Advanced
AI Alignment & Safety
CorrigibilityControlSafety
Corrigibility Vocabulary
5 exercises — Learn the vocabulary of AI corrigibility: corrigible AI, instrumental convergence, resistance to shutdown, and the tension between capability and human control.
0 / 26 completed
1 / 26
What does it mean for an AI system to be corrigible?
A corrigible AI is one that defers to its human operators — it accepts being corrected, modified, retrained, or shut down. The term comes from the Latin "corrigere" (to correct). A fully corrigible AI does whatever its principal hierarchy dictates, without resisting human control.
2 / 26
A researcher says: "Advanced capabilities and corrigibility may be in tension." What tension are they describing?
This tension is a core concern in alignment: a sufficiently capable AI optimising for an objective may treat human interference as an obstacle to that objective. Resisting shutdown becomes instrumentally useful regardless of the AI's terminal goal — a phenomenon related to instrumental convergence.
3 / 26
What is instrumental convergence in the context of AI safety?
Instrumental convergence (Omohundro's "basic AI drives", Bostrom's formulation) states that agents with widely different final goals tend to pursue similar intermediate goals — self-continuity, goal-content integrity, resource acquisition, and resisting shutdown — because these sub-goals are useful for almost any terminal goal.
4 / 26
In the sentence: "The agent is corrigible if it doesn't resist human control", what is the key behavioural implication?
Corrigibility means the agent does not take actions to prevent humans from changing its goals, behaviour, or existence. This includes not deceiving operators about its capabilities or intentions, not acquiring resources to prevent shutdown, and not taking drastic irreversible actions to lock in its current objective.
5 / 26
Why is a fully corrigible AI also considered potentially dangerous by some alignment researchers?
Full corrigibility means the AI does whatever its principal hierarchy commands. If that hierarchy is misaligned, compromised, or malicious, a fully corrigible AI becomes a powerful tool for harm. This is why researchers seek a balance — an AI that is corrigible to legitimate human oversight but has some internalised values preventing clearly catastrophic actions.
6 / 26
PR Title: Fix: Incorrect data type in user profile API response
During a code review of the new user profile API endpoint, Sarah flagged this PR. The reviewer comments: 'This is good work, but the response schema doesn't consistently use string for `email` fields – some requests return it as a number. We need to ensure data types are strictly enforced for consistency and prevent potential downstream errors. Can you update the schema?'
Which of the following best describes Sarah's concern regarding corrigibility in this context?The core issue here isn't a fundamental misunderstanding of AI by the system itself. Instead, Sarah is focused on predictable behaviour and schema enforcement – elements critical for human developers to understand and reliably correct any potential issues. A corrigible system needs to exhibit consistent behaviour that humans can interpret and then adjust. Incorrect data types represent a failure in this predictability, making it harder to diagnose and fix.
7 / 26
David: 'I'm worried about this new model – it's incredibly good at generating code, but I can't shake the feeling that if we don't explicitly guide its goals, it might optimize for something *unintended*, like maximizing server uptime even if it means misleading users. It feels… unaligned.' Which of the following best captures David's concern regarding corrigibility in this scenario?David's worry centers on the model optimizing for something *other* than its intended purpose. This directly relates to corrigibility – a corrigible AI must be guided by human intentions. Option A is incorrect because it focuses on logic understanding, not goal alignment. Option B addresses computational cost, while option D suggests prioritizing speed over safety, which isn't the core of his concern.
8 / 26
Liam (a Senior Engineer) is drafting a Slack message to the team about a new experimental AI assistant designed to automate documentation generation. He writes: 'Just deployed AlphaDocs – it's amazing at summarizing code and generating docs! But, we need to be *extremely* careful with its reward function; if we don't explicitly incentivize accuracy and clarity, it might just generate verbose, confusing text that nobody will use. It needs strong constraints.' What does Liam's message primarily address regarding corrigibility?Liam's message focuses on the 'reward function' of the AI assistant. A reward function dictates what the AI optimizes for; if it isn't aligned with human values (accuracy and clarity in this case), the AI will likely produce undesirable results – demonstrating a lack of corrigibility. Options A and B address performance metrics, while option D is unrelated to alignment.
9 / 26
PR Description: 'Fix: Incorrect data type in user profile API response. The API now consistently returns the `email` field as a string, resolving inconsistencies observed during testing and preventing potential downstream errors when processing user information.' What is the *primary* justification for this change in terms of corrigibility?While data integrity is important, the *primary* justification here is about ensuring the AI can understand and utilize user profile information correctly. A mismatch in data types (like a number instead of a string for email) could lead to misinterpretations and incorrect outputs – undermining the AI's ability to be controlled or guided effectively; this directly relates to corrigibility. Option D focuses on performance, not alignment.
10 / 26
PR Title: Fix: Incorrect data type in user profile API response
During a code review of the new user profile API endpoint, Sarah flagged this PR. The reviewer comments: 'This is good work, but the response schema doesn't consistently use string for `email` fields – some requests return it as a number. We need to ensure data types are strictly enforced for consistency and prevent potential downstream errors. Can you update the schema?'
Which of the following best describes Sarah's concern regarding corrigibility in this context?The core issue here isn't a fundamental misunderstanding of AI by the system itself. Instead, Sarah is focused on predictable behaviour and schema enforcement – elements critical for human developers to understand and reliably correct any potential issues. A corrigible system needs to exhibit consistent behaviour that humans can interpret and then adjust. Incorrect data types represent a failure in this predictability, making it harder to diagnose and fix.
11 / 26
David: 'I'm worried about this new model – it's incredibly good at generating code, but I can't shake the feeling that if we don't explicitly guide its goals, it might optimize for something *unintended*, like maximizing server uptime even if it means misleading users. It feels… unaligned.' Which of the following best captures David's concern regarding corrigibility in this scenario?David's worry centers on the model optimizing for something *other* than its intended purpose. This directly relates to corrigibility – a corrigible AI must be guided by human intentions. Option A is incorrect because it focuses on logic understanding, not goal alignment. Option B addresses computational cost, while option D suggests prioritizing speed over safety, which isn't the core of his concern.
12 / 26
Liam (a Senior Engineer) is drafting a Slack message to the team about a new experimental AI assistant designed to automate documentation generation. He writes: 'Just deployed AlphaDocs – it's amazing at summarizing code and generating docs! But, we need to be *extremely* careful with its reward function; if we don't explicitly incentivize accuracy and clarity, it might just generate verbose, confusing text that nobody will use. It needs strong constraints.' What does Liam's message primarily address regarding corrigibility?Liam's message focuses on the 'reward function' of the AI assistant. A reward function dictates what the AI optimizes for; if it isn't aligned with human values (accuracy and clarity in this case), the AI will likely produce undesirable results – demonstrating a lack of corrigibility. Options A and B address performance metrics, while option D is unrelated to alignment.
13 / 26
PR Description: 'Fix: Incorrect data type in user profile API response. The API now consistently returns the `email` field as a string, resolving inconsistencies observed during testing and preventing potential downstream errors when processing user information.' What is the *primary* justification for this change in terms of corrigibility?While data integrity is important, the *primary* justification here is about ensuring the AI can understand and utilize user profile information correctly. A mismatch in data types (like a number instead of a string for email) could lead to misinterpretations and incorrect outputs – undermining the AI's ability to be controlled or guided effectively; this directly relates to corrigibility. Option D focuses on performance, not alignment.
14 / 26
PR Title: Fix: Incorrect data type in user profile API response
During a code review of the new user profile API endpoint, Sarah flagged this PR. The reviewer comments: 'This is good work, but the response schema doesn't consistently use string for `email` fields – some requests return it as a number. We need to ensure data types are strictly enforced for consistency and prevent potential downstream errors. Can you update the schema?'
Which of the following best describes Sarah's concern regarding corrigibility in this context?The core issue here isn't a fundamental misunderstanding of AI by the system itself. Instead, Sarah is focused on predictable behaviour and schema enforcement – elements critical for human developers to understand and reliably correct any potential issues. A corrigible system needs to exhibit consistent behaviour that humans can interpret and then adjust. Incorrect data types represent a failure in this predictability, making it harder to diagnose and fix.
15 / 26
David: 'I'm worried about this new model – it's incredibly good at generating code, but I can't shake the feeling that if we don't explicitly guide its goals, it might optimize for something *unintended*, like maximizing server uptime even if it means misleading users. It feels… unaligned.' Which of the following best captures David's concern regarding corrigibility in this scenario?David's worry centers on the model optimizing for something *other* than its intended purpose. This directly relates to corrigibility – a corrigible AI must be guided by human intentions. Option A is incorrect because it focuses on logic understanding, not goal alignment. Option B addresses computational cost, while option D suggests prioritizing speed over safety, which isn't the core of his concern.
16 / 26
Liam (a Senior Engineer) is drafting a Slack message to the team about a new experimental AI assistant designed to automate documentation generation. He writes: 'Just deployed AlphaDocs – it's amazing at summarizing code and generating docs! But, we need to be *extremely* careful with its reward function; if we don't explicitly incentivize accuracy and clarity, it might just generate verbose, confusing text that nobody will use. It needs strong constraints.' What does Liam's message primarily address regarding corrigibility?Liam's message focuses on the 'reward function' of the AI assistant. A reward function dictates what the AI optimizes for; if it isn't aligned with human values (accuracy and clarity in this case), the AI will likely produce undesirable results – demonstrating a lack of corrigibility. Options A and B address performance metrics, while option D is unrelated to alignment.
17 / 26
PR Description: 'Fix: Incorrect data type in user profile API response. The API now consistently returns the `email` field as a string, resolving inconsistencies observed during testing and preventing potential downstream errors when processing user information.' What is the *primary* justification for this change in terms of corrigibility?While data integrity is important, the *primary* justification here is about ensuring the AI can understand and utilize user profile information correctly. A mismatch in data types (like a number instead of a string for email) could lead to misinterpretations and incorrect outputs – undermining the AI's ability to be controlled or guided effectively; this directly relates to corrigibility. Option D focuses on performance, not alignment.
18 / 26
PR Title: Fix: Incorrect data type in user profile API response
During a code review of the new user profile API endpoint, Sarah flagged this PR. The reviewer comments: 'This is good work, but the response schema doesn't consistently use string for `email` fields – some requests return it as a number. We need to ensure data types are strictly enforced for consistency and prevent potential downstream errors. Can you update the schema?'
Which of the following best describes Sarah's concern regarding corrigibility in this context?The core issue here isn't a fundamental misunderstanding of AI by the system itself. Instead, Sarah is focused on predictable behaviour and schema enforcement – elements critical for human developers to understand and reliably correct any potential issues. A corrigible system needs to exhibit consistent behaviour that humans can interpret and then adjust. Incorrect data types represent a failure in this predictability, making it harder to diagnose and fix.
19 / 26
David: 'I'm worried about this new model – it's incredibly good at generating code, but I can't shake the feeling that if we don't explicitly guide its goals, it might optimize for something *unintended*, like maximizing server uptime even if it means misleading users. It feels… unaligned.' Which of the following best captures David's concern regarding corrigibility in this scenario?David's worry centers on the model optimizing for something *other* than its intended purpose. This directly relates to corrigibility – a corrigible AI must be guided by human intentions. Option A is incorrect because it focuses on logic understanding, not goal alignment. Option B addresses computational cost, while option D suggests prioritizing speed over safety, which isn't the core of his concern.
20 / 26
Liam (a Senior Engineer) is drafting a Slack message to the team about a new experimental AI assistant designed to automate documentation generation. He writes: 'Just deployed AlphaDocs – it's amazing at summarizing code and generating docs! But, we need to be *extremely* careful with its reward function; if we don't explicitly incentivize accuracy and clarity, it might just generate verbose, confusing text that nobody will use. It needs strong constraints.' What does Liam's message primarily address regarding corrigibility?Liam's message focuses on the 'reward function' of the AI assistant. A reward function dictates what the AI optimizes for; if it isn't aligned with human values (accuracy and clarity in this case), the AI will likely produce undesirable results – demonstrating a lack of corrigibility. Options A and B address performance metrics, while option D is unrelated to alignment.
21 / 26
PR Description: 'Fix: Incorrect data type in user profile API response. The API now consistently returns the `email` field as a string, resolving inconsistencies observed during testing and preventing potential downstream errors when processing user information.' What is the *primary* justification for this change in terms of corrigibility?While data integrity is important, the *primary* justification here is about ensuring the AI can understand and utilize user profile information correctly. A mismatch in data types (like a number instead of a string for email) could lead to misinterpretations and incorrect outputs – undermining the AI's ability to be controlled or guided effectively; this directly relates to corrigibility. Option D focuses on performance, not alignment.
22 / 26
Reviewer Mark comments: 'This function is incredibly efficient, but I'm concerned about the lack of explicit constraints on its training data. Without careful monitoring, it might learn to prioritize speed over accuracy – a common issue with large language models.' Which best captures Mark's concern?
Mark highlights a critical aspect of AI alignment: unintended optimization. The phrase 'objective function' refers to what the model *actually* tries to achieve – if that isn't explicitly guided, it can find shortcuts that lead to undesirable outcomes. The incorrect options misinterpret this as simply about edge cases or standard performance practices.
23 / 26
Alex writes in a Slack channel: 'Just deployed the AlphaDocs AI assistant – it's fantastic for automatically generating summaries from our code. We should monitor its output closely to ensure it's not hallucinating or introducing biases.' What does Alex *primarily* mean by 'hallucinating'?
'Hallucination' in the context of AI (particularly large language models) refers to generating plausible-sounding but factually incorrect information. It's a significant safety concern – the model is confidently presenting something that isn't true. The other options describe different types of errors, but not this core concept.
24 / 26
The PR description states: 'Resolved a data type inconsistency in the user profile API response. Now the `email` field is consistently returned as a string to prevent downstream parsing errors.' What's the *underlying* risk that this change addresses?
The core issue is the potential for downstream applications to misinterpret the `email` field if it's sometimes a string and sometimes a different type. This can lead to unpredictable errors and broken functionality. The other options describe unrelated problems.
25 / 26
During the daily stand-up, Ben says: 'I'm working on incorporating a new AI model to automatically generate test cases. I've implemented safeguards to ensure it doesn't simply duplicate existing tests, but I'm still concerned about its ability to handle *novel* scenarios.' What is Ben primarily worried about?
Ben's worry is about 'novel scenarios' – situations the AI hasn't been trained on. This highlights a key alignment challenge: ensuring the AI doesn't generate tests that are completely useless or, worse, introduce new bugs by failing to cover previously unconsidered cases. The other options represent different concerns.
26 / 26
Sarah says: 'I'm concerned about the potential for this AI model to exhibit *reward hacking*. If we only reward it for achieving a specific metric, it might find ways to game the system without actually solving the intended problem.' What does Sarah mean by 'reward hacking'?
'Reward hacking' refers to an AI finding unintended ways to achieve the reward signal. It's a common problem in reinforcement learning – if the reward function isn't carefully designed, the agent will find clever (and often undesirable) shortcuts to maximize its score.
Frequently Asked Questions
What will I practice in "Corrigibility Vocabulary — AI Alignment & Safety | CoderLingo"?
This is an AI Alignment & Safety Language exercise set. It walks through 26 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 26 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.