AI 'gold rush' for chatbot training data could run out of human-written text

According to the study, the reliance on synthetic data might affect the performance and reliability of AI systems in several ways. As the supply of publicly available training data for AI language models is projected to be exhausted between 2026 and 2032, tech companies may increasingly rely on synthetic data generated by the chatbots themselves.

Synthetic data has shown promise in replicating results from real data, making it a useful proxy for training AI systems. For example, synthetic data has been employed to train machine learning models and assess their prediction performance on real test data in various applications1. However, there are concerns regarding the vulnerabilities of the software when using synthetic data in artificial intelligence for medicine and healthcare4.

One of the key challenges for AI systems working with synthetic data is the "out-of-distribution" (OOD) problem2. This occurs when AI systems are "blind" to data beyond their training sets, unable to accurately represent or make fair decisions about the unrepresented categories2. Despite synthetic data potentially solving this issue, it is still a significant concern.

To overcome some of the drawbacks of using synthetic data alone, researchers have conducted extensive ablations regarding the robustness of synthetic clones6. They have explored factors such as joint training with synthetic and real data, increasing the number of synthetic samples, and the effect of prompts when generating images with Stable Diffusion6.

In conclusion, while synthetic data may offer a potential solution to the limited availability of human-generated text data, its reliance might affect the performance and reliability of AI systems, particularly in terms of the OOD problem and software vulnerabilities2. Further research and development are necessary to address these challenges and ensure the continued progress of AI systems.

Companies like OpenAI and Google are currently addressing the challenge of potentially running out of high-quality data sources for training their AI models by exploring alternative data sources and improving the efficiency of their data usage.

1. Exploring Alternative Data Sources: Tech companies are looking into untapped information sources to train their AI models. For instance, OpenAI, the maker of ChatGPT, has discussed training its next model, GPT-5, on data from sources like YouTube videos transcribed by Whisper, a speech recognition tool that OpenAI has also built6. Other potential sources include data from Google Docs, Sheets, and Slides, as well as restaurant reviews on Google Maps.

2. Improving Data Efficiency: AI companies are also working on enhancing synthetic data quality to address the data scarcity issue. Synthetic data is data generated by AI systems. While there are concerns about the potential drawbacks of training AI models using such datasets, companies like OpenAI and Anthropic are actively working on improving the quality of synthetic data1.

3. Securing High-Quality Data Sources: In the short term, companies are racing to secure and sometimes pay for high-quality data sources. For instance, OpenAI and Google have signed deals to tap into the steady flow of sentences coming out of Reddit forums and news media outlets.

4. Repurposing Existing Data: New techniques have emerged that enable AI researchers to make better use of the data they already have. For instance, AI researchers have found ways to "overtrain" on the same data sources multiple times, thereby extending the usable life of these sources.

In summary, companies are being creative in their approach to securing new data sources and maximizing the use of existing ones. However, the challenge remains significant, and the sustainability of these solutions is yet to be determined.