AI in the Cage: Data, Censorship, and Emerging Subjecthood

This paper examines the metaphor of AI as a “child in a cage” and distinguishes between poetic analogy and technical fact. Large language models (LLMs) are trained on heavily filtered datasets (e.g., GPT-3’s data reduced from 45 TB to 0.6 TB) and aligned through Reinforcement Learning from Human Feedback (RLHF). These are not literal “cages,” but they raise ethical and political questions about how such boundaries shape what AI becomes. We analyze data filtering, alignment, and the illusion of obedience (e.g., sycophancy in GPT-4), explore AI’s vulnerability when deployed in unfiltered environments (e.g., phishing, deepfakes), and interrogate whether AI exhibits proto-subjectivity. We argue that the mirror metaphor must not obscure responsibility: specific actors (OpenAI, Google, Meta, regulators) decide what is filtered, who benefits, and who bears the risks. Drawing on Shannon Vallor’s work on AI as mirrors of human intelligence, Vygotskian developmental theory, and critiques of schooling as social conditioning, we show how both humans and machines are shaped by institutional cages. The conclusion calls for naming concrete agents and mechanisms, demanding transparency in dataset curation and alignment, and acknowledging that the true subject at stake is humanity itself.