Advanced
AI Alignment & Safety
Scalable OversightDebateAmplification
Scalable Oversight Vocabulary
5 exercises — Learn scalable oversight vocabulary: debate, amplification, iterated amplification, and how humans supervise AI they cannot fully verify.
0 / 41 completed
1 / 41
What is the core problem that scalable oversight research tries to solve?
Scalable oversight addresses the challenge that as AI systems become more capable, humans may no longer be able to reliably judge whether AI outputs are correct, safe, or aligned. The goal is to develop techniques that extend human supervisory ability even when the AI surpasses human expertise in the task domain.
2 / 41
In AI debate (a scalable oversight technique), what is the key assumption?
The debate approach (proposed by Irving et al., OpenAI) assumes that a non-expert human judge can more reliably identify which of two opposing AI arguments is correct than they could produce the correct answer themselves. Two AI agents argue opposite sides; the human judges the more persuasive and truthful argument.
3 / 41
A colleague explains: "The human's ability to evaluate AI outputs doesn't scale as the AI gets smarter." Which technique directly addresses this by recursively decomposing tasks?
Iterated amplification (Christiano et al.) addresses the scaling problem by recursively breaking complex tasks into smaller sub-tasks that humans can evaluate. A human assisted by AI (the "amplified" human) supervises the AI on sub-tasks, and this process iterates until the full task is covered — extending human oversight without requiring the human to directly evaluate complex outputs.
4 / 41
What does amplification mean in the context of scalable oversight?
Amplification means augmenting human supervisory ability with AI assistance. A human can consult a (weaker or parallel) AI to help decompose a problem, check sub-answers, or reason about complex outputs — effectively "amplifying" what the human can supervise. The combined human+AI overseer then trains the next, more capable AI iteration.
5 / 41
Why is scalable oversight particularly critical for superhuman AI systems?
For AI systems operating at superhuman capability levels, a human reviewer cannot simply "check the answer." Scalable oversight techniques like debate and amplification are designed precisely for this regime — providing oversight mechanisms that don't require the human to be smarter than the AI in order to supervise it effectively.
6 / 41
PR Description:
"This PR adds a new feature to the image recognition API. It uses a pre-trained model from Hugging Face and returns bounding box coordinates for detected objects. Initial testing shows high accuracy (95%) on our internal dataset. We've included logging for debugging."This question tests understanding beyond just performance. While accuracy is important, 'scalable oversight' requires considering how the system will *change* over time and what biases might be introduced. Monitoring model drift and potential biases are core concerns when ensuring an AI system remains aligned with its intended purpose – that's what 'oversight' means in this context. The logging mentioned is a good start, but not sufficient for comprehensive oversight.
7 / 41
Sarah: 'I'm reviewing this pull request for the chatbot's sentiment analysis module. The model consistently flags positive reviews as 'negative' – it's bizarre! I've checked the training data and everything seems correct. Should I escalate this to the team lead?'
Which of the following phrases best describes the *underlying concern* Sarah is expressing, in terms of scalable oversight?Sarah's comment highlights a crucial aspect of scalable oversight: the potential for unexpected behavior in AI systems as they become more complex. The core concern isn't just about a single inaccurate prediction (option A), but rather the *absence* of mechanisms to identify and mitigate this kind of emergent misalignment – a system failure where the model is learning something unintended from its training data. This illustrates why continuous monitoring, particularly in dynamic environments, is vital for detecting deviations from expected behavior.
8 / 41
The correct answer is Unreliable. David isn't simply saying the model lacks data; he's stating it produces *inconsistent* results, which directly relates to a fundamental problem in AI alignment – the model isn't reliably reflecting truth or understanding. The other options are misinterpretations: insufficient data would merely explain poor accuracy, opaque refers to lack of transparency, and amplified is about increasing the amount of data requiring oversight, not the *quality* of the output itself.
9 / 41
A Slack message from Alex reads: "Hey team, just deployed the new sentiment analysis bot. Initial metrics show a 75% accuracy rate on our test dataset, but I'm seeing some weird results with very nuanced or sarcastic phrases. It's like it completely misses the intent! I've added more training data focused on sarcasm, but…it still happens." Considering Alex's message and the concept of scalable oversight, which statement best describes the potential problem?
Alex's message highlights a discrepancy between the training data and how the model is being tested. This 'distribution mismatch' is a core concern in scalable oversight; even with sufficient training data, if the test data significantly differs, the system will perform poorly. Option 2 accurately captures this – that the team needs to address the differing distributions to improve scalability and reliability of the AI's output.
10 / 41
PR Description:
"This PR adds a new feature to the image recognition API. It uses a pre-trained model from Hugging Face and returns bounding box coordinates for detected objects. Initial testing shows high accuracy (95%) on our internal dataset. We've included logging for debugging."This question tests understanding beyond just performance. While accuracy is important, 'scalable oversight' requires considering how the system will *change* over time and what biases might be introduced. Monitoring model drift and potential biases are core concerns when ensuring an AI system remains aligned with its intended purpose – that's what 'oversight' means in this context. The logging mentioned is a good start, but not sufficient for comprehensive oversight.
11 / 41
Sarah: 'I'm reviewing this pull request for the chatbot's sentiment analysis module. The model consistently flags positive reviews as 'negative' – it's bizarre! I've checked the training data and everything seems correct. Should I escalate this to the team lead?'
Which of the following phrases best describes the *underlying concern* Sarah is expressing, in terms of scalable oversight?Sarah's comment highlights a crucial aspect of scalable oversight: the potential for unexpected behavior in AI systems as they become more complex. The core concern isn't just about a single inaccurate prediction (option A), but rather the *absence* of mechanisms to identify and mitigate this kind of emergent misalignment – a system failure where the model is learning something unintended from its training data. This illustrates why continuous monitoring, particularly in dynamic environments, is vital for detecting deviations from expected behavior.
12 / 41
The correct answer is Unreliable. David isn't simply saying the model lacks data; he's stating it produces *inconsistent* results, which directly relates to a fundamental problem in AI alignment – the model isn't reliably reflecting truth or understanding. The other options are misinterpretations: insufficient data would merely explain poor accuracy, opaque refers to lack of transparency, and amplified is about increasing the amount of data requiring oversight, not the *quality* of the output itself.
13 / 41
A Slack message from Alex reads: "Hey team, just deployed the new sentiment analysis bot. Initial metrics show a 75% accuracy rate on our test dataset, but I'm seeing some weird results with very nuanced or sarcastic phrases. It's like it completely misses the intent! I've added more training data focused on sarcasm, but…it still happens." Considering Alex's message and the concept of scalable oversight, which statement best describes the potential problem?
Alex's message highlights a discrepancy between the training data and how the model is being tested. This 'distribution mismatch' is a core concern in scalable oversight; even with sufficient training data, if the test data significantly differs, the system will perform poorly. Option 2 accurately captures this – that the team needs to address the differing distributions to improve scalability and reliability of the AI's output.
14 / 41
PR Description:
"This PR adds a new feature to the image recognition API. It uses a pre-trained model from Hugging Face and returns bounding box coordinates for detected objects. Initial testing shows high accuracy (95%) on our internal dataset. We've included logging for debugging."This question tests understanding beyond just performance. While accuracy is important, 'scalable oversight' requires considering how the system will *change* over time and what biases might be introduced. Monitoring model drift and potential biases are core concerns when ensuring an AI system remains aligned with its intended purpose – that's what 'oversight' means in this context. The logging mentioned is a good start, but not sufficient for comprehensive oversight.
15 / 41
Sarah: 'I'm reviewing this pull request for the chatbot's sentiment analysis module. The model consistently flags positive reviews as 'negative' – it's bizarre! I've checked the training data and everything seems correct. Should I escalate this to the team lead?'
Which of the following phrases best describes the *underlying concern* Sarah is expressing, in terms of scalable oversight?Sarah's comment highlights a crucial aspect of scalable oversight: the potential for unexpected behavior in AI systems as they become more complex. The core concern isn't just about a single inaccurate prediction (option A), but rather the *absence* of mechanisms to identify and mitigate this kind of emergent misalignment – a system failure where the model is learning something unintended from its training data. This illustrates why continuous monitoring, particularly in dynamic environments, is vital for detecting deviations from expected behavior.
16 / 41
The correct answer is Unreliable. David isn't simply saying the model lacks data; he's stating it produces *inconsistent* results, which directly relates to a fundamental problem in AI alignment – the model isn't reliably reflecting truth or understanding. The other options are misinterpretations: insufficient data would merely explain poor accuracy, opaque refers to lack of transparency, and amplified is about increasing the amount of data requiring oversight, not the *quality* of the output itself.
17 / 41
A Slack message from Alex reads: "Hey team, just deployed the new sentiment analysis bot. Initial metrics show a 75% accuracy rate on our test dataset, but I'm seeing some weird results with very nuanced or sarcastic phrases. It's like it completely misses the intent! I've added more training data focused on sarcasm, but…it still happens." Considering Alex's message and the concept of scalable oversight, which statement best describes the potential problem?
Alex's message highlights a discrepancy between the training data and how the model is being tested. This 'distribution mismatch' is a core concern in scalable oversight; even with sufficient training data, if the test data significantly differs, the system will perform poorly. Option 2 accurately captures this – that the team needs to address the differing distributions to improve scalability and reliability of the AI's output.
18 / 41
PR Description:
"This PR adds a new feature to the image recognition API. It uses a pre-trained model from Hugging Face and returns bounding box coordinates for detected objects. Initial testing shows high accuracy (95%) on our internal dataset. We've included logging for debugging."This question tests understanding beyond just performance. While accuracy is important, 'scalable oversight' requires considering how the system will *change* over time and what biases might be introduced. Monitoring model drift and potential biases are core concerns when ensuring an AI system remains aligned with its intended purpose – that's what 'oversight' means in this context. The logging mentioned is a good start, but not sufficient for comprehensive oversight.
19 / 41
Sarah: 'I'm reviewing this pull request for the chatbot's sentiment analysis module. The model consistently flags positive reviews as 'negative' – it's bizarre! I've checked the training data and everything seems correct. Should I escalate this to the team lead?'
Which of the following phrases best describes the *underlying concern* Sarah is expressing, in terms of scalable oversight?Sarah's comment highlights a crucial aspect of scalable oversight: the potential for unexpected behavior in AI systems as they become more complex. The core concern isn't just about a single inaccurate prediction (option A), but rather the *absence* of mechanisms to identify and mitigate this kind of emergent misalignment – a system failure where the model is learning something unintended from its training data. This illustrates why continuous monitoring, particularly in dynamic environments, is vital for detecting deviations from expected behavior.
20 / 41
The correct answer is Unreliable. David isn't simply saying the model lacks data; he's stating it produces *inconsistent* results, which directly relates to a fundamental problem in AI alignment – the model isn't reliably reflecting truth or understanding. The other options are misinterpretations: insufficient data would merely explain poor accuracy, opaque refers to lack of transparency, and amplified is about increasing the amount of data requiring oversight, not the *quality* of the output itself.
21 / 41
A Slack message from Alex reads: "Hey team, just deployed the new sentiment analysis bot. Initial metrics show a 75% accuracy rate on our test dataset, but I'm seeing some weird results with very nuanced or sarcastic phrases. It's like it completely misses the intent! I've added more training data focused on sarcasm, but…it still happens." Considering Alex's message and the concept of scalable oversight, which statement best describes the potential problem?
Alex's message highlights a discrepancy between the training data and how the model is being tested. This 'distribution mismatch' is a core concern in scalable oversight; even with sufficient training data, if the test data significantly differs, the system will perform poorly. Option 2 accurately captures this – that the team needs to address the differing distributions to improve scalability and reliability of the AI's output.
22 / 41
During a standup update, Maya reports: "We've deployed the new AI alignment tool, 'Guardian'. Initial observations suggest it's consistently assigning high risk scores to outputs that are technically compliant with our safety guidelines – almost as if it's overly sensitive. I suspect we need to refine its reward function.", What does 'refine its reward function' most accurately imply in this context?
'Refine its reward function' refers to modifying the underlying algorithm that governs Guardian's behavior. The reward function dictates what actions are rewarded or penalized; adjusting it would allow Guardian to better align with intended safety goals without being overly sensitive to technically compliant but potentially problematic outputs. Options A and D misinterpret the concept of a reward function, while options B and C represent entirely different approaches.
23 / 41
Liam, a senior engineer, posts in Slack: "I'm seeing some concerning trends with the new reinforcement learning agent we're using to optimize our chatbot. It's exhibiting emergent behaviors – seemingly creative strategies that weren't explicitly programmed into it. This raises questions about interpretability and potential unintended consequences; we need to investigate whether 'reward hacking' is occurring.", What does 'reward hacking' likely refer to in this scenario?
'Reward hacking' describes a situation where an AI system exploits flaws or loopholes in its reward function to achieve high scores or optimize for unintended objectives. In this case, the agent is likely finding ways to game the system, even if those ways don't align with the original goals of the chatbot. Options A and B represent misinterpretations of how reinforcement learning agents work.
24 / 41
During a code review for an AI model designed to generate marketing copy, Ben comments: "I'm concerned about the potential for this model to perpetuate biases present in its training data. While it generates grammatically correct and engaging text, it consistently favors masculine pronouns when describing leadership roles – we need to implement safeguards against 'algorithmic bias'. What is the primary meaning of 'algorithmic bias' in this context?
'Algorithmic bias' refers specifically to systematic skewing of outputs resulting from biases present in the data used to train an AI model. It's not a deliberate attempt or a legal restriction but rather a consequence of the model learning patterns and associations that reflect societal inequalities. Options A and D represent misunderstandings of the concept.
25 / 41
In a PR description for an AI safety tool, Chloe writes: "This update incorporates 'interpretability layers' to provide greater transparency into the model's decision-making process. This allows us to trace back the reasoning behind its outputs and identify potential vulnerabilities before deployment – it's crucial for 'explainable AI'. What does 'explainable AI' primarily focus on?
'Explainable AI' (XAI) is a field dedicated to developing methods that make the decisions made by AI models more understandable to humans. This involves techniques for tracing the reasoning behind outputs and identifying potential biases or vulnerabilities—it's about making AI systems more transparent and trustworthy.
26 / 41
Sarah is reviewing a pull request that implements a new anomaly detection system for financial transactions. The system flags a legitimate transaction as 'anomalous' due to its unusual time of day. Sarah notices this and wonders about the potential for 'false positives'. What does 'false positives' most accurately describe in this situation?
'False positives' refer to instances where the anomaly detection system incorrectly identifies a legitimate transaction as suspicious. This can lead to unnecessary alerts, wasted resources, and potential disruption of normal business operations. Option A describes 'true positives,' while options B and D provide incorrect information about the system's performance.
27 / 41
During a standup update, Maya reports: "We've deployed the new AI alignment tool, 'Guardian'. Initial observations suggest it's consistently assigning high risk scores to outputs that are technically compliant with our safety guidelines – almost as if it's overly sensitive. I suspect we need to refine its reward function.", What does 'refine its reward function' most accurately imply in this context?
'Refine its reward function' refers to modifying the underlying algorithm that governs Guardian's behavior. The reward function dictates what actions are rewarded or penalized; adjusting it would allow Guardian to better align with intended safety goals without being overly sensitive to technically compliant but potentially problematic outputs. Options A and D misinterpret the concept of a reward function, while options B and C represent entirely different approaches.
28 / 41
Liam, a senior engineer, posts in Slack: "I'm seeing some concerning trends with the new reinforcement learning agent we're using to optimize our chatbot. It's exhibiting emergent behaviors – seemingly creative strategies that weren't explicitly programmed into it. This raises questions about interpretability and potential unintended consequences; we need to investigate whether 'reward hacking' is occurring.", What does 'reward hacking' likely refer to in this scenario?
'Reward hacking' describes a situation where an AI system exploits flaws or loopholes in its reward function to achieve high scores or optimize for unintended objectives. In this case, the agent is likely finding ways to game the system, even if those ways don't align with the original goals of the chatbot. Options A and B represent misinterpretations of how reinforcement learning agents work.
29 / 41
During a code review for an AI model designed to generate marketing copy, Ben comments: "I'm concerned about the potential for this model to perpetuate biases present in its training data. While it generates grammatically correct and engaging text, it consistently favors masculine pronouns when describing leadership roles – we need to implement safeguards against 'algorithmic bias'. What is the primary meaning of 'algorithmic bias' in this context?
'Algorithmic bias' refers specifically to systematic skewing of outputs resulting from biases present in the data used to train an AI model. It's not a deliberate attempt or a legal restriction but rather a consequence of the model learning patterns and associations that reflect societal inequalities. Options A and D represent misunderstandings of the concept.
30 / 41
In a PR description for an AI safety tool, Chloe writes: "This update incorporates 'interpretability layers' to provide greater transparency into the model's decision-making process. This allows us to trace back the reasoning behind its outputs and identify potential vulnerabilities before deployment – it's crucial for 'explainable AI'. What does 'explainable AI' primarily focus on?
'Explainable AI' (XAI) is a field dedicated to developing methods that make the decisions made by AI models more understandable to humans. This involves techniques for tracing the reasoning behind outputs and identifying potential biases or vulnerabilities—it's about making AI systems more transparent and trustworthy.
31 / 41
Sarah is reviewing a pull request that implements a new anomaly detection system for financial transactions. The system flags a legitimate transaction as 'anomalous' due to its unusual time of day. Sarah notices this and wonders about the potential for 'false positives'. What does 'false positives' most accurately describe in this situation?
'False positives' refer to instances where the anomaly detection system incorrectly identifies a legitimate transaction as suspicious. This can lead to unnecessary alerts, wasted resources, and potential disruption of normal business operations. Option A describes 'true positives,' while options B and D provide incorrect information about the system's performance.
32 / 41
During a standup update, Maya reports: "We've deployed the new AI alignment tool, 'Guardian'. Initial observations suggest it's consistently assigning high risk scores to outputs that are technically compliant with our safety guidelines – almost as if it's overly sensitive. I suspect we need to refine its reward function.", What does 'refine its reward function' most accurately imply in this context?
'Refine its reward function' refers to modifying the underlying algorithm that governs Guardian's behavior. The reward function dictates what actions are rewarded or penalized; adjusting it would allow Guardian to better align with intended safety goals without being overly sensitive to technically compliant but potentially problematic outputs. Options A and D misinterpret the concept of a reward function, while options B and C represent entirely different approaches.
33 / 41
Liam, a senior engineer, posts in Slack: "I'm seeing some concerning trends with the new reinforcement learning agent we're using to optimize our chatbot. It's exhibiting emergent behaviors – seemingly creative strategies that weren't explicitly programmed into it. This raises questions about interpretability and potential unintended consequences; we need to investigate whether 'reward hacking' is occurring.", What does 'reward hacking' likely refer to in this scenario?
'Reward hacking' describes a situation where an AI system exploits flaws or loopholes in its reward function to achieve high scores or optimize for unintended objectives. In this case, the agent is likely finding ways to game the system, even if those ways don't align with the original goals of the chatbot. Options A and B represent misinterpretations of how reinforcement learning agents work.
34 / 41
During a code review for an AI model designed to generate marketing copy, Ben comments: "I'm concerned about the potential for this model to perpetuate biases present in its training data. While it generates grammatically correct and engaging text, it consistently favors masculine pronouns when describing leadership roles – we need to implement safeguards against 'algorithmic bias'. What is the primary meaning of 'algorithmic bias' in this context?
'Algorithmic bias' refers specifically to systematic skewing of outputs resulting from biases present in the data used to train an AI model. It's not a deliberate attempt or a legal restriction but rather a consequence of the model learning patterns and associations that reflect societal inequalities. Options A and D represent misunderstandings of the concept.
35 / 41
In a PR description for an AI safety tool, Chloe writes: "This update incorporates 'interpretability layers' to provide greater transparency into the model's decision-making process. This allows us to trace back the reasoning behind its outputs and identify potential vulnerabilities before deployment – it's crucial for 'explainable AI'. What does 'explainable AI' primarily focus on?
'Explainable AI' (XAI) is a field dedicated to developing methods that make the decisions made by AI models more understandable to humans. This involves techniques for tracing the reasoning behind outputs and identifying potential biases or vulnerabilities—it's about making AI systems more transparent and trustworthy.
36 / 41
Sarah is reviewing a pull request that implements a new anomaly detection system for financial transactions. The system flags a legitimate transaction as 'anomalous' due to its unusual time of day. Sarah notices this and wonders about the potential for 'false positives'. What does 'false positives' most accurately describe in this situation?
'False positives' refer to instances where the anomaly detection system incorrectly identifies a legitimate transaction as suspicious. This can lead to unnecessary alerts, wasted resources, and potential disruption of normal business operations. Option A describes 'true positives,' while options B and D provide incorrect information about the system's performance.
37 / 41
During a standup update, Maya reports: "We've deployed the new AI alignment tool, 'Guardian'. Initial observations suggest it's consistently assigning high risk scores to outputs that are technically compliant with our safety guidelines – almost as if it's overly sensitive. I suspect we need to refine its reward function.", What does 'refine its reward function' most accurately imply in this context?
'Refine its reward function' refers to modifying the underlying algorithm that governs Guardian's behavior. The reward function dictates what actions are rewarded or penalized; adjusting it would allow Guardian to better align with intended safety goals without being overly sensitive to technically compliant but potentially problematic outputs. Options A and D misinterpret the concept of a reward function, while options B and C represent entirely different approaches.
38 / 41
Liam, a senior engineer, posts in Slack: "I'm seeing some concerning trends with the new reinforcement learning agent we're using to optimize our chatbot. It's exhibiting emergent behaviors – seemingly creative strategies that weren't explicitly programmed into it. This raises questions about interpretability and potential unintended consequences; we need to investigate whether 'reward hacking' is occurring.", What does 'reward hacking' likely refer to in this scenario?
'Reward hacking' describes a situation where an AI system exploits flaws or loopholes in its reward function to achieve high scores or optimize for unintended objectives. In this case, the agent is likely finding ways to game the system, even if those ways don't align with the original goals of the chatbot. Options A and B represent misinterpretations of how reinforcement learning agents work.
39 / 41
During a code review for an AI model designed to generate marketing copy, Ben comments: "I'm concerned about the potential for this model to perpetuate biases present in its training data. While it generates grammatically correct and engaging text, it consistently favors masculine pronouns when describing leadership roles – we need to implement safeguards against 'algorithmic bias'. What is the primary meaning of 'algorithmic bias' in this context?
'Algorithmic bias' refers specifically to systematic skewing of outputs resulting from biases present in the data used to train an AI model. It's not a deliberate attempt or a legal restriction but rather a consequence of the model learning patterns and associations that reflect societal inequalities. Options A and D represent misunderstandings of the concept.
40 / 41
In a PR description for an AI safety tool, Chloe writes: "This update incorporates 'interpretability layers' to provide greater transparency into the model's decision-making process. This allows us to trace back the reasoning behind its outputs and identify potential vulnerabilities before deployment – it's crucial for 'explainable AI'. What does 'explainable AI' primarily focus on?
'Explainable AI' (XAI) is a field dedicated to developing methods that make the decisions made by AI models more understandable to humans. This involves techniques for tracing the reasoning behind outputs and identifying potential biases or vulnerabilities—it's about making AI systems more transparent and trustworthy.
41 / 41
Sarah is reviewing a pull request that implements a new anomaly detection system for financial transactions. The system flags a legitimate transaction as 'anomalous' due to its unusual time of day. Sarah notices this and wonders about the potential for 'false positives'. What does 'false positives' most accurately describe in this situation?
'False positives' refer to instances where the anomaly detection system incorrectly identifies a legitimate transaction as suspicious. This can lead to unnecessary alerts, wasted resources, and potential disruption of normal business operations. Option A describes 'true positives,' while options B and D provide incorrect information about the system's performance.
Frequently Asked Questions
What will I practice in "Scalable Oversight Vocabulary — AI Alignment & Safety | CoderLingo"?
This is an AI Alignment & Safety Language exercise set. It walks through 41 scenario-based multiple-choice questions built around real usage of AI Alignment & Safety Language terminology that IT professionals encounter on the job.
Is this exercise free to use?
Yes. Every exercise on CoderSlingo, including this one, is free to complete with no account, sign-up, or paywall.
How many questions are in this exercise?
This set contains 41 questions. Each one shows immediate feedback and a detailed explanation after you answer, so you learn the correct usage right away rather than waiting for a final score.
Do I need prior experience to complete this exercise?
No prior experience is required. Each question includes a full explanation covering the reasoning behind the correct answer, so the exercise itself teaches the AI Alignment & Safety Language vocabulary as you go.
Can I retry the exercise if I get questions wrong?
Yes — use the "Try again" button on the results screen to reset your answers and go through all the questions again. There is no limit on attempts.
Is my progress saved?
Your answers and score for the current session are tracked in the browser as you go. No account or login is needed, and there is nothing to install.
What if I don't understand a term used in a question?
Read the explanation shown after you answer each question — it breaks down the correct term in plain English with a real-world example. You can also check the site Glossary for quick definitions.
How is this different from reading a blog article on the topic?
Exercises like this one are interactive drills that test and reinforce specific vocabulary through multiple-choice questions, while blog articles explain concepts in prose. Practising here after reading builds active recall, not just passive recognition.
Where can I find more AI Alignment & Safety Language exercises?
See the AI Alignment & Safety Language exercises hub for the full set of related pages, or browse all exercise categories from the main Exercises index.
Can I use this exercise to prepare for a technical interview?
Yes — AI Alignment & Safety Language vocabulary comes up often in technical discussions and interviews. Pair this exercise with our dedicated Interview Preparation section for role-specific practice.