1. Based on the 2024 Reuters Institute survey, what is the PRIMARY use of AI in most news organization editorial workflows?
Correct. The 2024 Reuters survey showed AI in newsrooms primarily handling structural workflow tasks — transcription, translation, tagging — while final published content remained human-judged.
The Reuters data showed transcription, translation, and metadata tagging as primary uses — AI handling structural labor while humans retain editorial voice and judgment.
2. What does the EU AI Act, passed in March 2024, use as its primary regulatory mechanism?
Correct. The EU AI Act classifies systems into risk tiers — prohibited, high-risk, limited-risk, and minimal-risk — with obligations that scale with risk level.
The EU AI Act uses risk-based classification: prohibited uses (e.g., social scoring by governments), high-risk systems with strict obligations, and lighter requirements for lower-risk applications.
3. The "Brussels Effect" in AI governance is most analogous to which earlier regulatory precedent?
Correct. The GDPR is the most direct predecessor — global companies adopted GDPR-aligned privacy practices worldwide because serving EU users required compliance regardless of where the company was headquartered. The EU AI Act is expected to replicate this dynamic for AI governance.
Incorrect. The GDPR is the most direct analogy — companies worldwide adopted EU privacy standards to serve EU users, and the EU AI Act is expected to produce the same dynamic for AI compliance.
4. What is the primary distinguishing characteristic of Layer 1 (Frontier Research) in the AI labor market?
Correct. Layer 1 is defined by frontier research — novel work published at top venues — and typically requires a PhD or equivalent research record.
Layer 1 is frontier research: generating novel architectures and techniques, publishing at top venues, typically requiring a PhD. Layer 2 is the largest layer; Layer 3 is growing fastest in percentage terms.
5. ICLR's open peer review on OpenReview.net means that:
Correct. OpenReview.net publishes all submissions, their reviews, ratings, and author rebuttals publicly — making the review process itself a transparent research artifact.
Not quite. ICLR's open review means all submissions, reviews, scores, and author responses are publicly visible on OpenReview.net — even for rejected papers.
6. The BLEU score metric is primarily criticized because:
Correct. BLEU measures n-gram overlap — a surface form match — not semantic accuracy or factual correctness. A fluent but factually wrong output can score high on BLEU.
Not quite. BLEU's core limitation is that n-gram overlap doesn't capture meaning. A model can score high on BLEU while being factually inaccurate.
7. An ML engineer's primary focus is best described as:
Correct. ML engineers bridge the gap between research and production — building the systems that train, deploy, and sustain models at scale.
ML engineers focus on production systems — deploying and sustaining models reliably — rather than research, visualization, or finance.
8. LinkedIn's ML infrastructure processes over how many feature values per day?
Correct. LinkedIn's real-time ML platform processes over 3 trillion feature values per day — an illustration of the scale at which AI infrastructure engineers operate in large consumer platforms.
LinkedIn's platform processes 3 trillion feature values per day — a number that illustrates why dedicated AI infrastructure engineering exists as a specialized discipline.
9. In the context of AI policy careers, what distinguishes a role in an in-house ethics function from a role at an independent think tank?
Correct. The fundamental distinction is in work products and audiences: in-house roles produce internal governance deliverables affecting specific products, while think tanks publish public-facing research designed to influence legislative and regulatory processes.
Incorrect. The key distinction is in output type and audience — internal governance products versus public research designed to influence policy — not primarily in technical requirements or compensation.
10. Which body was created in February 2024 to oversee enforcement of the EU AI Act?
Correct. The EU AI Office was created in February 2024 to coordinate enforcement of the EU AI Act, develop technical standards for GPAI models, and liaise with member state authorities.
Incorrect. The EU AI Office was established in February 2024 — the enforcement and coordination body for the EU AI Act.
11. Jason Allen's 2022 Colorado State Fair win with an AI-generated image was notable primarily because it demonstrated what about AI art direction?
Correct. Allen's weeks of iterative prompting and selection process was analogous to the work of an art director — making creative decisions through a new kind of interface.
The significance was that Allen's careful prompting and curation process demonstrated AI direction as a legitimate form of creative authorship.
12. In professional image generation workflows, a "negative prompt" is used to:
Correct. Negative prompting is a core professional technique — instructing the model what not to include, which shapes results as importantly as positive guidance.
A negative prompt tells the model what to exclude — it is instruction about unwanted elements, not an inversion of the main prompt.
13. NeurIPS requires authors to include which section, added as a requirement from 2021 onward?
Correct. NeurIPS added a Broader Impacts requirement to encourage authors to consider societal implications and honestly address limitations of their work.
Not quite. NeurIPS introduced the Broader Impacts section requirement — addressing potential societal effects and limitations — as a mandatory component of submissions.
14. What did Airbnb's data science team discover that directly led to a major product initiative?
Correct. Photo quality analysis led directly to Airbnb's professional photography program — a canonical example of data science shaping product decisions.
Airbnb's data team found that professional photos were the key driver of booking success, which led to the Airbnb photography initiative.
15. The concept of "content stratification" in AI-augmented writing describes:
Correct. Content stratification describes how AI reshapes the writing market — compressing demand at the high-volume/low-differentiation tier while increasing demand for human judgment, original reporting, and AI direction at the top tier.
Content stratification describes the market split between AI-susceptible (high volume, low differentiation) and human-essential (high judgment, original perspective) writing tiers.
16. The CNET AI article controversy demonstrated what core limitation of AI writing tools?
Correct. CNET's articles were stylistically plausible — the problem was factual errors in financial content that the AI could not catch, demonstrating the verification gap in AI writing at scale.
The CNET lesson was specifically about plausible-sounding content with undetected factual errors — the generation/verification gap in AI writing.
17. The "Attention Is All You Need" paper (Vaswani et al., 2017) introduced:
Correct. The 2017 paper introduced the Transformer — a sequence-to-sequence architecture based entirely on attention mechanisms, abandoning recurrent and convolutional layers.
Not quite. "Attention Is All You Need" introduced the Transformer architecture — based purely on attention, with no recurrence or convolutions.
18. A paper that shows only cherry-picked qualitative examples without reporting systematic evaluation statistics:
Correct. A figure with four impressive outputs is always selected from many outputs. Without systematic evaluation, you have no information about failure rate, which is essential for assessing the method.
Not quite. Cherry-picked examples are uninformative about typical performance and failure rate — they are always selected from the best outputs of many.
19. The major record labels' 2024 lawsuits against Suno and Udio alleged primarily:
Correct. Universal, Sony, and Warner alleged that Suno and Udio had trained on their protected music catalogs without permission or compensation — the same training-data dispute seen in image generation.
The lawsuits alleged training-data infringement — using protected music catalogs without licensing to train generation systems.
20. What three core provisions did the WGA's 2023 strike settlement establish regarding AI?
Correct. The three WGA provisions — authorship prohibition, disclosure requirement, and minimum protection — established a template that other creative labor agreements subsequently referenced.
The WGA's three provisions: AI cannot author literary material, studios disclose AI material provided to writers, and writer minimums apply regardless of AI use.