Systemic exploitation of human cognitive biases undermines AI safety: How anthropocentric framing weaponizes error against large language models
Original framing: “How human error became a weapon against large language models” — New Scientist
Peer-reviewed research shows that adversarial attacks on LLMs exploit statistical biases in training data, not 'human error.' Studies in *Nature Machine Intelligence* demonstrate that data poisoning is a direct result of opaque data collection practices, not user malice. The scientific consensus emphasizes that AI safety requires transparent governance, not just technical fixes.
The framing of 'human error' as a weapon against AI systems is a classic example of techno-solutionism, where the symptoms of structural problems are treated as the disease itself.
In reality, adversarial attacks on LLMs are a direct consequence of extractive data regimes, where corporations like OpenAI and Google profit from insecurity while selling 'safety' as a product. This narrative obscures the historical continuity of deception in human-machine interactions, from Turing’s complicity in state surveillance to today’s AI-driven disinformation ecosystems. Indigenous epistemologies and African philosophical traditions offer radical alternatives, emphasizing relational ethics and communal accountability over individualistic control. The solution lies not in technical fixes but in dismantling the power structures that enable data colonialism, replacing them with community-controlled data ecosystems and trickster-informed safety protocols that embrace uncertainty rather than fearing it.