‘Data poisoning’: How artists are fighting back against Artificial Intelligence image generators
Data poisoning, exemplified by tools like Nightshade, allows artists to subtly alter images so they disrupt AI training while appearing normal to humans. When unauthorized models ingest these corrupted datasets, they produce erratic outputs, such as misidentifying objects or applying incorrect artistic styles. This serves as a defensive mechanism against the large-scale, unconsented scraping of copyrighted work that fuels many generative AI systems. This phenomenon highlights a critical tension in open data practices: the ethical implications of indiscriminately harvesting public information for commercial model training. It challenges the assumption that online data is free for any use, arguing that such practices infringe on creators' moral rights. Consequently, data poisoning acts as a form of technological governance, forcing technology vendors to reconsider their reliance on unverified and potentially infringing data sources. To address these vulnerabilities, stakeholders propose stricter data provenance audits and ensemble modeling to identify outliers. Ultimately, this issue underscores that open data must be balanced with respect for intellectual property and user rights. Recognizing data poisoning not merely as a technical glitch but as a response to systemic privacy and copyright violations is essential for developing sustainable and ethical AI frameworks.
Source: scroll.inPublished on 2023-12-26