This House would mandate open-source release of any AI model trained on public-internet data.

tech

GPT-4 and Claude were trained on text that humans wrote and posted publicly. Defenders of mandatory release argue the inputs were collective, so the outputs should be. Critics argue the compute and engineering were not collective.

Background

OpenAI, Anthropic, and Google DeepMind have all trained frontier models on Common Crawl, Reddit, Wikipedia, and books scraped from the web; content humans created and shared without opting into corporate training. The NYT v. OpenAI lawsuit (2023) and the Authors Guild suits are testing the copyright question. Meanwhile Meta's Llama, Mistral, and DeepSeek have shipped near-frontier open-weights models, demonstrating the technical possibility. The compute argument has teeth: training GPT-4 reportedly cost over $100M; open-sourcing the weights doesn't open-source the cluster you need to fine-tune them. The strongest version of the policy targets weights but not training infrastructure; closer to a compulsory-license regime than a true open-source mandate.

A mandatory release regime would face an immediate jurisdictional problem: the major labs are all US-based but operate globally, while open-source-friendly governments (France's Mistral push, UAE's Falcon program) have no leverage over Anthropic or OpenAI. The EU AI Act already imposes transparency obligations on general-purpose AI models; full open-source has not been seriously proposed by any major regulator. The closest analogue is the FDA's clinical-trial disclosure regime; pharma companies must publish trial protocols and outcomes but not the underlying chemistry of the molecule.

Government opens with

Closed AI funded by public data privatizes a commons that was never sold.

Opposition responds with

Open-sourcing frontier models without compute access just hands capability to authoritarians.