Solving AI’s biggest problem: Microsoft, Google, Meta are using fake data to train their AI models

AI companies for quite some time now have been are grappling with the challenge of obtaining high-quality data to train their systems, leading them to explore alternative methods such as synthetic data.

Traditionally, AI systems relied on vast amounts of data extracted from various sources like articles, books, and online comments to understand user queries and generate responses. However, the availability of such high-quality data on the internet is limited, which has prompted AI firms to seek alternative solutions.

Synthetic data, essentially artificial data generated by AI systems, is emerging as a promising approach to address this issue. By leveraging their own AI models, tech companies are producing synthetic data to train future iterations of their systems. This method, dubbed an “infinite data generation engine” by Anthropic CEO Dario Amodei, aims to mitigate legal, ethical, and privacy concerns associated with traditional data acquisition methods.

More from Tech

How ChatGPT is becoming everyone’s BFF and why that’s dangerous America ready for self-driving cars, but it has a legal problem

Although synthetic data in computing is not a novel concept, the rise of generative AI has facilitated the creation of higher-quality synthetic data at scale. Major AI companies like Meta, Google, and Microsoft have started using synthetic data to develop advanced models, including chatbots and language processors.

For instance, Anthropic used synthetic data to power its chatbot, Claude, while Google DeepMind employed this method to train a model capable of solving complex geometry problems. Meanwhile, Microsoft has made its small language models, developed using synthetic data, publicly available.

The process of generating synthetic data involves setting specific parameters and prompts for AI models to create content. For example, researchers at Microsoft tasked an AI model with generating children’s stories using a predefined list of words. This approach allows for more precise control over the data used to train AI systems.

Impact Shorts

More Shorts

America ready for self-driving cars, but it has a legal problem

Alibaba, Baidu begin using own AI chips as China shifts away from US tech amid Nvidia row

However, some AI experts have raised concerns about the risks associated with synthetic data. Researchers at prominent universities observed instances of “model collapse,” where AI models trained on synthetic data exhibited irreversible defects and produced nonsensical outputs. Additionally, there are concerns that synthetic data could exacerbate biases and toxicity in datasets.

Despite these challenges, proponents argue that synthetic data, when properly implemented, can yield accurate and reliable models. Nonetheless, there is no consensus on the best practices for generating synthetic data, highlighting the need for further research and development in this area.

Furthermore, there is a philosophical debate surrounding the reliance on synthetic data, with questions arising about the nature of AI intelligence and its potential divergence from human understanding. Stanford University professor Percy Liang emphasized the importance of incorporating real human intelligence into the data generation process, highlighting the complexity of creating synthetic data at scale.

While synthetic data can be a promising solution to the data quality dilemma faced by AI companies, its implementation requires careful consideration of ethical, technical, and philosophical implications. As the field continues to evolve, collaboration between AI researchers and domain experts will be crucial in harnessing the full potential of synthetic data for AI development.

(With inputs from agencies)