MI 598/Summer 2026
Instructor K. Garimella
A 10-week online seminar

Beyond the Hype.

A course about what AI can do, what it can't, and how to tell the difference.

Course
MI 598, 3 credits
Term
27 May to 1 August, 2026
Format
Online, asynchronous
Institution
Rutgers · School of Communication & Information
Canvas
rutgers.instructure.com/courses/400181
Instructor
Kiran Garimella · kg766@rutgers.edu
01
Course

What the course is about.

Predictions about AI keep being wrong, in both directions. This course asks you to update your views as the evidence comes in. By the end you should have a defensible sense of where the evidence is, where it isn't, and how to tell the difference.

We start with the technical arc that produced today's models: predictive AI, scaling and pretraining, reinforcement learning, and agents. Then we turn to the substantive impacts that matter. The material footprint of AI infrastructure. The evidence on AI and labor markets. What happens when AI generates most of our information. And how hundreds of millions of people now use chatbots for emotional support and self-understanding.

The last two weeks separate misuse from misalignment, and build habits for reading AI claims critically. Throughout, we try to keep arguments tied to evidence rather than vibes.

By the end of this course you will be able to:
  1. 01 Distinguish AI methods from applications from impacts, and explain why conflating them produces bad predictions in both optimistic and pessimistic directions.
  2. 02 Explain how today's frontier models were built (pretraining, RLHF, reinforcement learning on verifiable rewards, the rise of agents) in plain language.
  3. 03 Reason about the material costs of AI: data centers, electricity, water, capex, grid interconnection, and what they imply for deployment.
  4. 04 Use primary evidence to evaluate claims about AI's effects on labor, the information ecosystem, and emotional reliance on chatbots.
  5. 05 Distinguish AI misuse (deepfakes, bioweapon uplift, cyber automation, election influence) from misalignment (scheming, alignment faking), and weigh the empirical evidence on each, not the speculative case.
  6. 06 Run a hands-on task with a coding or browsing agent, document its trajectory and failures, and identify “AI snake oil” in marketing and product claims.
02
Schedule

The ten weeks.

Click any row to expand. Keyboard: J/K step through.

Wk
Topic
Title
Marker
    03
    Work

    Five assignments.

    Each contributes 20% of the grade. There are no participation points. I assume you are here because you want to be.

    1. 01
      Reflection
      20%
      Due after Week 02

      Methods, applications, impacts.

      2 to 3 pages. Pick a predictive or generative AI system you actually use, and apply the methods / applications / impacts distinction to a public claim about it.

    2. 02
      Hands-on · Agent task
      20%
      Launched Wk 04 · Due end of Wk 05

      Run a real agent on a real task.

      4 to 6 pages, plus an appendix with your verbatim prompt log. Pick Claude Code, OpenAI Codex CLI, Google Antigravity, Gemini CLI, or a browser-based agent. Give it a task with real stakes for you. Document the trajectory, including the failures.

      • i.Methods. Graded as heavily as the analysis. Which agent, which model, which tools were enabled, what failed, what you did about it.
      • ii.Trajectory. The path the agent took, including dead ends and recoveries.
      • iii.Analysis. What worked, what didn't, and what this changes about your prior on AI capability.
    3. 03
      Reflection
      20%
      Due after Week 06

      Labor and the empirical question.

      2 to 3 pages. Find a published claim about AI's effects on labor, information quality, or emotional support. Weigh what the claim rests on. What would change your mind? What would change the author's?

    4. 04
      Reflection
      20%
      Due after Week 08

      AI in your own life.

      2 to 3 pages. How AI tools have shaped what you do at work, how you decide, what you read, what you write, and how you feel after a long session with a chatbot.

    5. 05
      Final · Policy proposal
      20%
      Due end of Week 10

      An AI policy proposal for an institution you belong to.

      6 to 8 pages. Use the course frameworks. Be specific and implementable. Recommend what the institution should do, what it should not do, what evidence would change your recommendation, and what monitoring it should put in place. Build on the course materials, and more importantly on current sources from outside the course.

    04
    Instructor

    Kiran Garimella.

    KG
    Kiran Garimella · Rutgers SC&I

    I research how information moves through encrypted and low-resource platforms, and what happens when AI gets added to that pipeline. I work on misinformation on WhatsApp, AI-generated political media, and how AI can be applied in solving everyday problems.

    I teach this course because the AI conversation veers between rapture and ruin every few months, and most of the people I talk to (including researchers I respect) feel under-equipped to argue the middle. The middle is where I feel the evidence lives. We should think critically and act rationally based on evidence in order to prepare and equip ourselves for a future disrupted by AI.

    Email
    kg766@rutgers.edu
    Office
    Huntington House 204
    Phone
    come on
    Office hours
    By email appointment
    Canvas
    course site
    Canvas help
    help@oit.rutgers.edu · 833-648-4357
    05
    Readings

    What we'll read.

    About 2 hours per week. You don't need to buy the book; chapters and excerpts will be provided.

    Anchors
    • Narayanan & Kapoor, AI Snake Oil (2024). Selected chapters.
    • Narayanan & Kapoor, AI as Normal Technology (2025). The running essay.
    Selected primary & current (will change)
    1. Sutton, The Bitter Lesson.
    2. A plain-language explainer on the DeepSeek-R1 / o1 reasoning lineage.
    3. Anthropic Economic Index reports.
    4. Brynjolfsson, Li & Raymond on generative AI in customer support.
    5. Acemoglu, The Simple Macroeconomics of AI.
    6. Kapoor & Narayanan, Can AI Do Your Job Yet?
    7. EPRI load-growth forecasts and the RAND data-center inventory.
    8. Greenblatt et al. (Anthropic, 2024) on alignment faking.
    9. Apollo Research, scheming evaluations on frontier models.
    10. MIT / OpenAI affective-use study of heavy ChatGPT users.
    11. Anthropic, Claude Code launch post.
    12. Simon Willison's recent writing on coding agents.
    06
    Policies

    House rules.

    AI tools policy

    This course is about understanding what AI can and cannot do. That means learning to use these tools thoughtfully, not just avoiding them.

    You can: brainstorm, clarify concepts, get feedback on drafts, probe failure modes.

    You cannot: paste AI output as your own work, outsource the reflection itself, or skip the disclosure.

    Disclosure rule: if you use ChatGPT or a similar tool, include a short note (footnote or end-of-doc) saying how and why you used it. The Week 04 to 05 agent assignment requires agent use; the disclosure rule still applies for every other use.

    Grading scale

    GradeRangeDescription
    A> 90 – 100%Outstanding, polished, on time.
    B+> 85 – 90%Substantially above minimum; error-free.
    B80 – 85%Good, above minimum.
    C+> 74 – 80%Minimum standard.
    C> 70 – 74%Barely adequate; errors.
    D> 65 – 70%Disorganized; many errors.
    F< 65%Inadequate.

    Attendance & late work

    Online class. Watch the lectures, do the assignments. If you need an extension and have a real reason, just email me. I am flexible with late submissions as long as it doesn't happen over and over.

    Students with disabilities

    Rutgers Office of Disability Services: ods.rutgers.edu.