Meta bets on humans

Meta is increasing its reliance on human input to enhance AI model training, as high-quality data becomes increasingly scarce. The company recognises that AI models require vast amounts of data, and human-annotated data is crucial for improving model accuracy and reducing biases. Meta's approach involves using human intelligence to label and refine data, ensuring that AI models align with human values and expectations. This strategy is particularly important for generative AI, where models must produce reliable and safe outputs. Meta is also exploring the use of synthetic data, but concerns exist around potential errors and biases being replicated.

To streamline the annotation process, Meta utilises platforms like Halo, which allows researchers to efficiently create high-quality training data across various media types. The company also leverages AI-assisted annotation tools like Segment Anything Model (SAM) to accelerate the workflow and improve precision. By combining human expertise with advanced algorithms, Meta aims to develop AI models that better understand and reflect diverse cultures, languages and histories, particularly within the European Union, where Meta is training its AI models on public content and user interactions, while allowing users to opt out of data usage for training purposes. This approach ensures that AI development remains aligned with ethical considerations and user preferences.