Using bigger AI training data sets may produce more racist results
Contrary to the widespread industry belief that scaling data improves AI fairness, new research demonstrates that larger training sets can actually amplify racial biases in artificial intelligence. By comparing models trained on datasets of different magnitudes, researchers found that the system utilizing more data was significantly more likely to associate Black faces with criminal categories than the one using less data. This finding directly challenges the assumption that quantity inherently leads to better diversity, suggesting instead that the underlying composition of internet-scraped data carries inherent prejudices that are reinforced rather than diluted by volume. This discovery is critically relevant to the open_data movement because it highlights the urgent need for rigorous quality control in publicly available training resources. The study indicates that many organizations fail to perform basic ethical checks to remove hateful or biased content, even when such data is accessible. Since open-source datasets allow for public scrutiny, they offer a unique opportunity to identify and correct these systemic flaws. Without proactive curation and transparency, the open_data community risks perpetuating harm under the guise of neutrality, making it essential to prioritize data ethics over mere scale. Furthermore, the research underscores the dangers of proprietary, closed datasets, which remain hidden from public evaluation and may harbor even deeper biases. The authors argue that transparency is the only viable path to accountability, urging major tech corporations to emulate the openness of non-profit initiatives. By advocating for broader access to training data, the study reinforces the open_data principle that public oversight is necessary to mitigate harmful algorithmic outcomes. Ultimately, ensuring ethical AI development requires not just more data, but higher-quality, transparent, and critically examined data sources.
Source: newscientist.comPublished on 2023-07-14