When Does Distilling GPT Become Theft? (legallayer.substack.com)

A recent article explores the legal implications of training AI models using outputs from competitors, raising critical questions about ownership and intellectual property. The piece highlights a significant case involving DeepSeek, which claimed to develop a powerful AI model for considerably less money than OpenAI’s ChatGPT. OpenAI accused DeepSeek of “distilling” its model outputs, prompting discussions on how data acquisition methods define legal boundaries. This analysis reveals that traditional copyright laws may not apply, and instead, trade secret laws become relevant concerning how developers access and utilize these outputs, emphasizing the potential for legal exposure even in common AI development practices. Technical implications are wide-ranging. The article underscores the importance of compliance with licensing agreements, especially those containing anti-distillation clauses, as breaching these can transform lawful engineering into potential misappropriation. Additionally, it delves into the nuances of reverse engineering protections under U.S. law, emphasizing that while reverse engineering is generally lawful, violations of specific contractual terms could negate that protection. This ongoing legal discourse impacts developers across the AI/ML community, highlighting the need for clear understanding of both the legal landscape and ethical considerations in AI model training.