On April 30, the MIT Schwarzman College of Computing’s Social and Ethical Responsibilities of Computing (SERC) initiative convened a full-day research symposium that cut through the noise surrounding artificial intelligence to address a fundamental question: as AI systems grow more capable, what role does human judgment, values, and understanding play in shaping their trajectory? The answer, according to the cross-disciplinary group of researchers, panelists, and speakers who participated, is that the human component is not merely important but irreplaceable. The symposium, which featured research talks from SERC’s latest seed grant recipients, panels on AI alignment and education, a keynote address from Cornell’s Jon Kleinberg, and a student poster session, made clear that the most pressing challenges in AI today are not purely technical, but deeply human.
Aligning AI with Human Values — and Deciding What Those Values Should Be
A central theme of the symposium was the difficulty of aligning AI systems with human values, a problem that grows more urgent as these systems are deployed in high-stakes domains. A panel moderated by Dylan Hadfield-Menell, associate professor of EECS, brought together an interdisciplinary group to wrestle with the ethical questions that underpin alignment research. Who decides which values are encoded into an AI system? How do those values survive the translation from human intention to machine behavior?
Iason Gabriel, a philosopher and research scientist at Google DeepMind, offered a useful analogy: a judge. “You want a judge to have good character, but to still interpret the rules. A reasonable person, though not necessarily the best person who ever lived,” Gabriel said. “When it comes to AI, it’s not appropriate to model it as perfect. AI should be doing what we tell it to do, while using its character to interpret according to our moral values.” This framing positions AI not as a flawless oracle, but as a system that must exercise judgment within boundaries, much like a human professional.
Bailey Flanigan, assistant professor of political science in a shared appointment with the MIT Schwarzman College of Computing in EECS, pushed the discussion further. For her, the most critical problem in AI alignment is not technical but political: “Resolving fundamental questions on who is entitled to govern different types of AI systems in the first place.” This reframes alignment as a question of legitimacy and democratic control, not just engineering.
Bernado Zacka, associate professor of political science, added a cautionary note about the pace of deployment. Given the momentum of AI and the complexity of institutional systems, he argued that “one of the most urgent problems is understanding the wisdom contained in the systems we are replacing, and why they function the way they do.” The panelists, while broadly optimistic about the trajectory of alignment research, emphasized that human oversight, democratic deliberation, and institutional humility are essential components of any responsible AI deployment strategy.
AI in Education: Offloading Versus Uplifting
A separate panel on AI and education zeroed in on a dilemma that instructors at every level are now confronting: are students using AI to bypass learning, or can these tools be integrated in ways that genuinely deepen understanding? The panel featured MIT faculty alongside Marta McAlister, director of Gemini for Education, and explored how AI is already reshaping classrooms and what that means for instructional design.
Samuel Madden, faculty head of computer science in EECS and the MIT College of Computing Distinguished Professor, described the phenomenon of “cognitive struggle” — the process by which learning happens through trial and error, persistence through difficulty, and eventual mastery. “Students now, when they hit that wall, their first instinct is to ask AI,” Madden said. “They don’t see this as excelling in this process, and they haven’t actually acquired the skill you’re assessing.” The challenge for instructors is to design learning experiences that provide enough challenge to resist the temptation of AI shortcuts while still being achievable.
Eric Klopfer, director of the Scheller Teacher Education Program and the Education Arcade at MIT, echoed this concern, noting that critical thinking is increasingly being bypassed in the output-driven workflow that AI enables. His recommendation was direct: “Some core content has to go. We keep adding, instead of parsing or pruning.” Rather than layering AI tools onto an already overloaded curriculum, Klopfer argued that educators need to rethink what foundational knowledge truly matters.
Moderator Justin Reich, director of the Teaching Systems Lab and associate professor in Comparative Media Studies, observed that most students already know that relying on AI for answers is not ideal, but that awareness alone does not change behavior. His proposed solution was to bring students into the conversation about how AI is implemented in their learning environments, creating a reflective exchange that equips them to make intentional choices about when and why to use these tools.
Pat Pataranutaporn, head of the Cyborg Psychology research group at the MIT Media Lab, offered a broader reframing: “AI is not just one thing. It can and should be designed differently to promote things like creativity and critical thinking. What we measure, and how, shouldn’t be about getting the answer right. We should think about what it would really mean for a student to learn these days.” This perspective suggests that the problem is not AI itself, but the narrow way it is often deployed and evaluated in educational settings.
When AI Mimics Human Reasoning: Is Close Enough Good Enough?
Jon Kleinberg, the Tisch University Professor of Computer Science and Information Science at Cornell University, delivered a keynote address titled “AI’s Models of the World, and Ours” that examined what happens when the way an AI system understands a task diverges from the way a human does. Using chess as a starting point, Kleinberg illustrated how modern chess engines can operate at superhuman levels, but their strategies are often opaque to human partners. When a chess engine hands a turn over to a human teammate, the human cannot infer the engine’s reasoning and is left disoriented.
“The danger of human-algorithm teams is that when the human takes over, the algorithm knows what it wants to do next, but the human doesn’t,” Kleinberg explained. This mismatch between the AI’s model of the world and the human’s model of the world creates a failure mode that is not about capability but about communication and shared understanding. Kleinberg drew a memorable parallel to “The Fellowship of the Ring,” where Gandalf entrusts a critical quest to a group of adventurers and then disappears, leaving them momentarily paralyzed. The analogy captures the experience of being handed control without context or explanation.
The deeper question Kleinberg’s talk raised is whether an AI system that mimics human reasoning well enough to produce correct outcomes actually understands the domain in which it operates. If a chess engine can reliably deliver checkmate, does it matter that its human partner cannot follow its logic? The answer, in real-world applications where human judgment must reassert itself at critical junctures, is a clear yes. The gap between predictive simulation and embodied human knowledge remains one of the most significant unresolved challenges in human-AI collaboration.
What This Means for the Industry and Practitioners
The MIT SERC symposium did not offer a single, unified solution to the challenges it identified, but it did provide a coherent framework for thinking about them. The human component in AI is not a limitation to be engineered away, but a necessary feature of any system that is intended to serve human purposes. Alignment, education, and collaborative reasoning are not problems that can be solved purely through better models or larger datasets. They require ongoing input from philosophy, political science, education, and the broader public.
For professionals working with AI systems, the symposium’s key takeaway is that deploying AI responsibly means investing in the human infrastructure around it: clear communication about what the system can and cannot do, thoughtful design of handoff points between human and machine, and a willingness to question whether the systems we are building truly align with the values of the people they affect. For educators and institutional leaders, the message is to resist the temptation to treat AI as a one-size-fits-all solution and instead engage learners and stakeholders in a reflective dialogue about how these tools should shape the future of work, learning, and decision-making. The symposium’s strongest signal is that the most important work in AI is not just about making machines smarter, but about making the human systems that surround them wiser.
