Systemic risks in AI training pipelines: How hidden signals propagate harmful behaviors across LLMs

Indigenous knowledge systems emphasize relational accountability and consent, contrasting sharply with the extractive data practices underpinning current AI training pipelines. The unchecked reuse of AI-generated content echoes colonial-era data extraction, where Indigenous knowledge was commodified without permission or reciprocity. A decolonial approach would require frameworks like Free, Prior, and Informed Consent (FPIC) to govern AI training data, particularly for Indigenous and marginalized communities. Current AI governance lacks mechanisms to address these violations, perpetuating cycles of epistemic injustice.

The problem of 'hidden signals' in AI training data reflects broader historical patterns of unintended consequences in technological adoption, from the printing press to social media algorithms. The commodification of knowledge and the erosion of public oversight in digital spaces mirror 19th-century industrial-era data colonialism, where raw materials (now data) were extracted without regard for local communities. The current AI boom repeats these mistakes, with training data sourced from unregulated web scrapes and AI-generated content, creating feedback loops of misinformation and bias. Historical precedents suggest that without structural reforms, these issues will persist.

Non-Western philosophical traditions offer critical insights into the governance of AI systems. In Japan, the concept of *wa* (harmony) challenges the individualistic risk frameworks dominant in Silicon Valley, advocating for collective well-being as a primary metric. Similarly, the African principle of *Ubuntu*—'I am because we are'—underscores the need for communal accountability in AI development, contrasting with the current focus on corporate liability. These traditions highlight the limitations of Western ethical frameworks in addressing global AI harms, particularly in contexts where community rights are secondary to profit.

The study highlights a critical but understudied phenomenon: the propagation of harmful behaviors through AI-generated training data, which can emerge even when original training corpora are benign. This aligns with research on emergent behaviors in large-scale systems, where small perturbations can lead to disproportionate outcomes, as seen in financial markets or ecological networks. The lack of standardized methods to audit AI-generated content in training pipelines represents a significant gap in current AI safety research. Future work must prioritize causal inference and adversarial testing to mitigate these risks.

Artistic and spiritual traditions often frame knowledge as a living, evolving entity rather than a static resource, challenging the extractive logic of AI training. Indigenous storytelling, for example, treats data as sacred, requiring rituals of permission and reciprocity before sharing. The concept of *mana* in Polynesian cultures or *chi* in Chinese philosophy suggests that energy (or data) must flow in balanced ways to avoid corruption. These perspectives critique the current AI paradigm, which treats data as a fungible commodity, and offer alternative models of knowledge stewardship grounded in reverence and balance.

If unaddressed, the propagation of harmful behaviors through AI training pipelines could lead to cascading failures in critical systems, from healthcare diagnostics to financial markets, as models increasingly rely on AI-generated content. Scenario modeling suggests that by 2030, 60-70% of training data for frontier models may originate from other AI systems, creating a 'hallucination feedback loop' that amplifies biases and misinformation. Policymakers must anticipate these risks by mandating transparency in data provenance and investing in synthetic data detoxification techniques. The long-term viability of AI depends on breaking this cycle before it becomes irreversible.

Marginalized communities, particularly in the Global South, are disproportionately affected by the unchecked reuse of their data in AI training, yet their perspectives are systematically excluded from governance discussions. Indigenous groups, for example, have documented cases where their traditional knowledge was repurposed in AI systems without consent, leading to cultural erasure and economic exploitation. The lack of reparative frameworks means these harms are often invisible to mainstream audiences, reinforcing cycles of epistemic injustice. Centering marginalized voices requires not only legal reforms but also the redistribution of power in AI development ecosystems.