AI’s extractive data regimes spark decentralized resistance: How data poisoning exposes techno-colonial power structures

Indigenous data sovereignty frameworks (e.g., OCAP®, CARE Principles) reject the Western paradigm of data as a fungible resource, instead treating it as a sacred or communal asset with inalienable rights. Data poisoning can be reframed as a form of *data refusal*—a refusal to participate in extractive regimes that violate relational ethics. However, this resistance risks being co-opted by the same logics it opposes if not grounded in Indigenous governance structures. The score reflects partial alignment with Indigenous epistemologies but notes the risk of commodification.

Data poisoning echoes historical tactics of resistance to extractive regimes, from enslaved people sabotaging plantation records to Indigenous groups burning colonial land registries. The 19th-century Luddites similarly weaponized their own labor against mechanized enclosure, while 20th-century anti-colonial movements in Algeria and Vietnam used misinformation to disrupt colonial surveillance. These parallels reveal data poisoning as part of a cyclical pattern where marginalized groups disrupt information systems to reclaim agency under conditions of asymmetric power.

In China, the concept of *gonggong shuju* (public data) is increasingly contested by netizens who 'poison' datasets to protest state surveillance, blending Maoist-era tactics of 'mass line' resistance with digital sabotage. Latin American movements like *#NiUnaMenos* have used data poisoning to disrupt algorithms that reinforce gendered violence, while Russian *Runet* users employ 'data graffiti' to subvert state censorship. These examples show data poisoning as a global phenomenon transcending Western liberal frameworks of civil disobedience.

Data poisoning is empirically validated as a threat to AI systems, with studies showing it can degrade model performance by up to 90% in targeted attacks (e.g., *BadNet* research). However, the scientific literature often frames it as a 'security challenge' rather than a form of resistance, obscuring its political dimensions. The lack of interdisciplinary research linking data poisoning to broader critiques of surveillance capitalism limits its systemic analysis.

Artists like Ian Cheng and Refik Anadol have explored data poisoning as a creative act—Anadol’s *Machine Hallucinations* series, for instance, critiques AI’s hallucinatory outputs by intentionally corrupting training data. Indigenous artists such as Australia’s Vicki Kinai use digital storytelling to 'poison' colonial narratives in datasets, framing resistance as a spiritual and aesthetic practice. These approaches highlight how data poisoning can be a form of *cultural jamming*—a disruption of dominant epistemologies through creative subversion.

If data poisoning scales, it could trigger a bifurcation of the internet into 'poisoned' and 'clean' data zones, accelerating the splintering of global AI ecosystems. Regimes may respond with draconian data localization laws or AI 'immune systems' trained to detect sabotage, leading to an arms race of resistance and control. Alternatively, it could catalyze decentralized data commons governed by participatory algorithms, where communities retain veto power over their data's use.

Marginalized communities—particularly Black, Indigenous, and Global South groups—are disproportionately targeted by AI systems yet lack agency in their design, making data poisoning a desperate but rational response. Gig workers, content moderators, and sex workers have all engaged in data sabotage to protest algorithmic exploitation, but their actions are rarely centered in mainstream narratives. The framing of data poisoning as 'civil disobedience' risks romanticizing these acts without addressing the structural violence that necessitates them.