Towards a Sane Compromise on AI Part 1

90/10 Problems and 90/10 Solutions

The political reality is that policymakers are looking to address potential harms around AI, particularly problems around the content produced by Large Language Models (LLMs) such as ChatGPT, image models, and video models. The vast scientific[1, 2, 3, 4] and economic[1, 2, 3, 4, 5] benefits to these models are largely unaffected by these content questions. Consequently, these policy discussions are aiming to address content questions while minimizing the damage to economic and scientific applications of AI. 

All questions about the political content of AI output can be addressed with fine-tuning. Fine-tuning is one of the later steps of producing an AI model, used to target attributes of models including tone, ideology, or purpose. These behaviors are well suited to address model content, which covers attributes of model output including but not limited to:

List 1

• CBRN (Chemical, Biological, Radiological, and Nuclear) risks

• Privacy and data 

• Medical information

• Fraud and impersonation 

• Misinformation

• Mental health 

• Threats

• Discrimination against protected classes, such as race, sex, or sexual orientation. 

By establishing a clear distinction between research and values questions, we can draw a red line against economically destructive and anti-competitive policies while addressing the vast majority of potential harms in AI that policymakers identify.

With an agnostic view of what the answer to these political questions should be, how to answer them is clear: apply a single point of intervention for all content regulations at the end of the fine-tuning stage. This dramatically reduces the financial burden on both developers and regulators themselves.

In summary, fine-tuning is uniquely positioned to address List 1 content because it is effective, targeted, and cost-efficient relative to other policy options.