Well, my sabbatical is officially over. I’m really grateful for a whole semester to reflect on what it means to teach in 2026 and translate that into concrete suggestions for my own courses. But it’s finally time to put my conclusions into practice!

In my first post in this series, I stated my priors going into this redesign. Here’s a summary of where I’ve landed since January:

• This is an ongoing adaptation. I need to prepare my students for the job market they’re entering, not the one I think will exist in 2-3 years. A one-time break from the past will not work.
• Developing statistical judgment is now the primary goal of my teaching. I used to treat this as a happy consequence of learning the mechanics of practical statistics. I now think that’s wrong, and I need to explicitly teach the judgment calls involved in specifying, understanding, and critiquing an analysis.
• Developing AI skills is an insufficient component of statistical expertise, but also a necessary one. I don’t think facility with AI is a substitute for domain expertise, at least not yet. But I also believe that learning how to use these machines is an important step toward making them work well. I have seen too many domain experts get misled or frustrated by AI to believe that you can just “figure it out.”
• Unfortunately, developing statistical expertise and AI literacy at the same time can undo one another. These tools are too helpful, too opinionated, and too opaque to use while developing the knowledge required to push back against poorly specified analyses or bad judgment calls by an AI. We’re all just a few “sure, go aheads” from removing ourselves from the process entirely.

Here’s how I’m putting these takeaways into practice. My sabbatical was primarily about changing what I teach, but I’d be remiss if I didn’t also change how I teach. I’ll start with that.

Changing how I teach

First, a hard truth shared by others I spoke with this semester: at this point, there are very few things that we can reliably assess through take-home work. I know many incredible educators who create high-trust class environments where students are genuinely invested in their own learning and there are strong norms against cheating. Even then, the pernicious ubiquity of AI in our browsers, our word processors, and even our spell checkers makes it incredibly difficult to capture what a student is thinking and how they are growing.

The silver lining, in my view, is that a whole lot of learning can happen outside the classroom if incentives are aligned. I may not be able to use a take-home paper as an artifact to assess my students, but my students can write a take-home paper knowing that they’ll need to defend and extend their choices via an in-person conversation. In some ways, this is a further realization of the flipped classroom, wherein students must take more ownership of their learning outside the classroom to succeed within it. This can be a hard thing to do if you’re not used to being metacognitive about your learning, but I can’t help but think this will be one of the most important skills that humans will need in the coming decades.

So here’s what I’ve changed:

• Pre-class exercises, where students engage with material coming into class, are being expanded via AI-narrated presentations that we’ve been building on our platform, PingPong. The videos include slides that I would normally deliver in class, checks for understanding that students respond to, and a way for students to ask the AI questions about the slides in real time. Our developer Evangelos Kassos has done amazing work, and I’m really excited to put it in front of my students.
• Weekly problem sets are now weekly study problems that are not graded and exist as a learning tool. I’m using a platform called Pensive to host them, where students can work on problems and have an AI check their answers against the real solutions, giving them feedback, suggestions, and encouragement to meet with the course teaching team.
• Every two weeks, there will be a 45-minute quiz based on the content from the last two weeks of study problems. This is where the assessment happens, and the best way for students to prepare is to simply do the weekly problems carefully and authentically. There will also be a final exam at the end of the semester, which we deliver as a two-stage exam.
• We do have a final project involving a statistical analysis for a policymaker based on a real-world dataset. In addition to students using AI to complete this analysis (more in the next section), the project will now be accompanied by an oral defense.

Changing what I teach

I’m feeling pretty confident about the changes to the course’s delivery and assessment, but that’s not the hard part. The hard part is changing my curriculum to adapt to this changing landscape without overcorrecting. I strongly believe that faculty should be spending as much time worrying about whether they should be teaching what they’re teaching as they do worrying about cheating. It might be true that what we teach is essential, timeless, and unaffected by the advent of the most powerful technology of our lifetimes. But I doubt it.

Rather than list what I’m changing, here is the new course overview from my redesigned course syllabus:

API-201 introduces a range of analytic tools commonly used to inform public policy issues. The course focuses on probability theory, decision analysis, statistical inference, and assessing evidence, with an emphasis on how they are applied to practical policy questions.

The course develops two distinct but related sets of skills: the statistical reasoning you need to be able to exercise yourself, and the skills required to direct AI tools to do analytic work on your behalf. The first is what makes the second possible, so the course will slowly transition from the first to the second throughout the semester. An AI tool can execute an analysis in seconds, but you will first need to develop the judgment required to design, direct, interpret, and critique an analysis.

Statistical reasoning and judgment, developed without AI assistance. By the end of this course you will be able to:

1. Frame a broad descriptive policy question (such as “What has happened to crime rates in the U.S. in the past 30 years?”), determine the most appropriate analysis to answer it, plan and map out the components of the analysis, identify the most salient findings and patterns that emerge, and present them effectively to policymakers.
2. Use probability and decision analysis tools to better tackle real-world personal and policy decisions involving uncertainty.
3. Critically consume policy studies, papers, and reports in which statistical analysis is used — including recognizing what a given statistic does and does not tell you, and generating plausible alternative explanations for a reported result.
4. Exercise statistical judgment in unfamiliar settings: recognize when a policy question can be informed by data, articulate what a plausible result would look like before you see one, and identify the assumptions an analysis rests on.

Working with AI tools. By the end of this course you will be able to:

1. Direct an AI tool in conducting a statistical analysis, including effective prompting, managing context and model choices, and refining outputs.
2. Use AI to produce analyses that are high quality and reproducible, documenting the data, decisions, assumptions, and discarded approaches so that another person, or your future self, can follow the work and verify the result.

Here are some of the considerations that went into this new syllabus:

• Since this is an introductory statistics course, I am focusing more on the first set of skills (statistical reasoning) and somewhat less on the second (working with AI tools). I will be teaching a course on Agentic Data Analysis in the spring, which will be the reverse of this.
• I am frontloading content that I believe does not require facility with AI to understand well: probability and decision analysis. While both will benefit from exercises showing simulations and sensitivity analysis with AI, I don’t view facility with AI as a necessary part of being able to think about uncertainty in a decision-making context. Having this up front will make it easier for me to keep AI tools out of student work as they start to build statistical judgment.
• The second half of the course is generally about working with data in the real world. This will include the biggest shifts in the course. In particular, I will not be teaching a coding language, and I will not be teaching Excel. Instead, my plan is to teach data structures in a more abstract sense, and ways in which we can manipulate data to get the information we want. This will be loosely based on the verbs of the tidyverse, but will not be wedded to that language or even those operations.
• I envision this data unit including a lot of pre-specifying expectations before working with AI (e.g., after I ask AI to collapse this data, how many observations should I expect to see in the output?), and I will assess students based on their ability to pre-specify an analysis in the right way. This will include building in measures of sampling fluctuations, and eventually some practice in reading actual papers, culminating in the final analysis that they do themselves.
• In order to make room for all this, I will be spending much less time on sampling distributions, calculations of standard errors and p-values, and the mechanics of hypothesis testing. This may sound sacrilegious for an introductory statistics course, but given the limited time I have with students, I believe that this is a tradeoff worth making.

My writing going forward

So that’s where I’m at after grappling with this stuff for eight months. With the semester starting, I don’t expect to write as often on here as I have during my sabbatical. My intention is to send short updates on how things are going, and reflections on new AI capabilities as they emerge.

I’ll also be experimenting with using AI tools to streamline my own teaching setup. I’m not sure whether I should post those more productivity-oriented notes here as opposed to just my blog — please let me know if you have strong feelings either way!

If you’re making substantial changes to your course, be it quantitative methods or otherwise, I would love to hear about it in the comments or by message me directly. We’ve also put together a small Slack group to share these things as the semester progresses, so please drop me a line if you’d like to join.

Deep breaths; here we go.