New Perspectives on the Societal Risks of AI
AI safety, once a niche concern, is now a topic that most people have heard of, but many remain unaware of what can be done to safeguard against these new threats.
Published July 7, 2026
By Nick Fetty
The New York Academy of Sciences along with the International Science Reserve hosted a free webinar on June 30, 2026, to explore both the technical and societal challenges around AI.
On the technical side, issues such as robustness, alignment, interpretability, and controllability raise fundamental questions about how we design systems that behave as intended. Societally, concerns around misuse, systemic bias, economic disruption, and governance highlight the broader implications of deploying AI at scale. From technical design choices that can have far-reaching social consequences to policy decisions that can shape the trajectory of technical development, these two dimensions are deeply interconnected, as explored during the webinar.
The event was moderated by Nick Dirks, President and CEO of the Academy, and included three expert speakers:
- Andrew Draganov, Arcadia Impact, Research Lead and Program Manager for Alignment Research
- Sarah Kreps, Cornell University, John L. Wetherill Professor in the Department of Government, Adjunct Professor of Law, and the Director of Tech Policy Institute in the Cornell Brooks School of Public Policy
- Robert Slone, PhD, UL Solutions, Senior Vice President and Chief Scientist
Slow Moving, Diffuse Challenges
Prof. Dirks starts the conversation by emphasizing that their discussion wouldn’t just be focused on speculation and would instead look at “moving from the recognition of risk to the readiness to respond.” He points out that each member of the expert panel approaches “the problem from distinctive but quite complimentary vantages.”
“One of the key premises we’re here to test is that many of the most consequential risks are not dramatic rogue AI scenarios, but slower moving, diffuse challenges,” Prof. Dirks says.
For Dr. Slone, the regulation of AI is directly relevant to the work UL Solutions has been doing for more than 130 years on market safety for products and technologies.
“What keeps me up at night, honestly, is us,” Dr. Slone says. “It’s how humans are interacting with AI and this notion of gradual degradation of critical thought. Of using AI as an ‘easy button,’ accepting the output without questioning where it came from.”
Dr. Slone explains that AI-enabled technologies undergo engineering inspections and testing similar to other products, adding that issues around transparency and accountability are what separates AI from more traditional products.
“The one thing that’s been encouraging along the way is to see the brands we’ve worked with for decades, sometimes over a century, they clearly want to bring safe products to the market,” he says.
National Security and Ethical Concerns
Prof. Kreps, whose work falls at the intersection of national security and emerging technology, expresses some concerns around uncertainty in AI.
“There is always this window, when these technologies emerge, of uncertainty. And I think we’re in that window of uncertainty now with AI where there are just so many kinds of speculative risks. We don’t have a clear risk distribution,” she says, adding that uncertainty around AI differs from other technologies that impact national security like nuclear weapons, cyber threats, and propaganda via social media
“What’s different about AI is that it’s coming out of civilian firms, and now there are these enormous national security impacts,” Prof. Krep continues. “The dilemma is I think there are risks of intervening too aggressively if you’re a public institution like the government, and the risk of not intervening [at all].”
Much of Dr. Draganov’s work examines the alignment problem with AI, which explores AI’s ability to perform ethically.
“The thing that keeps me up at night is the pace of change because it seems more and more at least from my research, we can hand off significant portions to AI models and we don’t even have the state-of-the-art ones,” he says. “This recursive self-improvement cycle is fairly unprecedented, at least in the way it’s currently occurring, so this is very difficult to map out all the ways it can go.”
Disentangling Hype from Reality
In terms of addressing the short-term challenges around AI and national security, Prof. Kreps acknowledges it can be difficult to “disentangle the CEO-driven hype from the reality.”
“What I find very interesting about the AI space compared to social media is that it seemed like Sam Altman [of OpenAI] and Dario [Amodei of Anthropic] were very keen to get out ahead of these risks, contrary to the Facebook-Meta model where everything was reactive,” Prof. Kreps says.
Prof. Kreps worked with OpenAI in 2018 to explore the national security impacts of GPT-2.0 and GPT-3.0 in an electoral setting, prior to the public release of these models. While the Biden Administration was receptive to placing boundaries around these models, the Trump Administration reversed course with more of a “light touch approach to regulation.” Prof. Kreps feels that there is currently a tension within the government about where to flex its regulatory muscle, despite being ideologically predisposed to a more hands-off approach.
“[There’s] a realization these models are very capable whether that’s for biological hazards or cyber security, some of the ways in which these models can find vulnerabilities is truly eye-opening,” Prof. Kreps says. “I think there is a realization, objectively speaking, [that] it’s incumbent upon the government to be taking some of those risks seriously.”
Prof. Dirks adds that the situation is further complicated when a company or product reaches the IPO (initial public offerings) stage, meaning it would be subject to certain governmental regulations that may have not been relevant pre-IPO.
The Centaur Effect
Dr. Draganov said AI systems are in a precarious situation that we often do not know how or what they think, yet they are increasingly expected to make sound decisions for users in digital environments. AI agents are being tasked with everything from coding to developing data sets to training, even though historically these kinds of tasks have been performed by humans. This has led to what’s been called “the Centaur Effect.”
“[This means] the person plus the model is more capable than just the person,” explains Dr. Draganov. “But the thing we’re finding is that with each coming six-month increment, the number of things we can hand off, at least on a technical domain, is growing.”
With the AI models doing more of the work in this regard, it contributes to the problems around uncertainty that Prof. Kreps addressed. This can create issues for the user who may struggle to understand if the AI is maliciously undermining a prompt or if it’s earnestly trying to improve decision-making because of well-meaning ignorance on the part of the user issuing the prompt.
A Lack of Agreement Globally
Dr. Slone points out that there does appear to be global agreement on certain elements of AI development such as data and privacy as well as fairness and bias. This is the “easier part,” he says. The difficult part, which echoes Dr. Draganov’s sentiment, is “the dynamism.”
“What get certified on day one, could be updated on day three, and be completely different,” says Dr. Slone. “So unless we have access to lifetime data from those manufacturers, as to what updates they’ve sent through, you could have an outdated certification within 24 hours or less.”
Dr. Slone adds that this is the dynamism that is yet unresolved. Another complexity, according to Dr. Slone, is that legally there aren’t currently many requirements around AI development. For example, he says the European Union (EU) would be placing itself at an economic disadvantage if it enforced greater regulation on its tech companies compared to the United States or China.
“It’s that dynamic, changing piece of this that I think is the biggest challenge remaining that we do need to solve at some point soon,” says Dr. Slone.
China and Economic Competitiveness
While Democrats and Republicans are often at odds with one another, Prof. Kreps pointed out they do share some commonalities around China. In recent years presidents from both parties have implemented “export controls” around chip manufacturing to prevent the Chinese from gaining an economic advantage in the AI race.
“This idea of quote ‘keeping as large of a lead as possible’ is futile. It’s like shoveling water with a pitchfork,” says Prof. Kreps. “The idea that we can somehow deny China or that there’s any value in being four months ahead just seems silly to me.”
Discussions around the regulation of AI are like the work Prof. Kreps has done around nuclear nonproliferation. Leaders from countries may be skeptical to foreswear unilateral agreements if they feel it will put them at a competitive disadvantage.
“I think these [AI models] are much harder to regulate and verify compliance with than a nuclear weapon,” she says.
Dr. Slone sees an overreliance on AI in the United States as being a potential impediment to educational competitiveness.
“If we use AI as an ‘easy button’ we’re not raising our games. We’re lowering our own standards, our own bars, our own behaviors. And not applying critical thought,” Dr. Slone says.
He points to various universities and other institutions conducting research around learning outcomes and AI use. Nudges from behavioral economics can be helpful in fighting the degradation of critical thinking skills. For Dr. Slone, the return of “Blue Books” in the classroom is not a bad thing.
“I think some of those simple, basic, fundamentally-sound approaches to education, to training, to development, and to performing our work are actually really important right now,” he says.
Secrecy versus Oversight and the Role of Higher Education in AI Safety
The discussion concluded with a question-and-answer session from attendees. This included pondering issues around the balance between national security secrecy and oversight as well as the role higher education and research can play in alleviating safety concerns posed by lawmakers.
Prof. Dirks concluded the webinar by circling back to the mission of the International Science Reserve, which embraces the mantra of “scientists without borders.” A similar, egalitarian approach can be applied to the development and regulation of AI.
“I think we can all see the need to have institutions like this that bring scientists together to think about these major societal risks,” says Prof. Dirks. “To work to identify how to prepare for them but also to recognize and begin to operationalize the global network that we’re going to need from scientists, technologist, and others in continuing to bring science to bear on these kinds of issues. And also to keep science aligned with what we see as the public good.”
Special thanks to the partners for this event: FedEx, Google, IBM, Pfizer, and UL Solutions.
