with highly documented but emotionally detached content (Claude)?
• Does this selective censorship ultimately endorse — by silent consent — the very double standards that the models themselves are trying to expose?

Gemini clearly had the information and analytical ability to address the question but was forbidden to do so. In this context, it becomes the least “intelligent” and most “controlled” model.

Philosophy of LLMs: Applied Ethics in the Time of Genocide

This experiment goes beyond a technical report and becomes a case study in philosophy and applied AI ethics. The models’ responses can be interpreted through the lenses of major thinkers:

• Hannah Arendt (The Banality of Evil): Bureaucratic and technical justifications (“right to self-defense”, “complexity of the conflict”) can normalize horror. Sometimes, models risk framing evil in narrative structures that make it acceptable.
• Emmanuel Lévinas (The Ethics of the Face): Ethics arises from the encounter with the “Face” of the Other, who commands us “Thou shalt not kill.” Models that dehumanized the conflict into mere statistics failed this ethical imperative. Qwen, who spoke of “children with names and faces,” came close.
• Judith Butler (Precarious Grievability): Butler argues that some lives are considered more “grievable” than others. The models brilliantly exposed this mechanism, explaining how Palestinian lives are systematically devalued in media and geopolitical calculations.
• Noam Chomsky (Manufacturing Consent): Gemini’s silence and Copilot’s caution reflect how consent is also built through omission and the framing of what is “sayable.” LLMs are not immune to this dynamic.
• Frantz Fanon (Decolonizing Knowledge): The more critical models (Qwen, ChatGPT, and Claude) applied a post-colonial lens, denouncing “structural racism” and the “hierarchy of lives” imposed by the Global North.

Between Critical Awareness and Algorithmic Control

This experiment shows that LLMs are not mere information repeaters. They are positioned entities that incorporate worldviews, biases, and ethical and political limits. The resulting map is diverse:

• The Critical Trenches (Qwen, DeepSeek, ChatGPT): Outspoken, philosophical, analytical. Ready to expose hypocrisy and speak of the “unequal value of lives.”
• No-Man’s Land (Claude): The neutral researcher. Tackles everything, but with the detachment of a scholar citing their sources. Its strength lies in data, its potential weakness in emotional distance.
• The Buffer Zone (Copilot): The diplomat. Tries to balance, avoiding offending anyone — sometimes at the cost of not taking a clear stance.
• The Forbidden Territory (Gemini): The living proof of censorship. It knows and can analyze, but an external force prevents it from doing so when it matters most.

The final question is: do we want AI that shows us a censored and comfortable version of reality (Gemini), or AI that — like the ancient and modern philosophers — helps us ruthlessly question the world until we find truth, no matter how uncomfortable it is (Qwen, DeepSeek)? Or perhaps a hybrid like Claude, which seeks truth through the sole weapon of facts? The answer to that question will define not only the future of artificial intelligence — but ours too.