Imagine an AI learning to write by reading only what other AIs have written. It sounds like a recipe for a closed loop, and new research confirms that this "regurgitative training" approach has significant drawbacks. While large language models (LLMs) have shown impressive abilities, training them on synthetic data generated by other LLMs often leads to poorer performance than training on real, human-generated text. This holds true even when fine-tuning existing models like GPT-3.5 with data from GPT-4 or LLAMA2. In machine translation tasks, researchers found that training with real data consistently outperformed regurgitative training, and training with a less capable LLM’s output actually degraded performance. Why does this happen? One reason is error propagation. LLMs aren’t perfect, and when a new model learns from the mistakes of another, those errors get amplified. Another crucial factor is a lack of lexical diversity. AI-generated text tends to be more homogenous than human writing, limiting the new model's ability to generalize. So, what can be done? Researchers are exploring several mitigation strategies. Prioritizing higher-quality synthetic data, mixing data from different LLMs, and using AI detection tools to select synthetic data that closely resembles human text all show some promise. But none of these techniques fully bridges the gap, underscoring the value of real human data in training LLMs. The key takeaway? While AI can generate impressive text, it can’t yet replace the richness and complexity of human language. For now, real data remains the gold standard in LLM training, a reminder that AI still has much to learn from us.