Chapter 1. Instructional Design and Learning Materials

Learning objectives

• Include learning objectives and achievement standards in prompts so that AI does not stray outside the scope of the lesson.
• Explain the verification sequence for identifying factual errors, fabricated sources, and stereotypes in AI-generated learning materials.
• Design differentiated materials that adapt phrasing to student levels while preserving learning objectives.
• Distinguish the regulations and source-attribution obligations applicable when using published works in class materials across schools, universities, and private academies.
• Establish a risk-based verification plan that checks information with the highest consequences of error first.

Introduction

One Friday afternoon, a science teacher entered into an AI: "Create an 8th Grade 'Changes of State of Matter' worksheet." In 30 seconds, a plausible worksheet was generated. The teacher pressed print in admiration, and during Monday's class, three problems came to light all at once. The sentence on the worksheet stating that "the boiling point of water is 100°C regardless of altitude" was incorrect, half of the content belonged to what is taught in the next grade level, and students who needed foundational academic support got stuck at the very first paragraph.

In this story, the AI did nothing wrong. The AI merely created what it was told to create. Determining what the worksheet was for, who would read it, and what facts had been verified was the teacher's responsibility from the beginning. AI changed the speed of producing learning materials, but it did not change who holds responsibility for the learning materials.

This chapter covers, step by step, how to fulfill that responsibility. First, starting from learning objectives and achievement standards, you will learn how to prompt AI, and organize verification methods for identifying factual errors and bias in AI-generated materials. Next, we look at how to adapt materials to student levels while maintaining objectives, as well as the copyright rules for using published works in class materials across schools, universities, and private academies. Finally, we cover risk-based verification to determine what to check first when time is short.

Key concepts

Learning materials start from objectives

Good learning materials start with "what will be learned." In our curriculum, that starting point is the achievement standards. Achievement standards are statements describing the knowledge, skills, and attitudes that students should attain through learning in each subject, and they can be found by grade and subject at the NCIC (National Curriculum Information Center) operated by the National Education Commission and the Korea Institute for Curriculum and Evaluation (KICE).

AI does not know your school, grade, or unit. From the phrase "8th Grade Science worksheet" alone, it cannot know which curriculum, which achievement standard, or which lesson in the unit this session corresponds to. Therefore, AI generates an average representation of "8th Grade Science" commonly found on the Internet. This is why content not in the current unit gets mixed in.

By analogy, AI is a capable teaching assistant on their first day of work. Anthropic's prompt writing guide also uses the same analogy. It suggests showing the prompt to a colleague who lacks context and asking whether they can follow it. If the colleague is confused, the AI will be confused as well.

Five elements of a prompt

Prompts for creating class materials benefit from including five elements.

Examples are particularly powerful. Anthropic's guidance explains that examples are one of the most reliable ways to guide the format, tone, and structure of output, and recommends providing 3 to 5 relevant and diverse examples. For a worksheet, simply showing a single question at the desired level makes a significant difference in the result.

Conversely, grandly specifying only a role, such as "As a top education expert with 20 years of experience," and leaving the rest to the AI is not as effective as expected. This is because assigning a role sets the tone and perspective, but does not specify the scope of the current unit or the level of students in your class.

Revising prompts: From symptoms to causes

When you are dissatisfied with the result, examine the symptoms to identify the missing element.

Below is an example of revising the same request.

[Before revision]
Create a 7th-grade English past tense worksheet.

[After revision]
Objective: (Paste the achievement standard statement for this unit verbatim from the curriculum document)
      The objective for this lesson is to write 3 sentences in the past tense about what one did last weekend.
      Cover only the past tense forms of 10 regular and irregular verbs. Do not cover perfect tenses or conditionals.
Audience: 7th grade. For the foundational group, use 8 words or fewer per sentence.
Format: 6 fill-in-the-blank questions, 2 short-writing questions. Place answers and explanations separately on the very last page.
Example: (Foundational group question example) I ___ (play) soccer yesterday.
Constraints: Do not use the names of real people. Use only figures found in the materials provided by the teacher; if none, leave them blank.

Factual errors in AI-generated learning materials

Generative AI creates sentences by stringing them together probabilistically. As a result, grammatically flawless and confident sentences can be factually incorrect, a phenomenon commonly called hallucination. In learning materials, three things are particularly dangerous:

1. Figures and years. Plausible numbers such as populations, percentages, growth rates, and years of events are not noticeable even when incorrect.
2. Sources. When asked for sources, AI sometimes attaches non-existent report titles or random page numbers. The fact that a source is attached is no substitute for verification.
3. Quotations. Sentences enclosed in quotation marks as if spoken by real people are particularly dangerous. A "fictional interview" with a historical figure can be read by students as primary historical evidence.

Anthropic cites several methods to reduce hallucinations: instructing the model to say it does not know when it does not know, quoting directly from provided documents, requiring citations for every claim and retracting claims if no citation is found, and preventing the use of general knowledge outside the provided materials. It also adds that even though these techniques significantly reduce hallucinations, they cannot eliminate them, so important information must always be verified. In learning materials, this last sentence is essential.

Things that are not verification

There are methods commonly used during review that do not qualify as verification.

Verifying with primary sources

Verification is done using primary sources. Population and household figures are found in primary sources of public statistics such as the Korean Statistical Information Service (KOSIS); statutes in the provisions of the National Law Information Center; achievement standards in curriculum documents; and scientific concepts in textbooks and materials from authoritative institutions. After verifying, note the source and reference year on the worksheet. The source and reference year are helpful both when another teacher reuses the materials and when students learn the habit of verifying evidence.

Unverified sentences are handled in one of three ways: remove them, mark them as "Needs verification," or turn them into an activity where students locate primary sources and verify them directly. The last method makes for a particularly good classroom activity in the AI era.

Bias and stereotypes

Even if all facts are correct, AI-generated materials can still contain bias. For example, in word problem character setups, nurses and kindergarten teachers may all appear as women, while engineers and drivers all appear as men, or a specific region or generation may always take on the role of making mistakes. This is because common patterns in the training data are reproduced as-is.

The NIST AI RMF identifies "fairness with harmful bias managed" as one of the characteristics of trustworthy AI, dividing bias into three categories: systemic bias, computational and statistical bias, and human-cognitive bias. Applying this to learning materials: long-accumulated societal patterns of discrimination reproduced in occupational descriptions is an example of systemic bias, while skew resulting from underrepresentation of cases from specific regions or groups in training data is an example of computational and statistical bias. Furthermore, it clarifies several points: bias can occur in the absence of prejudice or discriminatory intent, is a broader issue than demographic balance or data representation, and a system with reduced harmful bias is not necessarily fair.

Translating this into learning materials review: Balancing the gender ratio of characters to 50:50 alone is not enough. One must also look at who takes on which roles, who is portrayed as competent and who makes mistakes, and how particular regions and cultures are depicted. In addition, remember that reviewers are less likely to question depictions that align with their own expectations (human-cognitive bias).

Differentiated materials: Changing expression while keeping objectives

AI excels at rewriting the same content at multiple levels. However, a prompt saying "Make this easy" carries two risks: one is that it does not become easier, and the other is that it becomes so simple that core concepts are omitted.

In particular, if the materials are for students taking the same assessment, all three levels must contain the same core concepts. If core concepts are omitted from foundational-level materials, reading may become easier, but preparing for assessments becomes more difficult. When reviewing, place the three versions side by side and verify the list of core concepts using a comparison table.

In prompts, describe levels through student characteristics rather than adjectives. Rather than "foundational level," describing it as "20 characters or fewer per sentence, plain-language explanations in parentheses for every technical term, and symbols explained with diagrams upon first appearance" yields far more consistent results. At this time, do not include student names or diagnostic scores. This is because defining a level requires characteristics of that level, not a specific student's personal information (Chapter 3).

Copyright: Using others' writings in class materials

Even when creating worksheets with AI, published works such as published poems, novels, essays, and news articles are frequently included. In this case, the key provisions are Article 25 (Exploitation for School Educational Purposes, etc.), Article 28 (Quotation from Works Made Public), and Article 37 (Indication of Sources) of the Copyright Act.

Article 25, Paragraph 3 allows the following schools and institutions to reproduce, distribute, perform, exhibit, or publicly transmit a portion of works made public for class purposes. If using the whole work is inevitable in light of the nature of the work or the purpose and format of its use, the entire work may also be used.

1. Schools established pursuant to special acts
2. Schools pursuant to the Early Childhood Education Act, the Elementary and Secondary Education Act, and the Higher Education Act
3. Educational institutions operated by the State or local governments
4. Education and training institutions operating accredited curricula pursuant to the Act on Recognition of Credits, etc. (limited to curricula based on remote learning)

Article 25, Paragraph 6 requires those who exploit works in this manner to pay compensation to the holder of author's economic rights according to standards determined and announced by the Minister of Culture, Sports and Tourism, but specifies that in cases of reproduction, etc. by high schools and schools of lower or equivalent levels, no compensation shall be paid. The right to receive compensation is exercised through an organization designated by the Minister of Culture, Sports and Tourism (Paragraph 7). Article 25, Paragraph 12 requires schools, etc., when publicly transmitting works—such as posting them to online classrooms—to take necessary measures prescribed by Presidential Decree, including technological measures to prevent copying. Article 37 requires anyone exploiting works in this manner to indicate their source, and Article 25 is not included in the list of exceptions under Article 37, Paragraph 1.

This can be summarized in the following table:

In places where Article 25 does not apply, one must examine whether requirements of other provisions are met, such as quotation under Article 28 or fair use under Article 35-5; if not applicable, permission must be obtained from the copyright holder. Article 35-5 permits uses that do not conflict with the normal exploitation of works and do not unreasonably prejudice the legitimate interests of authors, evaluated by taking into account the purpose and character of the use, the type and purpose of the work, the proportion and substantiality of the portion used, and the effect upon the market. Article 28 allows quotation from works made public for news reporting, criticism, education, research, etc., within reasonable limits in conformity with fair practice. Quotation requires a relationship where one's own writing is primary and another's writing supports it; therefore, filling most of a worksheet with someone else's writing can hardly be considered quotation.

AI-rewritten text and AI-generated images

The assumption that "because AI summarized or rewrote the original text, the original text was not used" is dangerous. If a summary relies on the content and expression of the original text, copyright issues regarding using the original work do not disappear, and this is even more so if parts of the original text are included alongside it. Furthermore, if sentences output by AI are nearly identical to existing poems or song lyrics, issues regarding the use of those existing works arise regardless of the fact that the output channel was AI.

When using illustrations or sentences generated by an AI service, ask two separate questions. One is a question of copyright law, and the other is the terms of use set by the service's terms of service. While the Copyright Act defines a work as "a creative production that expresses human thoughts or emotions" (Article 2, Item 1), the service terms independently govern how outputs may be used through a contract. It is a safe practice to check the terms of use in the service agreement before evaluating copyright protection.

Risk-based verification

Time for verification is always scarce. If there is no time to verify all 30 sentences against primary sources, what should be done first? The criterion is "how great is the harm if it is wrong?"

1. Facts students will use as the basis for assessment or judgment: If students will calculate using these figures or cite them as evidence in a report, this is the top priority.
2. Content related to safety: For items such as laboratory safety rules, first aid, or age ratings, errors can lead to physical injury.
3. Statements regarding real people and institutions: Fabricated quotes and incorrect facts create issues of reputation and credibility.
4. Background explanations and examples: Even if incorrect, the harm is relatively minor.

Sentences that cannot be verified in time should be removed or marked as "Needs verification" so students do not use them as evidence. The method of randomly checking only a few items may appear fair, but it can miss sentences with high potential harm, making it unsuitable for this situation.

The process of producing AI learning materials

Organizing the content covered so far into a sequence yields the following. The outcome of each stage becomes the material for the next stage.

1. Determine the unit's achievement standards and the learning objectives for the lesson.
2. Create a draft using a prompt containing objective, audience, format, examples, and constraints.
3. Review and revise the draft for facts, level, bias, and use of copyrighted works.
4. Pilot the revised edition in one class first, and refine it based on student responses.
5. Note the sources and the fact that AI was utilized on the final version, and share it in the subject materials archive.

Going deeper

Does AI with search capabilities solve the problem?

Because AI with search capabilities attaches links to every answer, it may appear as though the issue of factual errors has been solved. It is also true that it actually reduces the time spent finding evidence. However, two limitations remain. First, whether that sentence is truly on the linked page can only be confirmed when the teacher cross-checks it directly against the original text. The AI may summarize the page content incorrectly or mix in figures from another page. Second, the page returned as a search result may not be the primary source itself. It is common for blogs or summary articles to misquote official statistics.

Therefore, use the search feature as a tool to quickly find out "where to look," and carry out the verification directly in the primary source by opening the link. When a teacher demonstrates this process in class, students also learn how to distinguish between a link and actual evidence.

Differentiated materials and the trap of low expectations

When creating differentiated materials, the most important thing to guard against is fixing low expectations on basic-level students. When asking an AI for materials "for a basic class," it often removes concepts and leaves only interest-driven activities. While this may reduce students' resistance, if they are taking the same assessment, it ultimately deprives them of the opportunity to learn core concepts.

Good differentiated materials share the same goal but provide different ladders. They give the basic level a tighter ladder with glossary definitions, shorter sentences, and step-by-step questions, while giving the advanced level a higher ladder with longer passages and open-ended questions. If a curriculum is designed so that the assessment scope itself differs by level, one may consider the option of reducing the scope, but that is a decision made by the teacher at the curriculum level, not something to casually slip into an AI prompt.

Borderline cases among private academies, universities, and schools

When applying Article 25 of the Copyright Act, "where you teach" changes the conclusion. Even if the same two paragraphs of an essay are posted to an online learning room in the exact same manner, a middle school does so without compensation, a university pays compensation, and a private academy is judged not under Article 25, but according to licensing permission or quotation requirements. One must also verify whether the institution is on the statutory list, such as educational institutions operated by the state or local governments, or educational and training institutions operating accredited distance learning courses.

There are also cases where the boundary is ambiguous: for example, when a school entrusts an after-school program to an external vendor, or when a university runs an open lecture for the general public. In these cases, one must separately determine whether the requirement of "use by a school for instructional purposes" is met, and if it is difficult to judge, it is safer to obtain licensing permission. On the exam, when such borderline cases appear, the nature of the institution is stated clearly in the question, so do not miss that sentence.

At work

Sandstream Middle School: Friday afternoon worksheet

A second-year science teacher at Sandstream Middle School was using AI to create a class worksheet on 'Changes of State of Matter' for the following week. The initial output contained 12 questions; of these, 3 covered 'Heat Transfer,' which was not addressed in this unit, one explanation regarding boiling points was incorrect because it omitted changes according to altitude, and the sentences were long, making them seem difficult to read for seven students receiving basic academic assistance.

The first method the teacher thought of was clicking the "Regenerate" button three times and keeping only the questions that appeared in common. However, all three attempts yielded the same explanation, and out-of-scope questions continued to be mixed in. Realizing that the results would not differ significantly if the instructions remained unchanged, the teacher abandoned this method.

The teacher revised the prompt. The teacher inserted the achievement standard statement, added the constraint "Do not cover Heat Transfer," and attached one example of a basic-level question and the format (8 questions, answers at the very end). In the new output, the scope issue disappeared. Next, the teacher conducted a risk-based review. Four figures for boiling points and melting points—which students would use as evidence in their lab reports—were cross-checked against the textbook, and one was corrected. Three laboratory safety rules were also cross-checked against the school laboratory safety guidelines.

Finally, while creating the basic-level version, the teacher created a table listing the core concepts across the three levels (changes of state and the transfer of heat, changes in particle arrangement) and verified that they were included in all versions. At the very end of the worksheet, the teacher wrote: "This worksheet was drafted with generative AI and reviewed by the teacher. Source of numerical data: Grade 2 Science Textbook."

Sunrise Korean Language Academy: Same essay, different rules

An instructor at Sunrise Korean Language Academy saw a reading comprehension worksheet shared by a teacher at a nearby middle school and wanted to create materials in the same way. That worksheet included two paragraphs from a published essay, and the school had posted it to a class learning room with source attribution and without paying compensation. The instructor thought, "Since the school did it, we just need to cite the source as well."

The director of the academy halted this decision. The legal basis that allowed the school to do so was Article 25, Paragraphs 3 and 6 of the Copyright Act, but this academy was not a school or institution on that list. The director compared three options. First, the method of having AI summarize the essay and using only the summary was discarded because, being a summary reliant on the original text, it did not avoid the issue. Second, the method of including the two paragraphs of the essay verbatim as part of a critique activity was judged difficult to satisfy the primary-secondary relationship required for quotation, because the core of the worksheet was the essay itself. Third, the director chose the method of obtaining licensing permission through the publisher.

While waiting for permission, the instructor prepared a different activity. It was an activity designed so that a single paragraph chosen directly by the student was briefly quoted, and the student's critique served as the center of the piece. In this case as well, the source of the quoted text was cited. The director added the following sentence to the academy's instructor training materials: "What is permitted in a school is not necessarily permitted in an academy."

Common mistakes

1. "Answers from three AIs are the same, so it must be correct." When multiple sources agree, it appears reliable. However, AIs can share the same incorrect information, and consensus does not substitute for checking the primary source.
2. "Sources are attached, so it has been verified." A citation is merely the starting point of verification. AI can fabricate nonexistent sources or attach completely unrelated pages, so open the link and verify the content in the primary source.
3. "If you give it an expert role, it will create it well on its own." Assigning a role seems plausible, but it cannot inform the model of the scope of this unit or the level of students in your class. You must write the objectives, target audience, format, and examples yourself.
4. "An easy version is a version with fewer concepts." This seems considerate of students, but removing core concepts from students who take the same assessment actually puts them at a disadvantage. Change the expressions and maintain the objectives.
5. "Balancing the gender ratio completes the bias check." This seems convenient because it can be verified with numbers. However, bias is also an issue of roles and portrayals, and it occurs even without intent.
6. "If it is for instructional purposes, you do not need to cite the source." Because use for instructional purposes is permitted, other obligations might also seem exempt, but Article 25 is not among the exceptions to source attribution in Article 37 of the Copyright Act.
7. "If it is permitted in a school, it is permitted in an academy." Because both are educational activities, it seems like the same rules would apply, but Article 25 defines eligible institutions as a specific list. Universities pay compensation, and private academies outside the list must evaluate licensing permission or quotation requirements.
8. "Checking a few at random is fair." Sample checking looks systematic. However, when time is short, it is correct to check sentences that carry greater potential harm first, rather than relying on a fair sample.

Summary at a glance

Check the official docs

• National Curriculum Information Center (NCIC) — Where to find achievement standards and curriculum documents by subject and grade level.
• Anthropic — Prompting best practices — Read the sections on clear instructions, providing context, and using examples (3–5 recommended).
• Anthropic — Reduce hallucinations — Read the techniques for reducing hallucinations and the note to "always verify important information."
• NIST AI RMF — Characteristics of trustworthy AI — Check the three categories of bias in "Fair – with Harmful Bias Managed."
• Korea Law Information Center — Article 25 of the Copyright Act — Check the list of eligible institutions (Paragraph 3), compensation (Paragraph 6), and public transmission measures (Paragraph 12).
• Korea Law Information Center — Article 28 of the Copyright Act — A statutory provision that defines the requirements for quotation in a single sentence.
• Korea Law Information Center — Article 37 of the Copyright Act — Check the obligation to cite sources and the list of exceptions.

Self-check

1-1. State five elements that should ideally be included in a prompt requesting instructional materials from AI, and state one problem that arises when each element is missing.

1-2. Explain why getting matching answers after asking the same question to multiple AI services cannot be considered fact-checking.

1-3. State the three categories of bias classified by the NIST AI RMF, and identify which category can appear in the textbook reviewer themselves.

1-4. When creating reading materials at three levels—basic, standard, and advanced—state two elements that may vary across levels and two elements that must not be changed.

1-5. Explain how compensation payment obligations differ when a high school, a university, and a private academy not listed in Article 25, Paragraph 3 of the Copyright Act each post portions of disclosed works online as instructional materials.

1-6. Explain how the statutory structure of the Copyright Act confirms that the obligation to indicate the source remains even when use for instructional purposes is permitted.

1-7. When you have only 20 minutes of review time and 25 statements to verify, state the criteria for determining the order of verification, and suggest two ways to handle statements that were not verified.

1-8. (Design task) Choose one class session from a subject you teach, and write a prompt of no more than 10 lines that contains all five elements from this chapter.

Hands-on tasks