For Tinglong Dai, an AI expert and the Bernard T. Ferrari Professor of Business at the Johns Hopkins Carey Business School, the future of artificial intelligence depends on how we design AI systems and, increasingly, on whether people keep the skills that make those systems useful.
His recent research and talks make two arguments: AI needs guardrails, and people need to keep exercising their own judgment.
AI is moving from advice to action
The stakes rise when AI stops generating answers and starts taking actions.
In a paper in the October issue of Production and Operations Management as the lead article, Dai and co-authors David Simchi-Levi of MIT, Michelle Xiao Wu of Purdue University, and Yao Xie of Georgia Tech develop what they call “assured autonomy,” a framework, grounded in operations research, for letting generative AI operate in the real world while keeping it safe and accountable.
An AI agent might place orders, route vehicles, allocate clinical resources, coordinate logistics, or help balance power grids. Each of those tasks comes with hard constraints and real uncertainty, and an error can be costly.
This shift creates what the authors call the “autonomy paradox”: the more autonomy an AI system has, the more structure it needs.
AI “is moving from advice to action,” the researchers write. As autonomy increases, AI systems need more explicit constraints, monitoring, stress testing, and clear rules for when to escalate a decision to a human.
Build safety in from the start
Dai argues that one of the biggest challenges in AI safety is preparing for failures that developers don't see coming.
“The key question isn’t really whether a safeguard can stop a rogue agent once something goes wrong,” he says. “It’s whether safety is built into the system architecture so that dangerous behavior is difficult to initiate, propagate, or escape in the first place. Retrofitting safety is always going to be harder than designing for safety from the beginning.”
Dai's answer is to stress-test AI against worst-case scenarios. In this adversarial approach, a second system tries to find every way the first one could fail.
The adversary looks for unusual interactions, shifts in the operating environment, and vulnerabilities that no one anticipated during development.
Most AI safety tools on the market, Dai notes, catch or contain rogue behavior that developers can anticipate. No one can predict every failure, so the harder task is building systems that hold up when something unforeseen happens.
“I see something like NVIDIA’s approach as useful,” Dai says, referring to the company's new AI safety platform, “but still only part of the answer. The bigger question is whether we can build AI systems that remain safe even when the failure mode is something we never anticipated.”
The human side of AI safety
Dai's thinking about AI safety extends to the people who use it. At a September seminar at the University of Houston's C.T. Bauer College of Business, he spoke about whether AI will take our jobs. The discussion quickly turned to a harder question.
If AI can write, calculate, analyze, research, and make increasingly sophisticated decisions, what happens to the human skills we no longer have to practice?
In a recent LinkedIn post reflecting on the discussion, Dai wondered what knowledge, skills, judgment, creativity, and responsibility people want to keep as machines take on more.
“As AI becomes more reliable, we get fewer opportunities to exercise our own skills and judgment,” he wrote. “If we’re not careful, some of that human capital can go extinct.”
This risk is quieter than a rogue agent or a system crash. People grow less capable because a machine has been doing their practicing for them.
Creating “AI-free islands”
Dai proposes what he calls AI-free islands.
The phrase refers to protected settings, such as a class session, an exam, or a hospital case conference, in which people put AI aside and think, write, calculate, argue, create, and connect on their own.
The islands are one half of an all-or-nothing approach. Inside them, people work without AI and build the judgment that only grows with practice; everywhere else, Dai would use AI radically. In his view, the middle is the worst place to be, because a little AI everywhere erodes skill without capturing the gains.
The idea applies directly to business education, where AI can speed up research, analysis, and writing. If a machine can produce a first draft in seconds or solve a problem with a prompt, learning to do those things yourself may seem less valuable. Dai thinks its value is growing.
The more capable AI becomes, the more people need to master the underlying skills well enough to question the machine, recognize when something is wrong, and make decisions when the machine encounters something it hasn't seen before.
“The future of human capital is more about being human than ever,” Dai wrote.
