Appendix B. Self-check answers and explanations

1-1. Objectives (achievement standards and lesson objectives), target audience (grade level and reading level), format (number of questions, layout, and answer key location), examples (examples of desired results), and constraints (what not to include). If objectives are omitted, content outside the unit becomes mixed in; if the target audience is omitted, the difficulty level will be inappropriate; and if format is omitted, the structure of the output varies each time. If examples are omitted, AI guesses the level and format; if constraints are omitted, difficult-to-verify figures or fabricated quotations are introduced.

1-2. Because AI services may have been trained on similar data or may rely on the same flawed sources, they can produce the same errors together. Agreement merely means that they are similar to each other; it does not mean that the information was verified against primary sources. Generating multiple times and comparing the outputs can be used to detect discrepancies, but it does not prove the facts.

1-3. Systematic bias, computational and statistical bias, and human cognitive bias. The one that can manifest in the reviewer themselves is human cognitive bias, an example of which is the tendency to be less skeptical of descriptions that align with one's own expectations.

1-4. Two of the following may be changed: sentence length and vocabulary, terminology explanations, types and number of examples, or the proportion of descriptions for figures and tables. Two of the following must not be changed: core concepts of the achievement standards, relationships between concepts, the level demanded by the assessment, or accuracy. For students taking the same assessment, core concepts must be maintained in particular.

1-5. High schools are schools under Article 25(3) of the Copyright Act and are permitted to use works for educational purposes; because they are schools at or below the high school level, they do not pay remuneration. Universities are also permitted to use works for educational purposes, but must pay remuneration. For private academies not on the list, Article 25 itself does not apply, so the matter cannot be resolved through remuneration; they must either satisfy the requirements for quotation under Article 28 or fair use under Article 35-5, or obtain permission for use. All three must specify the source.

1-6. Article 37(1) of the Copyright Act provides that "a person who exploits a work pursuant to this Subsection shall indicate the source," and its proviso enumerates the articles exempt from indicating the source. Because Article 25 is not included in that list, the obligation to indicate the source remains even for use for educational purposes.

1-7. The criterion is the extent of harm when incorrect. Facts that students will use as a basis for assessment or judgment and safety-related content are verified first, statements regarding real people and institutions second, and background explanations last. Sentences that could not be verified are either omitted, labeled "needs verification," or converted into an activity where students search for primary sources.

1-8. Multiple answers are possible. A good response must include all of the following: the achievement standard text (transcribed verbatim from curriculum documents) and lesson objectives, specific descriptions of the grade level and reading level, formats such as the number of questions and location of the answer key, one or two examples of the desired level, and constraints not to use unverified figures or statements by real people. If student names or scores are included, they must be corrected.

2-1. The phrase "early childhood education, primary education, and secondary education" leaves higher education and adult education outside the scope. The phrase "student assessment" leaves non-assessment activities, such as instructional material creation or learning content recommendations, outside the scope. In addition, the requirement of "concern of exerting a material impact or causing risk" is also considered together.

2-2. A developer business entity is an entity that develops and provides AI, and a deployer business entity is an entity that provides products or services using AI provided by a developer business entity. An enterprise that created a learning application using an external model is a deployer business entity, and bears legal obligations under the law.

2-3. The obligation is prior review of whether it qualifies as high-impact, and the discretion is requesting confirmation from the Minister. The Minister replies within 30 days from the date of receipt of the request, but may extend the period once by up to 30 days; if there is an objection to the reply, reconfirmation may be requested within 10 days from the date of receipt of the reply.

2-4. For example, human oversight and supervision can be implemented through a screen and permissions for teachers to review and modify scores before finalizing them; explanatory measures through guide materials communicating key scoring criteria and an overview of training data; risk management measures through responses by error type and regular inspection plans; and document retention through records of measure implementation. User protection measures can be implemented through channels for error reporting and correction.

2-5. Averages mask errors where overestimation and underestimation offset each other. For example, a tool that overestimates long responses and underestimates short responses can still yield a similar average. Instead, one must examine differences and their distribution at the individual response level, as well as differences by specific groups (response length, Korean language learners, etc.).

2-6. Change adjective-based performance level descriptions into observable criteria, such as "at least 1 claim and at least 2 pieces of supporting evidence," and attach benchmark responses for each level agreed upon through collaborative scoring among teachers. Both methods align interpretation among humans as well.

2-7. If teachers see the AI score first, that score acts as an anchor, easily pulling the teacher's judgment toward the AI. This is automation bias. If the teacher scores first, the AI score does not replace judgment, but serves as a signal indicating which responses to review again.

2-8. Multiple answers are possible. If results are reflected in grades, the procedure should include: the teacher finalizing all scores while using AI to identify responses with large discrepancies; notifying students in advance of AI-assisted use and teacher finalization; and having teachers rescore against rubrics to explain appeals. If it is unrecorded practice, immediate AI feedback, notification of AI use, sample checks, and reporting channels will suffice.

3-1. Reasonable consideration is given to the time, cost, and technology required to identify an individual, including the accessibility of other information. Because a classroom is a small group whose members know each other well, information to combine with clues—such as family stories or unique experiences—is close at hand. Therefore, even if names are removed, individuals can often be readily identified.

3-2. For example, one may choose: clarifying the purpose and processing only the minimum necessary (Paragraph 1); processing in a manner that minimizes infringement of privacy (Paragraph 6); or processing anonymously if the purpose can be achieved anonymously, or pseudonymously if not (Paragraph 7). Choosing the prohibition of use beyond the intended purpose (Paragraph 2) or safe management (Paragraph 4) is also acceptable under the Personal Information Protection Act (PIPA).

3-3. Obtaining the consent of a legal representative when consent is required and verifying whether consent was granted (Paragraph 1); that the minimum information necessary to obtain the legal representative's consent (name and contact information) may be collected directly from the child (Paragraph 2); and using easily understandable forms and language when notifying a child (Paragraph 3). Methods of verification include mobile phone identity verification, written signature and seal, receiving a declaration of consent via email, text notification of the fact of consent verification, and telephone consent.

3-4. Purpose of collection and use, items to be collected, period of retention and use, and the right to refuse consent along with any disadvantages resulting from such refusal. For students who do not consent, an alternative activity capable of achieving the same learning objectives is prepared.

3-5. Cases where information is provided in a form that does not allow identification of the data subject, for purposes such as compilation of statistics and academic research. A non-disclosure agreement is merely a protective measure restricting use by the recipient; it does not satisfy the requirements prescribed by law, such as consent from students and legal guardians or the "unidentifiable form" condition.

3-6. Use for training, retention period, account type, processing relationship and location (consignment and overseas transfer), and age criteria. The phrase "not used for training" does not indicate the retention period (how long data remains), nor does it reveal differences across account types or age criteria.

3-7. The statutory age of consent determines whose consent is legally valid, whereas age criteria in terms of service determine who can use the service and what conditions are required. For example, a 15-year-old student can legally consent on their own, but if the terms of service require parental permission for individuals under 18, parental permission must also be obtained. A 12-year-old student cannot use the service if the terms of service prohibit use by those under 13, even if a guardian consents.

3-8. Multiple answers are possible. For example, generalized prompts are appropriate for summarizing counseling logs; approved tools following the deletion of identifying information and clues are appropriate for feedback on student writing; and aggregated statistics are appropriate for item analysis. The reasoning for each choice should reflect two criteria: the minimum processing necessary for the purpose, and tools managed by the school.

4-1. Authorship (who wrote what and whether it was disclosed honestly), process (whether the competencies the assignment aims to develop were practiced), and accuracy and accountability (whether facts were verified and whether one can take responsibility for the content). Failure can occur on all three questions even without AI. For example, using inaccurate figures from an internet post without verification fails on accuracy, and submitting a paper written by someone else fails on authorship.

4-2. 0 No use (diagnostics assessing independent current proficiency and basic skills practice), 1 Exploration only (inquiry reports assessing both research and writing), 2 Drafting assistance (assignments teaching revision processes), and 3 Free use (assignments where AI literacy is the objective).

4-3. If only ideas were obtained, a record of use is kept. If facts provided by AI are used, primary sources are searched, verified, and cited. If AI sentences are quoted verbatim, quotation marks distinguish them from the student's own sentences, and the tool that generated the sentences is disclosed.

4-4. That they were not reliable enough to make decisions with lasting consequences for students; that they classified human-written text as AI-written; that they could disproportionately impact English language learners or students writing in a formulaic or concise style; and that detection could be evaded through minor modifications. They also noted that asking ChatGPT cannot confirm this either.

4-5. Because the false positive rate is multiplied across the entire pool of honest students. When there are hundreds of honest students, even a small percentage translates into multiple wrongful accusations, and the fewer students who actually used AI, the higher the proportion of wrongful cases among papers flagged as "AI-written." If those misclassifications are concentrated in specific groups, the problem becomes even more severe.

4-6. Any three of the following may be written: in-class writing of outlines and opening paragraphs, phased submissions, brief oral checks, or submission of conversation logs. When student numbers are large and time is limited, mechanisms with low cost per student that directly verify process and understanding—such as in-class drafting and 2-to-5-minute oral checks—are suitable. Mechanisms that teachers cannot fully review, such as full screen recordings, are unsuitable.

4-7. Criteria verification, evidence collection, notification and opportunity for explanation, decision, and notification and appeal. The stage that must not be omitted even if evidence is clear is the opportunity for explanation.

4-8. Multiple answers are possible. A good guide includes permitted levels, specific examples of what may and may not be done, records to be submitted, how to handle facts provided by AI, and procedures when a violation of standards is suspected. If it contains only ambiguous statements open to interpretation, such as "Do not rely on AI," it must be revised.

5-1. Whose problem it is and what the problem is, the current approach and its limitations (in numbers if possible), and what change defines success. For example, if the cause of the grading burden is not a shortage of personnel but ambiguity in the rubric, refining the rubric takes precedence over adopting AI tools.

5-2. Usage volume merely shows that it was used frequently; it does not indicate whether learning or tasks improved. Satisfaction merely shows perceptions and can diverge from actual performance, and because it is typically surveyed once at the end of a semester, it cannot be used to detect operational errors.

5-3. Any three of the following may be written: educational institutions are unprepared to validate tools, so schools must validate them themselves; data privacy protection and age thresholds (13 years for classroom AI tools) must be verified; teacher training must accompany adoption; and a human-centered, age-appropriate approach must be adopted.

5-4. The five dimensions are human-centered mindset, ethics of AI, AI foundations and applications, AI pedagogy, and AI for professional development; the three progression levels are acquire, deepen, and create. Training programs can be differentiated according to teachers' current levels, offering an acquire course to first-time users and a deepen course to existing users.

5-5. Govern, Map, Measure, and Manage. The cross-cutting function is Govern, meaning that governance is not finalized once at initial adoption, but must guide other functions throughout operation and adapt based on inspection findings.

5-6. Any three of the following may be cited: model replacement or feature additions; textbook revisions or changes in exam scope; changes in vendor terms of service or data conditions, or contract renewals; surges in error reports; or faculty turnover and grade-level progression.

5-7. For personal information exposure, suspend the relevant function, report to the privacy officer to determine the scope of exposure, and assess whether notification and regulatory reporting are required. For student safety or emotional incidents, a human contacts the affected student directly to connect them with support and suspends the function. For content errors, correct the cause, retest, and notify affected students of the corrections. An example of under-response is merely waiting for vendor reports; an example of over-response is suspending use across all grade levels until the following school year due to a single fixable error.

5-8. Multiple answers are possible. A good response must include a problem definition expressed in numbers, key metrics directly tied to the objective (not usage volume or satisfaction), a pilot scope of two to three classes and duration, and specific criteria for deciding whether to expand based on pilot results (including changes in target groups).