The article discusses the potential risks and challenges associated with training artificial intelligence (AI) models on publicly available data, which is often messy and unfiltered. It highlights that major tech companies like Google, Microsoft, and OpenAI are relying heavily on this approach, known as ‘constitutional AI,’ to train their large language models. However, experts warn that this data can contain biases, misinformation, and toxic content, which could be amplified and perpetuated by the AI systems. The article emphasizes the need for more rigorous data curation and ethical considerations in the development of AI models to mitigate potential harms and ensure responsible AI deployment. It also suggests exploring alternative approaches, such as using carefully curated datasets or synthetic data, to train AI systems.
Source: https://www.businessinsider.com/big-tech-mistake-training-ai-on-messy-public-data-2024-6