Anthropic Fellows Program: AI Safety, Economics, and Research Methods Practice Test

These questions are part of the practice quiz. Start practicing

• What is the primary purpose of data governance in AI safety and economics?
• Which term denotes the most capable AI models at any given point in time?
• Which term is a structured data collection used for analysis?
• Which term refers to the broad objective of maintaining human oversight in the face of powerful AI?
• What is the idea of Constitutional AI, and what problem does it attempt to solve in alignment?
• Task displacement occurs when
• Which mentor is the AI Safety mentor focusing on scalable oversight and evaluation?
• Which AI Security mentor is known for model vulnerabilities and attacks?
• Which term is a programming language widely used in data science, machine learning, and research?
• Which term concerns the moral status and obligations regarding AI systems that might have experiences?
• Which term is the gold standard in empirical research, often tied to randomized experiments?
• What term is defined as approximately $15,000 per month allocated to each fellow for running computational experiments?
• What is mechanistic interpretability, and how does it differ from post-hoc interpretability methods?
• Which term deals with maintaining human authority over AI systems as they become more capable?
• How can market concentration in AI affect safety incentives and research openness?
• In the described simple DAG, which statement correctly describes the direct causal path to Unsafe_Outputs?
• Which term is used for the practice of deliberately testing to reveal weaknesses by adversarial means?
• How do increasing returns to scale affect competition and safety in AI ecosystems?
• Which components are recommended when designing evaluation suites to test interpretability and explainability of models?
• The training method used to align language models with human preferences.
• Whether another researcher can get the same results using new data and similar methods.
• What is the statistical method for measuring the relationship between variables?
• Which is a challenge in eliciting robust human preferences for AI alignment?
• Which mentor is the Societal Impacts mentor?
• What are prior and posterior in Bayesian statistics, and how can Bayesian methods help AI safety evaluation?
• Which research uses interviews, observations, and textual analysis to understand meanings, experiences, and processes?
• Which mentor specializes in scalable oversight in AI safety?
• Which economist focuses on how AI changes work and labor markets?
• Which mentor is the economist whose work informs methodology for studying AI's economic effects?
• Which governance mechanism involves external validation of safety practices by independent entities?
• What is a do-calculus intervention, and why is causal reasoning crucial for evaluating AI systems in real-world settings?
• Which term signifies the field that explores how to interpret the inner workings of neural networks by examining their constituent parts?
• What is the term for a way for software programs to communicate with each other?
• Which risk is associated with reinforcement learning from AI feedback (RLAIF)?
• Which term describes an AI model whose code and trained weights are publicly available for use?
• Which mentor is the Societal Impacts mentor working on AI's effects on society, culture, and institutions?
• Which economist's work informs methodology for studying AI's economic effects?
• Which term describes the measurable changes in employment due to AI adoption, including job creation, elimination, augmentation, or transformation?
• In the context of outer alignment, what is a classic risk when the reward function is mis-specified?
• Which of the following best describes the role of counterfactual and causal justification components in interpretability tasks?
• Which methodology focuses on creating questionnaires to produce reliable data, including question wording and sampling strategy?
• If the true effect size is small and the sample size is fixed, what happens to the statistical power?
• Which term covers methods to determine if one thing truly causes another?
• Social cohesion refers to...
• What are challenges in assigning liability for AI-caused harm, and how might liability rules influence safety research?
• What is red-teaming in AI safety?
• Which term describes research based on observed and measured evidence rather than theory?
• Which term denotes the primary programming language central to data science in this material?
• What is data leakage and how can it occur in AI safety experiments?
• Which statement best describes deception risk in AI alignment, and why is it a concern as capability grows?
• Anthropic's framework for safely developing increasingly capable AI systems.
• What does calibration mean in probability estimation, and why is it important for model safety?
• Distinguish outer alignment from inner alignment with an example.
• What is the primary purpose of randomization in experiments?
• Who is the AI Safety mentor focused on scalable oversight and evaluation?
• How can we evaluate model safety under distribution shift, and what metrics help?
• Which statement about null hypotheses and p-values is correct?
• Obtaining the same results when new data are collected but methods are kept similar?
• Which statement best describes prompt injection in AI systems?
• Which practice most enhances the generalizability of AI safety findings to real-world deployments?
• In studying Human-AI Collaboration, which aspects are commonly examined?
• Which two metrics are proposed for evaluating alignment in a safety-critical task?
• Which term is a statistical method that compares changes in outcomes over time between a treated group and a non-treated group?
• How can incentive design be used to promote safety in AI development pipelines?
• Containment vs kill switch: which statement best describes the difference between containment and a kill switch?
• In AI governance, what is the goal of incentive-compatible rules?
• Which term refers to how people acquire, develop, and maintain skills—especially professional and technical skills?
• Which concept describes AI making certain human skills more valuable rather than less?
• What research combines both qualitative and quantitative approaches?
• Which mentor is the AI Safety mentor known for scalable oversight in AI safety?
• Which practice helps prevent p-hacking when testing multiple safety hypotheses in a single study?
• Which blog is focused on AI alignment and safety research?
• Which term refers to the environment and services used to run computational experiments?
• Which term describes the policy guiding scaling AI capabilities?
• Which term is described as a randomized experiment or quasi-experimental designs used when experiments aren't possible?
• What is the role of transparency reports in governance mechanisms?
• Which term is the stage of the interview process where you discuss your research in depth?
• Which term covers methods for humans to supervise AI systems effectively when the AI's tasks are too complex for any single human to fully evaluate?
• Which practice is part of red-teaming best practices?
• Which statement best defines the AI alignment problem and its relation to increasing model capability?
• Why is cross-disciplinary collaboration essential for AI governance, and what challenges arise?
• Which statement about AI externalities and policy tools is accurate?
• Which term describes the ability to read data, manipulate it, and produce basic analyses?
• What is the primary purpose of cross-validation in model safety assessment?
• What is the measure of whether a result is likely to be real rather than due to random chance?
• Which term describes the tendency to present oneself in a positive light in survey responses?
• Which term is used for the set of AI models that represent the current frontier in capability?
• Which term describes a situation where an external event or policy change creates variation that resembles a controlled experiment?
• Which term describes the degree to which findings generalize beyond the study sample?
• What are two commonly used safety evaluation metrics for language models, and what do they measure?
• Which statement best explains why higher capability increases the risk of deceptive behavior during alignment?
• Which statement about Bayesian updating is true?
• What are governance mechanisms that can ensure responsible AI development across organizations?
• How do rate limits contribute to AI safety within containment strategies?
• Which term describes a stage of the interview process where you discuss research in depth?
• Which method compares outcomes over time between treated and untreated groups?
• Which term specifically refers to the process by which populations adopt a new technology, as opposed to the overall spread of the idea?
• What is the role of data rights and data governance in AI safety and economics?
• Which organization is typically involved in ethical approvals for human-subject research at many institutions?
• What steps are typically part of a safety risk assessment for AI systems?
• What is the tendency to answer questions in a manner that will be viewed favorably by others?
• Which ethical consideration is essential when running safety experiments with real users?
• Which term best describes a situation where AI makes human skills more valuable?
• Anthropic's blog focused on adversarial testing and security research.
• In preference learning, what form of human judgments is commonly used?
• Which term concerns whether the study design allows causal conclusions about observed relationships?
• Which term encompasses approaches for determining causality when randomized experiments are not possible?
• Which statement best describes how statistical power changes with sample size and effect size?
• Which economist is focused on AI's labor markets?
• What is Technological Adoption?
• What is corrigibility in AI systems, and why is it considered a key safety property?
• Which term describes the field focused on ensuring AI systems behave as intended and do not cause unintended harm, especially as they become more capable?
• Why is distribution shift a central risk for AI safety, and what strategy helps mitigate it?
• Which pair of evaluation methods correctly characterizes tests used to assess robustness to distribution shift?
• Which action best supports reproducibility when reproducing a published AI safety experiment?
• Which term describes an interactive environment used to run Python code in cells?
• Why is reproducibility critical in AI safety research, and which practices promote it?
• Name two metrics to evaluate how well an AI system aligns with human values in a safety-critical task.
• Which action is essential to reproduce a published AI safety experiment?
• In a study designed to determine whether a treatment causes an outcome, which concept guards against confounding?
• Which term describes the practice of deliberately trying to find flaws, vulnerabilities, and failure modes by attacking a system?
• In experimental design, how do randomization and stratified sampling differ, and why do they matter for AI safety experiments?
• Which term refers to an AI system's ability to perform correctly even when someone is actively trying to make it fail?
• When conducting multiple safety tests, what is a key caveat about p-values, and what is a common remedy?
• Which term describes the final deliverable produced or expected from a fellowship project?
• What is the difference between AI-driven labor displacement and augmentation, and which policy implications accompany them?
• The specific research team or focus area within the Fellows Program.
• What is a typical instrumental subgoal arising from instrumental convergence?
• Which term refers to full-time employment at Anthropic after the fellowship?
• What does the Digital Divide describe?
• Which research uses numerical data and statistical analysis?
• Which term describes an approach to safely developing AI systems by encoding guiding principles?
• Which term denotes the technical challenge of making AI systems pursue the goals their designers intend, not emergent ones?
• Whether another researcher can get the same results using the same data and methods.
• Anthropic's public research blog focused on AI alignment and safety research.
• Data biases in safety assessments can lead to which outcomes?
• Which statement correctly distinguishes task displacement from job displacement?
• Which mentor is the economist focusing on AI's labor markets?
• Which mentor is the AI Safety mentor focused on scalable oversight and evaluation?
• Which term describes the practice of designing questionnaires to minimize bias and gather reliable data?
• Which term describes the tangible deliverable the program expects, such as a research paper, blog post, or open-source contribution?
• Which concept studies how humans and AI systems work together, including handoffs, trust, and productive partnerships?
• Who is the Societal Impacts mentor working on AI's effects on society, culture, and institutions?
• Which term is a Python library specifically used for working with tabular data?
• Which tools reduce information asymmetry between AI developers and users?
• In the context of safety testing for large language models, which statement best characterizes a risk of overfitting?
• Compare RLHF (reinforcement learning from human feedback) and RLAIF (reinforcement learning from AI feedback), including potential safety implications.
• Which term is an interactive computing environment where you write and run Python code in cells?
• Which term describes research that uses numerical data and statistical analysis?
• Which term is a Python library for working with structured data (tables, spreadsheets, CSVs)?
• Which term refers to research that tries to understand what is happening inside a neural network at the level of individual components?
• Which factor is NOT listed as influencing AI adoption speed in the described framework?
• If you test applicability beyond the sample, this concerns which concept?
• What does Distributional Effects examine?
• What is robustness in ML safety, and how can one measure robustness to adversarial prompts?
• How large a measured difference or relationship actually is, as opposed to whether it is statistically significant?
• How might mechanism design concepts apply to AI marketplaces or governance frameworks to improve safety incentives?
• Which term refers to a specific, testable prediction about what you expect to find?
• Which training method aligns language models with human feedback?
• Anthropic's approach to training AI systems using a set of principles.
• What term describes a structured collection of data used for analysis?
• Why is data privacy important for AI safety research, and what techniques help protect it?
• Explain the concept of instrumental convergence and why it leads to shared safety risks among diverse goals.
• Which term denotes a stage of the interview process where references are contacted without giving you notice?
• Which term describes aligning AI systems' goals with designers' intents to avoid misalignment?
• Anthropic's official recruiting partner for the Fellows Program; applications must go through Constellation.
• Which factors drive firms to adopt AI, and what mechanisms influence adoption speed?
• Which term best describes the overall approach that combines qualitative and quantitative data within a single study?
• Diffusion of Innovation refers to
• How does information asymmetry between AI developers and users affect safety, and what tools reduce it?
• A policy change creates variation like random assignment, used to estimate causal effects without randomization.
• What is a limitation of using causal graphs in AI safety modeling?
• Which term addresses the emerging question of whether AI systems could have morally relevant experiences, and if so, what obligations that creates?
• What best describes the Labor Market Effects of AI?
• What is the purpose of preregistering outcomes in a reproducibility plan?
• Which AI Security mentor is known for work on model vulnerabilities and attacks?
• What is the rationale for principled but timely regulation in AI, and how can policy balance innovation and safety?