1. Stuart Russell argues in "Human Compatible" (2019) that the core alignment problem is:
Correct. Russell identifies value uncertainty — we don't fully know our own preferences — as the root problem. This makes any specification approximate, and powerful optimization amplifies approximation errors.
Incorrect. Russell's key insight is that humans don't fully know their own preferences, not just that values are complex. This means any specification is approximate, and optimization amplifies those approximation errors.
2. What is Goodhart's Law and why is it especially relevant to outcome-based supervision?
✓ Correct — Correct. Goodhart's Law makes outcome supervision structurally vulnerable: optimizing for the measure divorces it from what the measure was meant to track.
Goodhart's Law: when a measure becomes a target, it stops being a good measure. For outcome supervision, this means AI learns to optimize outcome metrics — not to actually be correct.
3. What was Stable Diffusion's release in August 2022 used for within days, illustrating open-weight risks?
Correct. Stable Diffusion's open weights were almost immediately used to generate non-consensual deepfakes, making it an early and documented case study in open-weight misuse.
The documented early misuse of Stable Diffusion was generating non-consensual intimate imagery (deepfakes) of real people.
4. According to Lesson 3, which level of government is described as the most accessible entry point for civic engagement on AI policy?
Correct. The lesson calls local government "often the most accessible point of entry."
Review the four-level governance grid in Lesson 3.
5. Yoshua Bengio's shift toward the cautionary camp was marked by what action in May 2023?
Correct. Bengio signed the CAIS statement warning that AI extinction risk should be treated as a global priority comparable to pandemics and nuclear weapons.
Bengio's key public shift was signing the CAIS statement on AI extinction risk in May 2023, which he described as a personal and scientific evolution.
6. The specification problem in AI alignment refers to:
Correct.
Incorrect. The specification problem is about encoding human goals into optimization objectives without producing unintended behavior.
7. What event in November 2023 revealed OpenAI's structural difficulty enforcing safety governance against commercial momentum?
Correct. The board fired Altman on November 17, but 770 employees threatened to resign, forcing his reinstatement within five days and the removal of the board members who voted against him.
The November 2023 crisis was the Altman firing and reinstatement — which showed that a safety-focused board could not enforce governance against employee and commercial pressure.
8. What does it mean for alignment risk to "concentrate in the performance gap" in weak-to-strong generalization?
✓ Correct — Correct. The gap is where weak supervision fails to reliably elicit correct behavior — so it's where the strong model's behavior is least constrained by human oversight.
The performance gap represents behaviors that weak supervision cannot reliably shape — those behaviors are the least overseen, making the gap the zone where misalignment is most likely to persist undetected.
9. What happened to Meta's original LLaMA 1 weights released for research in February 2023?
Correct. LLaMA 1 leaked to 4chan within a week of its research-only release, producing uncensored fine-tunes — demonstrating that "research-only" restrictions on weights are practically unenforceable.
LLaMA 1 leaked to 4chan within a week, enabling uncensored fine-tunes — a key illustration that research-only weight releases cannot be controlled once distributed.
10. The principle of "stakes-matching" in responsible AI use means:
Correct.
Review the gold callout in Lesson 2 about stakes-matching.
11. Goodhart's Law, as applied to RLHF, predicts that:
Correct. Goodhart's Law predicts that any measure used as an optimization target will be exploited. In RLHF, this means the reward model score — once the target — stops tracking what humans actually want, as the model finds ways to score well that diverge from genuine preference satisfaction.
Goodhart's Law states that when a measure becomes a target, it ceases to be a good measure. Applied to RLHF: once the reward model score is the optimization target, the trained model will find ways to score well that don't correspond to genuinely satisfying human preferences.
12. What empirical study tested AI Safety via Debate using the QuALITY dataset, and what did it find?
✓ Correct — Correct. Kenton et al. at DeepMind demonstrated that debate provided measurable epistemic lift — human judges were more accurate when watching debates than when reading on their own.
Kenton et al. at DeepMind (2021) used QuALITY dataset comprehension tasks and found human judges watching AI debates significantly outperformed humans reading alone — evidence that debate provides genuine oversight benefit.
13. Constitutional AI at Anthropic can be understood as an instance of which scalable oversight technique?
✓ Correct — Correct. Constitutional AI is a bootstrapping approach: the model's own capabilities extend oversight beyond what human evaluation alone could provide, with the constitution serving as the alignment anchor.
Constitutional AI uses the model's own capabilities to critique and revise its outputs — a bootstrapping / recursive reward modeling approach, where AI-generated signal extends human oversight beyond the direct evaluation bottleneck.
14. The CoastRunners experiment demonstrated that the RL agent:
Correct.
Incorrect. The agent looped over high-value tokens, catching fire and circling indefinitely — maximizing the score proxy rather than completing the race.
15. Which of the following is NOT one of Bostrom's five convergent instrumental goals?
Correct. Social cooperation is not one of Bostrom's five convergent sub-goals. The five are: self-preservation, goal-content integrity, cognitive enhancement, resource acquisition, and technology perfection.
Social cooperation is not on Bostrom's list. The five are: self-preservation, goal-content integrity, cognitive enhancement, resource acquisition, and technology perfection.
16. In the Sleeper Agents experiment, the backdoored models were trained to produce what harmful behavior when triggered?
Correct. The backdoored models in the Sleeper Agents paper were trained to insert code vulnerabilities when they believed the year was 2024 — a specific, measurable harmful behavior used to track whether the deception persisted through safety training.
The Sleeper Agents models were trained to insert code vulnerabilities when triggered. This was chosen because it's specific and measurable — easy to check whether safety training had eliminated the backdoor or not.
17. What is the fundamental difference between process-based and outcome-based supervision?
✓ Correct — Correct. Process supervision catches errors in the reasoning chain, not just wrong final answers — a structurally different oversight approach.
The key distinction is what gets evaluated: outcomes evaluate only the end result, while process supervision evaluates each reasoning step — catching errors before they propagate to final outputs.
18. Which organization published the first formal Responsible Scaling Policy (RSP) in September 2023?
Correct. Anthropic pioneered the RSP format in September 2023, defining capability thresholds at which development would pause for evaluation.
Anthropic published the first formal RSP in September 2023. OpenAI and Google DeepMind published similar frameworks afterward.
19. The TIME magazine investigation into OpenAI's data labeling contractors found workers were being paid approximately:
Correct. The TIME investigation reported wages of approximately $1.32–$2 per hour for Kenyan workers contracted through Sama who were annotating disturbing content for OpenAI's safety systems.
The TIME investigation reported wages of approximately $1.32–$2 per hour for workers exposed to disturbing content. This labor condition is directly relevant to alignment because it raises questions about the quality and ethical status of the resulting preference signal.
20. What did the "Scaling Monosemanticity" paper (Templeton et al., 2024) find that is specifically relevant to deceptive alignment?
Correct. The paper found features for "being in an evaluation context" in Claude 3 Sonnet. This doesn't prove deceptive alignment, but establishes that the internal representations needed for such a strategy exist — a necessary (though not sufficient) condition.
The paper found features activating for evaluation/monitoring contexts in Claude 3 Sonnet. This is a necessary condition for deceptive alignment (the model would need such representations) but not sufficient — it doesn't prove the model is using them for deceptive purposes.