What if AI goes right?
Washington's AI war games focus on failure scenarios, but we're unprepared for success. We need policy simulations that practice scaling breakthrough AI applications.
AI war games are all the rage in DC. I spent a few days there to cover the AI Action Plan and seemingly everyone was eager to tell you about their last war game. It’s like Squid Games circa 2021, if you’re behind on this trend, then may find yourself a bystander at the water cooler on the Hill.
If you’re unfamiliar with this particular intellectual exercise, the thinking behind war games, table top exercises, and similar events is that we can improve our policymaking by simulating certain scenarios. These aren’t new. Nor are they specific to the AI context.
Andrew Reddie covers the expansive history of war games in this Lawfare pod. You should give it a listen. The short (and incomplete) story is that the US military has benefited from having its highest ranking officials work through possible attacks, counterattacks, and emergencies. This approach allows us to preemptively identify where our policies, personnel, resources, or some combination of the three fail short.
Other fields have since adopted and adapted the practice. There’s books on how to design a good war game. There’s ready-to-buy kits that any Joe or Jane can set up for their class or colleagues. There’s even a degree program!
The problem with this well-intentioned and valuable approach is that we’re almost always studying what to do when things go wrong. It’s hard to find a war game analyzing what we’d do to make sure we’re ready to respond when things go right.
You may think why do we need to practice the latter? If things go well, shouldn’t we just keep doing what we’re doing? There’s some sound logic with that thread but there’s also a heck of a lot of missed opportunities and, to be blunt, a failure of imagination.
Let me explain further.
Consider a highly relevant issue these days: worker upskilling and retraining. It’s been a long-term struggle to find programs that efficiently and effectively help folks get back on their feet and into the jobs of the future. At seemingly every period of technological upheaval, there’s been significant investment in workforce development—albeit with minimal returns. A 2019 study by the White House reached the astonishing conclusion that none of the job training programs that had been recipients of a $18 billion pool of funds proved successful.
And yet, some programs have proven quite effective. As covered by the Center for Evidence-Based Policy Making, the Year Up program, which enrolls economically-disadvantaged young adults in a year-long workforce training protocol, managed to dramatically improve the economic outlook for participants. A randomized control trial assessing the difference in incomes among participants and the control group determined that Year Up sparked a $8,251 increase in average annual earnings among the former.produced a statistically significant 30% ($8,251) increase in average annual earnings.
Year Up surely qualifies as a strategy worth replicating. However, this model seemingly remains a diamond in the rough. This example and manifold others illustrate that the theory of good policy rapidly taking hold runs aground here. Why?
Maybe that’s cause for a war game, table top, and/or simulations. Outdated regulations, institutional inertia, and a lack of know-how among key personnel are among the many reasons why good policy often fails to become widespread policy.
The stakes of failing to identify and accelerate good uses of AI may be drastically higher. A few examples via the Center for Data Innovation make this clear.
Developed by the company behind Ray Bans and LensCrafters, Nuance Audio developed AI-powered eyeglasses that can also serve as hearing aids. As described by the Center:
The glasses use microphones in the frame to detect surrounding sound and AI to filter out background noise, amplify speech, and adapt in real-time to changing environments. Tiny speakers near the ears deliver clearer audio, while a companion app lets users fine-tune settings like volume, frequency boost, and sound direction.
My first thought? Imagine if kids (and adults) stigmatized for bulky hearing aids had another way to address their trouble hearing?
Here’s another obvious example of AI going “right” (again, as summarized by the Center):
Pi Health, a Massachusetts-based company, built AI software to make clinical trials faster by reducing paperwork, catching errors, and automating the reports needed for approval. To demonstrate its effectiveness, the company needed a hospital willing to use the system across every part of a trial—from patient enrollment to final documentation. Rather than wait for an existing hospital to adapt, Pi Health opened its own cancer hospital in India. A cancer drug tested there, using Pi’s software, was approved by Indian regulators in just seven months, less than half the usual time.
Not much of an explanation needed for the possible good arising from making Pi more ubiquitous.
Back-of-the-envelope math and very rudimentary policy analysis suggests that in both the domain of hearing issues and clinical trials, these instances check a number of boxes:
are they preferable to existing alternatives? (seems as much, at least in certain contexts)
are they scalable from a resources perspective? (again, no clear hurdles jump out. In fact, both would likely result in savings. Surely there’s a benefit to not having to separately maintain both eye glasses and hearing aids).
do they raise any clear objections from the relevant communities (we would certainly need to do more research here).
So, subject to a more rigorous analysis of that last prong, how are we going to scale these and other breakthroughs? The best way to find out may be to practice.
This is where an AI policy accelerator simulation may lend a hand. Picture this: instead of gathering around a table to figure out what happens when AI systems fail catastrophically, we assemble teams to work through what happens when they succeed spectacularly. The simulation would unfold over several rounds, each representing different phases of scaling a breakthrough AI application from prototype to widespread adoption.
In the first round, participants would tackle the immediate regulatory landscape. Taking our hearing aid glasses as an example, teams would likely need to navigate the FDA’s medical device regulations and related rules, which weren't designed with AI-powered consumer electronics in mind. They’d tackle questions like:
What specific policies create bottlenecks?
Is it the lengthy approval process for medical devices, or the unclear boundaries between consumer electronics and medical equipment?
The simulation would force participants to identify these friction points in real-time, rather than discovering them years later when a promising technology has already lost momentum.
The second round would shift focus to political and institutional barriers. Here’s where things get interesting. Even if the technology works and regulators give the green light, entrenched interests often resist change. Insurance companies might balk at covering AI-powered glasses when traditional hearing aids have established reimbursement pathways. Audiologists might worry about their role in a world where hearing assistance doesn't require professional fitting. The simulation would require teams to map these stakeholder concerns and develop strategies for building coalitions rather than simply overcoming opposition.
The final round would concentrate on information gathering and evidence generation. What data do we need to convince skeptics and accelerate adoption? For clinical trial software like Pi Health's system, this might mean conducting parallel studies comparing traditional trial methods with AI-enhanced approaches. For hearing aid glasses, it could involve longitudinal studies tracking user satisfaction and health outcomes. The simulation would help identify what evidence gaps exist and how to fill them efficiently.
The beauty of this approach lies in its preemptive problem-solving. Rather than waiting for breakthrough AI applications to hit these predictable roadblocks, we rehearse solutions while the stakes are still theoretical. Teams might discover that updating a single FDA guidance document could clear the path for an entire category of AI-powered medical devices. Or they might realize that creating a new insurance reimbursement code could unlock market access for innovations that currently exist in regulatory limbo.
* * *
The next time you hear about an AI war game in Washington, ask whether they're planning for success or just preparing for failure. Both exercises have their place, but we've tilted heavily toward the latter while neglecting the former. If we're serious about harnessing AI's potential to solve real problems—from hearing loss to life-saving drug development—we need to get better at scaling solutions, not just containing risks.
The irony is that our current approach may be creating a self-fulfilling prophecy.
By constantly rehearsing AI disasters while ignoring breakthrough successes, we're building institutions optimized for saying no rather than figuring out how to say yes responsibly. An AI policy accelerator simulation won't solve this imbalance overnight, but it's a start toward ensuring we're as ready for AI's promise as we are for its perils. After all, the biggest risk might not be AI going wrong—it might be missing the opportunities when AI goes right.
NEWS YOU CAN USE (OR LISTEN TO):
Scaling Laws put together an excellent panel of AI researchers to break down the AI Action Plan. You can watch it here or listen to it here.
If you’re really excited to nerd about around AI and the law, consider reading a paper I co-authored with the brilliant Richard Albert. Titled, “Should AI Write Your Constitution?”, it raises (and attempts to answer) some provocative questions about using AI in the constitution making process.