← Back to Contents
Note: The first drafts of these discussion questions were generated using Claude (Anthropic's AI assistant) and then reviewed and edited for the in-class use of this course.
Week 11 • Discussion

💬 Week 11 — Discussion Questions

To think about before class — for both halves of Week 11

These are example discussion points for you to think about before class. You are not expected to engage with all of them — pick the ones that speak most directly to your own research, and bring two or three rough answers to the in-class session. The full description of how to use these pages, including what the question tags mean, is on the Week 1 Discussion page.

Part A — The Future of AI in Research

11.1 — What the Future of AI in Research Might Look Like

  1. Calibrate Pick a striking AI-in-science announcement you have seen in the past month (a paper, a press release, a vendor blog, an X/Twitter thread). Place it in Real, Overclaimed, or Aspirational. Be explicit about which sentence in the announcement moves it between buckets, and what evidence would shift your placement.
  2. Apply Take Mollick's worked example: GPT-5 Pro finding a previously-undetected error in his own published job-market paper. If you ran the same exercise on your own most recent paper or research output, what would you expect the system to catch that a human reviewer wouldn't — and where would you be most suspicious of its output?
  3. Critical The “AI as a substantive collaborator is here” framing in the workflow-level info-box leans on practitioner reports (Mineault, Mollick, Karpathy) plus one early-stage randomised study (the Anthropic social-sciences sample). What is the strongest objection you can make to treating these sources as evidence for the claim, rather than as advocacy by people who already believe it?
  4. Connect How does the Real / Overclaimed / Aspirational framework in 11.1 relate to the calibrated-reading habit you developed in Weeks 9 and 10? Is there a useful distinction between calibrated reading of model behaviour (Week 9) and calibrated reading of news about model behaviour (11.1)?

11.2 — Speculative Futures: A Reading Guide

  1. Calibrate Of the speculative-futures works covered in 11.2 (Krenn, Wang, Morris, METR, AI 2027, Clune, Russell, the Royal Society piece, the Africa Declaration), pick one that makes a forecast you would actually bet against. State the bet, the timescale, and the condition that would settle it.
  2. Apply Imagine you are writing a five-year research roadmap for your own field. Which of Krenn's three dimensions — computational microscope, resource of inspiration, agent of understanding — is most useful as scaffolding for that roadmap, and which is least? Why?
  3. Critical AI 2027 commits to a specific year-by-year scenario that the authors themselves have already partially walked back. Is the lesson here that confident forecasting is always bad, or is there a kind of confident speculation that does productive work even when its specific predictions are wrong?
  4. Connect The METR doubling-every-seven-months result is the cleanest example of a falsifiable forecast in this literature. How would you design a similar falsifiable forecast for AI progress in your own field — one that could plausibly be checked by 2028?

11.3 — The Shifting Research Landscape: Policy, Peer Review, Integrity

  1. Calibrate The He & Bu (PNAS, March 2026) paper finds that journals with and without AI policies show statistically indistinguishable growth in detectable AI-written content. Read this finding in two ways: (a) “policies don't work”; (b) “policies have not yet been enforced.” Which reading do you find more honest, and what evidence would change your mind?
  2. Apply Pick the journal you would most realistically submit your first research paper to. Find its current AI policy. What is the gap between what the policy formally requires and what you would actually need to do to comply with the spirit of the ICMJE / COPE international consensus (no AI author / disclose / human accountable)?
  3. Critical The hidden-prompt-injection case in 11.3 shows researchers actively gaming the rules they expect their colleagues are also violating. What does this tell us about the limits of policy-as-disclosure? Is the right institutional response more enforcement, or a different kind of trust architecture entirely?
  4. Connect Week 4 asked you to commit to a personal disclosure practice; this lesson shows that the NRF currently leaves the disclosure question entirely to your own judgement. Hold the two together. Does the personal-disclosure commitment you made in Week 4 translate cleanly to the NRF context, or does the absence of a domestic framework change what you would commit to and how?

Part B — Africa's Sovereign AI Capacity

11.4 — Sovereign AI Capacity, and Why Compute Is the Floor

  1. Calibrate The 11.4 argument is that “sovereign AI” framed primarily around infrastructure ownership would not have prevented the Worldcoin Kenya violation. What kind of sovereignty framing would have? Sketch the actual policy instrument it would have required.
  2. Apply Suppose you had to write a one-page case for an African institution about which compute-access pathway to invest in over the next eighteen months. Pick from the 11.4 options (CHPC, Cassava GPUaaS, DataSpires, free-tier cloud) and defend the choice. What would change your mind?
  3. Critical The relational-sovereignty argument is grounded in Ubuntu philosophy. What is the strongest objection you can make to it from a different intellectual tradition? Does the objection threaten the argument's substantive content, or just its rhetorical positioning?
  4. Connect Moiloa's “design for efficiency from the start” position implies the right scale of African AI is small, efficient models on accessible compute. Awarri's N-ATLAS framing (introduced in 11.5) is closer to “African frontier model.” Are the two positions in tension, or are they doing different jobs?

11.5 — Data, Languages and African Model-Building

  1. Calibrate Hussen et al. find that 42 African languages receive “meaningful support” from current LLM families — out of roughly 2,000. The 98% gap is the headline. Make the strongest case that this number undersells the picture (more languages are supported than counted), and the strongest case that it overstates it (the 42 are not really supported either). Which case is stronger?
  2. Apply If you wanted to make a useful contribution to one of the African-language benchmarks in the 11.5 table (AfroBench, IrokoBench, AfriMTEB, AfriSpeech-200, NaijaVoices, etc.) in the next twelve months, which would you choose to extend? Be concrete about the language, the task, and the kind of data you would need.
  3. Critical The Esethu Framework asks commercial users — especially non-African ones — to pay back into the dataset community. Steelman the objection from a researcher who wants to use the data freely, then answer it. Where does your answer stop being a principled defence and start being an ad-hoc one?
  4. Connect The Indigenous-data-sovereignty arc covered in 11.5 (Te Hiku Kaitiakitanga → CARE Principles → Esethu) is explicit about borrowing from outside the African context. What does that lineage gain by being explicit about its borrowings, and what does it risk?

11.6 — Policy, Institutions, and Talent

  1. Calibrate The South African AI Policy withdrawal case (April 2026) is presented as a worked policy-document-scale instance of the policy/practice gap from 11.3. What other framings of the same case would you also defend — e.g. as a story about institutional capacity, journalistic accountability, or the political economy of consultancy?
  2. Apply Pick the African country whose AI policy or strategy is most relevant to your work. Find one specific commitment in that document that you could imaginably contribute to. What would your contribution look like in concrete terms over the next two years?
  3. Critical The 11.6 framing reads African AI work as “distributed across continent and diaspora” rather than “Africa losing its AI talent to the North.” Is the distributed framing accurate to where substantive AI research is actually being done by Africans, or is it putting an optimistic gloss on a structural problem?
  4. Connect The Lelapa AI six-co-founder team draws across Wits, UP, Brown, and the Masakhane network. The “quadrilateral of substantive PhD activity” framing names Wits, UP, UCT, and AIMS SA. Does the Lelapa example reinforce the quadrilateral or complicate it — and what does the Brown involvement say about how you should read “African AI capacity”?