What Can AI Do Now, Really?

FYSEMR 54J Subject & Catalog Number

Description

In the time since members of the Class of 2030 were born, artificial intelligence has advanced from barely recognizing a human face in a photograph to creating convincing images and videos of people—real or imagined—doing nearly anything. In just the past few years, the acceleration in the capabilities of Large Language Models (LLMs) has been astonishing, leaving nearly everyone who works with them wondering about the limits of their powers, dangers, and utility. The pace of change is now so fast that even experts are asking a basic question: What can AI do now, really? In this seminar, we will take that question seriously and explore it by applying current AI models to real problems. Students will suggest areas for inquiry across science, policy, economics, health, the arts, or everyday life, motivated by their own interests.

Working individually and in small groups, they will test AI systems on proposed questions, comparing results and tracking where models succeed, fail, or mislead. A mix of questions will be studied. Some will be small and concrete, answerable within a semester or less. For larger questions, the goal will not be to “solve” a problem, but to frame it in a way that makes real progress possible. Over the term, students will develop their questions into case studies, lead discussions, and evaluate results with peers and AI-expert reviewers. Student work will be included in a seminar-summary paper led by Prof. Goodman and submitted to the Harvard Data Science Review. This seminar treats AI not as a computer science subject, but as a system to be tested. The seminar is designed to support sustained investigation rather than quick exposure.

As leader, I will inquire about how questions are framed, how evidence is gathered, and what claims are actually supported by the results. Students will typically investigate the same question using multiple AI tools (e.g., ChatGPT, Gemini, and others), allowing differences in results to become visible. Small, concrete problems allow students to see clearly whether a result works. Larger problems—such as those involving policy or social outcomes—require students to refine the question itself until progress can be evaluated. This process of reformulating questions is an explicit part of what students will learn.

The seminar also spans different kinds of tasks. Some case studies will involve structured, code-heavy work, such as simulating a system and interpreting its behavior. Others will depend more on language and reasoning, where judgment and interpretation matter more than correctness in a narrow sense. Some will involve visual explanation, and a few may attempt time-based media such as video. These latter tasks are intentionally challenging: producing a coherent sequence over time is significantly harder than generating a single image, and often reveals current limits. With twelve students working in groups of three, the seminar will focus on a small number of case studies in depth. The class will narrow a longer list of questions to the strongest and most illustrative four, ultimately informing the seminar-based submission to the Harvard Data Science Review (HDSR).

Students will be invited to comment, optionally, on the HDSR submission, after the seminar concludes. To support this the students during the seminar, each group will meet with Prof. Goodman regularly in weekly office hours, and each student will meet with her individually at least twice during the semester. These meetings will help students refine their questions, interpret results, and stay on track as their projects evolve. Because the tools themselves are evolving, the specific systems used and some readings will remain flexible. Students will be part of deciding what is worth testing as the course progresses.