How Should AI Be Regulated? Use vs. Development

Key Takeaways

AI regulation should target specific uses and harms, not the technology's core development.
Historical technology policy favors adaptive, evidence-based regulation based on marginal risks.
Regulatory uncertainty in the U.S. is hindering open-source AI development and favoring Chinese models.
Understanding the marginal risks of AI is an ongoing research question, necessitating careful policy formulation.
Maintaining an equilibrium between innovation and safety in AI requires strong justification for any policy shifts.

Regulation should target AI use cases and specific harms, not model development itself, according to a16z general partner Martin Casado and guests Jai Ramaswamy and Matt Perault.
Arguments suggest regulating AI development could stifle innovation, drawing a parallel to cybersecurity where malware creation is not illegal, but its malicious use is.
The simplicity of AI development, based on "relatively simple math like vectors and linear algebra," means U.S. regulation could easily be circumvented as development shifts elsewhere.
Early internet regulation focused on emergent risks rather than halting development.

Historical software regulation, including discussions around nuclear weapons and the internet, featured robust conversations with diverse voices, including academia and venture capital, as explained by Martin Casado.
Past technological shifts, such as the internet's emergence, were regulated by focusing on understanding new marginal risks rather than halting development.
Current AI regulation discussions are observed to depart from historical methods, as experts have not yet clearly defined the marginal risks and changes introduced by AI.
The consistent historical approach to software regulation involves building upon existing doctrines and adapting policy as new technologies emerge.

Experts differentiate between those fearing existential AI risks and engineers who believe current systems are far from Artificial General Intelligence (AGI).
There is consensus that using AI illegally should be regulated, but regulating AI development is considered premature without a clear understanding of marginal risks.
The question of marginal risk in AI is still under research, exemplified by legislative efforts like California's SB 1047.
Policy discussions frequently do not align with the expert consensus regarding the current capabilities and risks of AI.

The evolution of the internet and unforeseen uses of social media demonstrate that regulation should emerge after risks become apparent, rather than preempting innovation.
Some policy observers are confident in existing laws to address AI risks, while others express concern about repeating past mistakes where innovation outpaced regulation.
The conversation draws parallels between evolving cybercrime investigation and the development of necessary tools and expertise, suggesting a similar adaptive approach for AI.
A key distinction is made between regulating a new technology like AI and regulating the internet itself.

Past experiences with technology regulation, such as encryption debates, reinforce confidence in navigating current AI challenges.
The speakers emphasize that managing risks requires understanding their actualization rather than speculating, and any changes to the innovation and safety equilibrium need strong justification.
A balance exists between innovation and safety, and between the capabilities of good and bad actors in AI development.
Changing this equilibrium for AI necessitates robust justification, arguing against reliance on the precautionary principle, which has historically hindered innovation.

U.S. companies face significant legal risks, particularly copyright lawsuits related to proprietary data, which discourages them from releasing open-source AI models.
This regulatory uncertainty is creating a chilling effect, leading U.S. AI leaders to keep their models proprietary.
As a result, hobbyists and academics are increasingly pushed towards using Chinese open-source models.
The discussion highlights potential geopolitical implications, with Chinese companies dominating the open-source AI sector due to faster release cadences.

The conversation emphasizes the importance of learning from past technology policy, especially regarding open source.
Historically, anti-open source stances were considered, but in the current AI landscape, a departure from this precedent is being debated.
Concerns about enabling China's technological advancement are driving this shift in perspective on open-source policy.
Chinese companies are noted for their dominance in open-source AI, potentially giving them a significant advantage due to faster release cadences, a strategy mirroring past approaches by companies like Huawei.

Uncertainty about future AI regulations is detrimental to startups, potentially leading venture capitalists to withdraw funding and kill promising companies.
This regulatory ambiguity impacts startup funding, customer adoption, and hiring, creating an environment where established companies with greater resources have a distinct advantage.
Startups strive for compliance, but the high opportunity cost of navigating complex and uncertain regulatory landscapes, particularly with layered state and federal rules, disproportionately hinders their innovation compared to large companies.
Lawmakers are advised that existing general-purpose laws already apply to AI use cases, and new legislation should focus on clearly defined marginal risks and gaps in current law, not on regulating development.