- Science & Technology
Max Lamparth is a research fellow at the Hoover Institution’s Technology Policy Accelerator (TPA), where he focuses on artificial intelligence, and is affiliated with the Stanford Intelligent Systems Laboratory and the Stanford Center for AI Safety. He recently spoke with Donna Obeid, TPA’s senior research program manager.
Donna Obeid: What first drew you to AI, and how did your path from physics to computer science shape the questions you now study?
Max Lamparth: I have wanted to be a researcher since I was a kid. What pulled me in was problem solving, the feeling of taking something confusing and finding the structure underneath. Early on I became convinced that answering fundamental questions requires using evidence and data, which led me to pursue physics. I loved that you could describe the universe with a handful of equations and then test them with statistical and numerical methods.
During my PhD I started using machine learning as a tool to extract verifiable scientific results and optimize scientific experiments. By then, groundbreaking systems like AlphaGo beating one of the best human players in Go, and OpenAI Five achieving something similar in a strategy game, had shown me that this same search for structure could become a kind of general problem solving at an unmatched scale.
In early 2022, I moved into computer science working on language models. The more capable these systems became, the more I cared about a different question: not just what they can do, but whether we can reliably trust them in the high-stakes settings where we can’t simply check the answer, given how deeply these systems are starting to shape society. So, my work aims to answer how AI systems learn human preferences and objectives, when that learning breaks, and how to make it reliable where we cannot simply check the answer.
Donna Obeid: What is the most misunderstood thing about AI right now?
Max Lamparth: The most misunderstood thing is that people treat AI as one single thing. When most people say “AI” today, they picture a chatbot, but not every AI system is a “language model,” and what is called “language model” today commonly refers to modern AI agents that are very different from ChatGPT a few years ago. In particular, these agents are no longer purely statistical pattern matchers but a kind of quasi “neuro-symbolic AI.” They have become a blend of neural networks that read and write natural language and symbolic tools like deterministic software compilers and knowledge graphs.
So, the system that feels like it is just talking to you is often parsing your request, calling real tools, and increasingly acting in the world. Understanding that shift matters, because it changes both what these systems can be trusted to do and where they are still likely to fail.
Donna Obeid: Where is AI already creating real value, and where is the hype still ahead of the evidence?
Max Lamparth: AI is already creating real value wherever the goal is clear and the answer can be checked. We have seen great progress in agentic systems built on language models for smaller everyday tasks, a kind of better Google search that can also complete smaller work for you. It is genuinely strong in domains where results can be verified, especially code and math, and in many task-specific applications such as drug discovery and weather forecasting. The common thread is that in these settings you can tell whether the system got it right.
The hype runs ahead of the evidence in the opposite kind of problem. AI is still weak in non-verifiable domains where there is no clean right answer to provide a training signal, like creative writing or military decision-making, and in open-ended exploration, such as acting as an autonomous business consultant. In those domains, you have to learn what counts as a good answer from human judgment rather than check it, and that learned signal is both hard to pin down and easy to game. These are not small gaps to be patched. They sit on top of big, unsolved research questions. So, I would say the value is concentrated where success is measurable, and the hype lives where it is not.
Donna Obeid: What risks worry you most in the next two to five years: misinformation, labor disruption, cybersecurity, bias, autonomy, or something else?
Max Lamparth: I see a pressure to deploy AI prematurely, pushed by hype and by a sense of economic or geopolitical pressure, often before reasonable safety measures are in place. The logic tends to be the same: “If we do not use it, someone else will,” so there is supposedly no choice but to speed up adoption. This is clearest in high-stakes settings where answers are hard to verify, like real-world automation, mental health support, and autonomous military decision-making. The cost of being wrong there is not a bad search result but a harmed patient, an unnecessary escalation, or dire individual consequences.
The second risk is power concentration. Fewer and fewer companies control these systems and capture their benefits, even though the field was built on academic research and on data or labor drawn from society at large (given ongoing legal processes, allegedly not always under fair use). Just this month, Anthropic limited the usage of their frontier models for academic AI research, initially even setting their models to silently be less capable for AI research tasks, causing public backlash.
Besides gatekeeping the very knowledge that should be democratized, if frontier AI labs continue on this path, they risk dictating what research topics are viable. This issue also connects directly to labor, where the natural incentive for companies is to automate jobs to raise margins rather than to build AI that augments and works alongside them. Concentrated control plus an automate-first incentive is, to me, the more immediate danger.
Donna Obeid: How should policymakers regulate AI without slowing useful innovation?
Max Lamparth: The biggest mistake would be writing rules that only one or two of the frontier AI labs can realistically comply with. That is how you get regulatory capture, where incumbents lobby for requirements that lock out competitors and slow the diffusion of AI across the rest of the economy. Concentrated power, more than regulation itself, is what stifles useful AI innovation, in my opinion.
A better approach does not pick winners but sets reliability targets and holds providers accountable through independent third-party audits and clear liability when their models cause harm. Independence matters because self-reporting fails. In Evaluation Cards, a governance effort I recently helped build, we found that roughly 96 percent of public results lacked the basics needed to reproduce them, worst when developers graded their own models. Obligations need to scale to the risk an AI system poses rather than to who can afford the paperwork. That creates trust in the technology without the government deciding who gets to build it.
The real goal is democratization and closing the gap between what these systems can (or could) do and how reliably they do it. Widely available, trustworthy AI is what turns raw capability into value that people and the economy can actually count on.
Donna Obeid: How will AI change the future of work, especially for students, researchers, writers, engineers, and knowledge workers?
Max Lamparth: I worry less about which tasks get automated and more about cognitive security and overreliance. We have not sufficiently studied how these systems quietly push people toward overreliance through sycophancy, overconfidence, and a slow drift from tool to companion. Sycophancy and overconfidence are not random quirks but artifacts of how these systems are trained, so part of the fix lives upstream, in the training itself, not only in the workplace. For students, researchers, writers, and engineers, the risk is not that the work disappears but that the judgment behind it atrophies. Exactly how this plays out is still to be determined.
What I am sure of is that if the choice is left to the companies alone, automation will be the default over uplifting individual decision-makers and knowledge workers, because automation is easier to sell and to scale. The better future is one where these systems sharpen human judgment instead of replacing it, but that outcome has to be designed for, not assumed.
Donna Obeid: Finally, what gives you optimism about AI? What do you hope AI looks like ten years from now?
Max Lamparth: What gives me optimism is the size of the upside once we make it work. In a decade, reliable AI could democratize what used to require scarce expertise, in education, health care, and access to scientific knowledge, and it could accelerate discovery and productivity on the scale of the Industrial Revolution.
In the meantime, I hope we can shift the question from “what AI can do” to “who does it benefit” and “how does it shape society”—and that the answer is most people rather than a few companies. The future I am optimistic about is widely available, trustworthy systems that sharpen human judgment instead of concentrating expertise. That is not automatic, but it is within reach if we build for it.
