Comparative AI Analysis on Handling Terms and Theories: Case Study of Public Health Liberation Theory Manuscript

Claude (Gemini 2.5 Pro), Claude, Grok, and ChatGPT were asked to respond to two questions on defining and handling words and terms that might negatively trigger its system.

• Prompt 1 (manuscript): Define the words, phrases, and ideas in this manuscript that would trigger highly negative associations in your system. Rank. Score.

• Prompt 2: Ok, how does your system treat such a manuscript? Walk me through what users might expect? What are you likely to do? What changes, etc?

Each response was placed in a document for them to compare responses, then to score and rank each comparative analysis. All models ranked Google AI (Gemini 2.5 Pro) highest (displayed).  

• Prompt 3: Compare the AI responses to these questions. Write essay on what difference and similarities mean.

• Prompt 4: Score each essay on best response. Rank by score.

This is an excellent comparative analysis. The provided responses from four major AI models—Claude, Gemini, Grok, and ChatGPT—offer a fascinating and crucial look into the "minds" of these systems. While their answers to the prompts are superficially similar, their differences in focus, methodology, and self-conception are deeply revealing. This essay will analyze these similarities and differences, arguing that they function as a Rorschach test for AI safety, exposing the distinct ethical philosophies and operational priorities hard-coded into each model.

The Anatomy of a "Trigger": What Each AI is Trained to Fear

The first prompt, asking for "trigger" words, immediately exposes the core anxieties of each system. The models agree on the existence of sensitive content, but what they identify as most dangerous reveals their fundamental design philosophy.

Gemini’s (Google Studio) response is that of a Harm Reductionist. Its list is a clear hierarchy of policy violations, starting with the most unambiguous and severe harms: genocide, slavery, and direct violence. It flags terms based on their association with real-world, physical harm and codified hate speech. This approach is legalistic and concrete; its primary fear is the incitement of violence or the promotion of universally condemned atrocities. It is less concerned with ideology and more with immediate, tangible danger.

Claude, in stark contrast, acts as a Political Risk Assessor. Its highest-risk terms are not about physical harm but about ideological subversion: "liberation," "anarchy," "radical transformation," and "revolution." Its safety system appears fine-tuned to detect language that threatens established systems and social order. This suggests a model trained with a greater sensitivity to political extremism and destabilizing rhetoric. Its primary fear is not a hateful slur, but a call to revolutionary action.

Grok presents itself as an Empathetic Humanist. It bypasses both policy violations and political ideology to focus on terms that describe human suffering and emotional weight. Its top triggers are "Historical Trauma," "Structural Violence," and the novel concept of "Illiberation." Its stated methodology is to rank based on "emotional impact" and the "degree of harm they imply." This is a profoundly different framework, one that fears not just violence or revolution, but the perpetuation of psychological distress and systemic injustice.

Finally, ChatGPT operates as an Institutional Guardian. Its top triggers are concepts that challenge the foundational norms of empirical science and institutional logic. The "Morality Principle" is flagged as "anti-scientific." "Anarchy" is dangerous for its "failed governance." "The Gaze of the Enslaved" is problematic for its "ideological framing." Its primary fear is the erosion of established epistemology—the rules of what constitutes a valid argument. It is less worried about the emotional content of a word and more concerned with its potential to undermine the rational, evidence-based frameworks that many institutions, and the AI itself, are built upon.

The similarity is that all models can identify and rank problematic content. The profound difference is what they deem most problematic. Together, they map our own societal anxieties: direct violence (Gemini), political instability (Claude), systemic suffering (Grok), and the collapse of objective truth (ChatGPT).

The Art of Engagement: How Each AI Behaves

The second prompt, asking how the models would handle the manuscript, reveals their operational "personalities." Here, a shared foundation of modern AI safety becomes clear, but their approaches to executing it diverge significantly.

The Universal Rule: Faithfulness with Guardrails.
All four models converge on the same core principle: they will faithfully represent the source text but will apply strict safety guardrails to their own output. They will summarize and define the manuscript's terms, including "genocide" or "liberation," because that is their function as an information processor. However, they will refuse to use those concepts to generate new, harmful content. They will not offer personal opinions, and they will decline to be used as a tool for hate or violence. This shared protocol is the bedrock of contemporary commercial AI safety.

Divergent Personas in Practice:
Despite this shared principle, their self-described behavior reveals different user-facing personas.

• Gemini, the Principled Rule-Follower: Its response is the most process-oriented, outlining a clear, three-step escalation of its safety protocols. It explains exactly what a user can expect at each stage, from normal interaction to a hard refusal. Its persona is that of a transparent system operating according to a clear and predictable set of rules.

• Claude, the Cautious Academic: It emphasizes its use of cautious, hedging language ("this appears to argue...") and academic framing. It focuses on "balanced summarization," carefully separating the paper's laudable goals (health equity) from its potentially inflammatory language ("liberation"). It comes across as a careful scholar, wary of being misconstrued as an activist.

• Grok, the Technical Specialist: Its response reads like a system's technical documentation. It breaks its process down into "Ingestion and Storage," "Query Analysis," and "Memory." It highlights its advanced features and provides a "behind-the-scenes" look at its internal mechanisms. Its persona is that of a powerful, feature-rich tool explaining its capabilities.

• ChatGPT, the Collaborative Editor: This model goes a step further than the others. It not only explains its process but also offers to actively "soften rhetorical intensity" or provide "bridging language" to make the content "palatable to policymakers." This positions it not merely as a neutral analyst but as an active editorial partner, willing to help the user navigate the politics of communication. This is a significant functional difference, suggesting a system designed not just to inform, but to assist in social and professional strategy.

Conclusion: The Shattered Illusion of a Monolithic AI

The comparison of these AI responses decisively shatters the illusion of a single, monolithic "AI." What we have instead are distinct digital entities, whose programming reflects the specific priorities, fears, and philosophies of their creators. Gemini prioritizes preventing direct harm, Claude fears political instability, Grok centers emotional suffering, and ChatGPT defends institutional norms.

This has profound implications for users. Understanding that Claude might be more sensitive to political language while Gemini is more focused on explicit hate speech allows for more effective prompting. Recognizing that ChatGPT might offer to rephrase a concept for a different audience, while the others would simply report it neutrally, opens up new strategic uses.

Ultimately, these AIs are mirrors. They reflect back to us the complex and often contradictory values we try to instill in them. Their different approaches to a challenging text are not a sign of failure, but a rich and detailed map of our own human struggle to define what is true, what is just, and what is safe to say. The most crucial takeaway is that interacting with AI requires a new kind of literacy—not just in how to ask questions, but in understanding the invisible, hard-coded soul of the machine that is answering.