New York’s RAISE Act Misses AI Safety Risks
The Responsible AI Safety and Education (RAISE) Act proposed by New York State aims to protect individuals from the harms associated with artificial intelligence (AI). However, it raises concerns by focusing primarily on the AI models themselves as the key points of leverage for ensuring safety. This approach risks transforming a technical challenge into a bureaucratic burden.
Overview of the RAISE Act
The RAISE Act, authored by State Assemblymember Alex Bores, establishes a set of requirements intended to ensure that AI technologies are deployed responsibly. It is currently under debate in the legislative committee. Legislators, including Bores, express fears that advanced AI may facilitate the creation of chemical, biological, and nuclear weapons. However, the greater risks are associated with the accessibility of dangerous precursor materials rather than the AI systems themselves.
Requirements and Compliance
Similar to California’s SB 1047, the RAISE Act targets advanced “frontier models”—AI systems that meet specific computational thresholds and cost over $100 million to train. Developers of these covered models must adhere to various stringent requirements, including:
- Mandatory testing procedures and risk mitigation strategies
- Regular third-party audits
- Transparency mandates
- Reporting of instances where a system has facilitated dangerous incidents
- Retention of detailed testing records for five years
- Annual protocol reviews and updates
- Prohibition against deploying “unreasonably” risky models
- Protection of employee whistleblower rights
Violations of these requirements carry significant penalties, starting at 5% of compute costs for a first violation and escalating to 15% for subsequent infractions, potentially resulting in fines ranging from $5 million to $15 million.
Challenges in Model Alignment
The RAISE Act’s premise is to align the profit motives of companies with public safety interests. However, aligning AI models to prevent misuse has proven to be a complex challenge. Model alignment tends to be more effective in mitigating accidental harms, such as generating misleading advice or incorrect information, rather than in curtailing malicious activities.
Experts from Princeton University, including computer scientists Arvind Narayanan and Sayash Kapoor, have highlighted the concept of model brittleness. They argue that even if AI models can be engineered to be “safe,” they can still be exploited for harmful purposes.
Emerging Approaches in AI Safety
Current strategies for maintaining model alignment involve the use of external systems that operate atop the models. Leading companies are investing in the development of:
- External content filters
- Human oversight protocols
- Real-time monitoring systems
These measures aim to detect and prevent harmful outputs, indicating that the market is advancing more rapidly than the existing regulatory frameworks.
Regulatory Burden vs. Actual Safety
The RAISE Act imposes extensive requirements to achieve its objectives. For instance, if robust safety protocols are functioning effectively, the necessity for five years of record-keeping comes into question. Furthermore, if a model successfully passes an independent audit, the rationale behind requiring developers to meet separate standards for “reasonableness” in deployment is unclear.
These inconsistencies go beyond mere bureaucratic inefficiencies. The RAISE Act attempts to address various complex issues including corporate transparency, employee protections, technical safety, and liability within a single regulatory framework. This broad approach risks prioritizing compliance over genuine safety outcomes.
Cost of Compliance
Policymakers often underestimate the compliance costs associated with AI legislation. Previous analyses have shown that official projections may not capture the true financial burden of compliance. Using advanced large language models (LLMs) to evaluate the RAISE Act, estimates suggest that initial compliance may require between 1,070 and 2,810 hours of labor—effectively necessitating a full-time employee. For subsequent years, ongoing compliance burdens could range from 280 to 1,600 hours annually.
This significant variance in compliance estimates underscores the inherent uncertainty surrounding the RAISE Act and similar legislation. The rapid evolution of sophisticated AI technologies indicates a pressing need for laws that prioritize effective risk mitigation rather than regulatory theater.
