AI isn’t magic and it isn’t doom. It’s a power tool—useful, fast and occasionally dangerous. The question isn’t whether we should build better tools; it’s whether we’re willing to bolt on the guard before plugging them in.
That’s the balance New York Assemblymember Alex Bores is trying to strike with the Responsible AI for Safety Evaluation Act—RAISE for short. When I sat down with him, we talked about the tension between competition and caution—the pressure to “win” the AI race and the risk that speed without brakes leads straight into the wall.
Bores doesn’t see the state’s role as slowing innovation. He sees it as keeping the floor from collapsing while everyone else is running. “States can move faster than the federal government. In 2023, New York passed about 774 bills while Congress passed 27,” he told me. “That’s not necessarily a good or bad thing—it just means we can adapt more quickly when the technology changes.”
That agility matters because AI is advancing at a pace Washington can’t match. Federal efforts like the AI Bill of Rights or voluntary safety pledges are valuable in spirit but thin on teeth. States, meanwhile, have started to do what they always have—treat policy as a prototype.
What the RAISE Act actually does
The RAISE Act isn’t sweeping, catch-all legislation for AI. It’s narrow by design. “The RAISE Act only applies to the absolute largest builders—those spending more than a hundred million dollars on the final training run of a model,” Bores explained. “It requires them to have a safety plan, disclose critical safety incidents and if their testing shows an unreasonable risk, they can’t release that model.”
That specificity is key. It doesn’t burden startups or open-source researchers tinkering on small systems. It focuses on frontier models—the kind with enough compute to alter markets or national security. It’s also unapologetically focused on catastrophic risk.
As Bores put it, “This bill is focused on catastrophic harms—bio, critical infrastructure, that kind of thing. It’s not about algorithmic bias or discrimination; there are separate bills working on that.”
That separation matters, according to Bores. Too many proposals lump every AI concern—from deepfakes to hiring bias to extinction—into one incoherent mess. The RAISE Act treats existential risk as its own category, while companion bills target data transparency and content provenance.
Closing the loopholes
The legislation also anticipates one of the trickier tactics in model development: knowledge distillation, where a smaller model learns from a larger one’s behavior. Bores wanted to be clear that scaling down shouldn’t mean ducking responsibility. “We specifically included language so that knowledge-distilled models are covered. If you’re releasing a smaller version trained from a frontier model, you still have the same responsibilities.”
That clause closes a key gap. Without it, a company could claim its product isn’t a “frontier model” because it’s technically derived rather than original—a recipe for endless shell games.
The broader push for transparency
Bores also pointed to what comes next: “The next step is training-data transparency and C2PA provenance. We need to know what goes into these systems and be able to verify what comes out.”
That two-part focus—understanding inputs and authenticating outputs—is where most safety conversations eventually land. If we know what data trained a model and can prove the origin of what it produces, we can reason about trust. If we don’t, we’re just hoping the machine behaves.
The Coalition for Content Provenance and Authenticity — or C2PA — is an organization that has developed an open technical standard for verifying the origin and history of digital content. Those provenance standards already exist and can be built into cameras, editing tools and AI generators to flag when an image or video has been synthetically altered. Mandating—or at least normalizing—that approach could make deepfakes less persuasive overnight.
States are writing the first draft of AI law
New York isn’t alone. California recently passed its own AI Accountability Act (SB 1047), signed by Governor Gavin Newsom in September. Like RAISE, it targets high-compute frontier systems, requiring developers to document safety practices, conduct risk assessments and implement “kill switches” if models behave unpredictably. Alongside that, SB 896 tackles generative AI’s role in deepfakes and misinformation by mandating clearer disclosure when synthetic media could mislead voters or consumers.
Together, these laws show a pattern: states aren’t waiting for Washington. They’re experimenting—setting practical boundaries for AI development while federal efforts remain mostly advisory.
It’s messy, sure. But it’s also how American governance has always worked. Laboratories of democracy don’t need permission to start testing.
My take: speed with structure
I’ve spent enough years covering cybersecurity to know that regulation doesn’t stop bad actors. Criminals generally know they’re doing something wrong and creating a new law to define it is the equivalent of sharing thoughts and prayers in the wake of a mass shooting at a high school. It’s wishful thinking that does nothing to address or resolve the issue.
Hopefully, though, regulations like these can stop the next catastrophe from becoming a case study. The RAISE Act is pragmatic, not performative. It gives regulators a pre-harm brake—authority to question releases that developers themselves admit are risky. That’s a rare and necessary shift from the post-incident blame game we usually play.
Security professionals see the potential. Vineeta Sangaraju, security solutions engineer at Black Duck, called RAISE “a long-awaited, landmark step toward responsible innovation in today’s AI rat race.” She believes it changes the conversation from speed to accountability. “RAISE provides companies and their customers assurance that frontier AI systems meet a standard of responsible development, allowing them to confidently unleash business innovation in an era of accelerating risk,” she told me. Sangaraju adds that the law could go further with independent third-party verification and more technical testing—adversarial simulations, prompt-injection checks and other stress tests—to prove resilience as well as capability.
Not everyone shares that optimism. John Watters, CEO of the cyber-risk intelligence firm iCOUNTER, sees RAISE as “theoretically elegant, but not executable.” As he put it, “You can buy a gun to protect yourself, to arm your military, or to rob a bank – AI is the gun.” His concern is that rules can’t easily constrain data sets or prevent dual-use: the same AI that strengthens a network’s defenses can also help attackers map it for intrusion. In cyberspace, he argues, the problem isn’t just speed—it’s the habit of shipping minimal viable defenses in pursuit of maximal innovation.
Both views ring true. The security engineer wants measurable standards; the risk analyst worries about enforceability. Between them lies the reality: progress needs friction. Guardrails don’t have to halt motion—they just keep it from veering off the cliff.
There’s a larger lesson here too. We talk about “AI races” as if the finish line is discovery itself. But as history shows—whether nuclear, chemical, or cyber—the real finish line is safe, stable use. Winning doesn’t mean being first; it means being sustainable enough to stay in the race.
State-level frameworks like RAISE and California’s SB 1047 won’t solve everything. They’ll evolve, fracture and iterate, just like the technologies they govern. But they create momentum and precedent—the scaffolding that lets smarter national policy take shape.
The alternative is the Wild West we already see forming online: untraceable models, synthetic media without provenance and no shared threshold for “too dangerous to deploy.” RAISE doesn’t outlaw innovation; it just asks builders to think like adults before lighting the fuse.
That shouldn’t be controversial. It should be the baseline for progress.
