Module Check: Open Problems and Hard Limits
You have completed Module H4: Open Problems and Hard Limits. Across nine lessons, you examined what frontier AI cannot do reliably, why hallucination is structural rather than incidental, where reasoning breaks under adversarial pressure, how data and compute impose hard ceilings on scaling, why lab performance fails to transfer to real-world deployment, what the interpretability gap means for accountability, how AI infrastructure strains energy and water resources, and which deep unsolved problems the research community is actively working on. This final lesson reviews and consolidates all of it. Begin with the key terms, then test yourself across the module, and close with a synthesis task that asks you to apply everything together.
Flashcards — click each card to reveal the answer
Module Check Quizzes
A language model generates a perfectly formatted academic citation — author names, journal, volume, year — for a paper that does not exist. Which combination of factors best explains this specific failure?
A model trained on 2020 English web data is deployed to classify medical records written in Hausa (a West African language with roughly 70 million speakers but limited digital text representation). Which concept from this module best explains the expected performance gap?
An autonomous drone navigation system passes 10,000 hours of simulated testing without failure. It is then deployed in a city with reflective glass buildings that create lighting conditions not present in the training simulation. It fails within the first week. This is best described as:
A medical AI company argues that its diagnostic model is safe to deploy without interpretability requirements because it has 94% accuracy on a large independent test set. What is the strongest counterargument specifically about interpretability?
A technology company trains each new model generation on text generated by the previous generation, because this avoids copyright issues with internet data. After five generations, researchers find that rare factual knowledge and low-frequency vocabulary have dramatically degraded. This phenomenon is called:
Which pairing of problem and proposed solution is INCORRECTLY matched?
Module Synthesis
The AI Limits Brief
- You will write a structured one-page policy brief addressed to a decision-maker — a hospital administrator, a city official, a school board, a financial regulator, or a military procurement officer. Choose the audience that most interests you.
- Your brief must advise this decision-maker on deploying an AI system in their domain. It must reference at least five distinct concepts from across the module.
- Structure your brief as follows:
- EXECUTIVE SUMMARY (2-3 sentences): What is the core recommendation? Is AI deployment advisable, with conditions, or not advisable at this time for this domain?
- KEY LIMITATIONS RELEVANT TO THIS DOMAIN (one paragraph per limitation, pick 3-4):
- For each limitation: name it precisely, explain how it manifests specifically in your chosen domain, and cite the realistic magnitude of the risk.
- WHAT MITIGATIONS EXIST (one paragraph):
- What technical and procedural safeguards can reduce the risks you identified? Be honest about what mitigations can and cannot guarantee.
- OPEN PROBLEMS THAT REQUIRE CAUTION (one paragraph):
- Identify one unsolved problem from Lesson 8 that is particularly relevant to your domain. Explain why its current intractability should affect deployment decisions.
- CONDITIONS FOR RESPONSIBLE DEPLOYMENT (bullet list, 4-6 bullets):
- Specific, enforceable conditions under which you would recommend deployment — required human oversight roles, mandatory monitoring metrics, restricted use cases, disclosure requirements, review timelines.
- When complete, exchange briefs with a partner who chose a different domain. Each reads the other's brief and writes a one-paragraph response: Do you agree with the core recommendation? What did the brief get right? What did it miss?
- Goal: demonstrate that you can apply the module's technical content to realistic governance and deployment decisions.