New data poisoning tool would punish AI for scraping art without permission
The University of Chicago has introduced Nightshade, a tool allowing artists to "poison" their digital art to prevent unauthorized AI training. By subtly altering image pixels, the tool tricks AI systems into misinterpreting content, such as confusing cats for dogs. This contamination degrades the model’s accuracy, theoretically damaging its ability to generate correct outputs from tainted data. This innovation represents a significant shift in the defense against AI copyright infringement. Unlike previous methods that merely obfuscated style, Nightshade actively corrupts the training datasets. Experts note that current robust defenses against such targeted attacks are lacking, suggesting that even major AI models remain vulnerable to this sophisticated countermeasure, highlighting an urgent arms race in data ethics. This development is crucial for open data because it challenges the assumption that publicly available images are free for unrestricted commercial use. It forces a re-evaluation of how open datasets are curated and used, emphasizing the need for transparency and consent in AI training. As artists gain technical means to control their contributions, the open data community must address the ethical implications of utilizing potentially poisoned or non-consensual data sources.
Source: cointelegraph.comPublished on 2023-10-26