An AI authoring tool just handed a junior designer a forty-screen module on data-handling policy in ninety seconds. It is fluent, well-formatted, and confident. It is also built on nothing: no stated objective, no Bloom's level, no alignment between what the quiz asks and what the job actually requires. The designer's manager glances at it and asks a single question that the tool cannot answer and the designer, three weeks into the role, also cannot answer: "What is this module supposed to make someone able to do, and how do we know the test proves it?" That question is a hundred years of instructional-design evidence compressed into one sentence, and no AI tool will ever teach it to you. This lesson does.

Why the Frameworks Survive the AI Wave

There is a comforting story circulating in 2026 that AI has made instructional design obsolete, that the models can now do "the whole thing," and that the dusty frameworks taught in graduate programs are relics. The story is exactly backwards. AI collapsed the cost of producing learning content, which means the production work is no longer where the value sits. The value moved to the judgment that decides what to build, whether it is correct, whether it measures the right thing, and whether it changed behavior. That judgment lives entirely inside the evidence-based frameworks AI cannot replace, because those frameworks are not about generating screens. They are about whether the screens should exist and whether they work.

Here is the term that anchors the lesson. Instructional design is the discipline of deciding what people need to be able to do, designing an experience that builds that capability, and proving it worked. Why you care: an AI tool can draft any of the artifacts in that sentence, but it cannot make the decisions the sentence describes, and the decisions are the job. A model will happily generate a module with no objective, a quiz that tests reading comprehension instead of the skill, and a structure that violates every principle of how adults learn, and it will do all of that in fluent, confident prose that looks finished. The frameworks are your defense against being fooled by fluency. They are also the vocabulary an AI tool will never teach you, which is why an instructional designer who knows them is worth more in the AI era, not less.

This lesson is a guided tour of the load-bearing frameworks: the process models that organize a build (ADDIE and SAM), the alignment discipline that keeps a course honest (backward design and constructive alignment), the taxonomy that calibrates difficulty (Bloom's revised), the event sequence that structures a lesson (Gagne's nine events), the multimedia rules that govern media (Mayer's principles), the design principles that anchor problem-centered learning (Merrill), the focusing method that kills unnecessary content (Cathy Moore's action mapping), and the heuristic that situates formal training inside the wider flow of work (70-20-10). You do not need to memorize every one cold. You need to recognize each one, know what job it does, and know the question it lets you ask of any AI-drafted build.

AI made content cheap, so the scarce skill is no longer building the module. It is knowing whether the module should exist, whether it is correct, and whether it works. That skill is the frameworks.

The Process Models: ADDIE and SAM

Every learning build follows a process, whether the designer names it or not. The two named processes worth knowing are ADDIE and SAM, and the difference between them tells you something real about how to use AI.

ADDIE: The Spine of the Profession

ADDIE stands for Analysis, Design, Development, Implementation, and Evaluation, the five phases of a classic instructional-design process. It originated around 1975, developed at Florida State University for the US Army to standardize how training was built across a large organization. It is not a brand or a proprietary method; it is the underlying grammar of nearly every design process in corporate learning. Analysis asks what the real performance gap is and who the learners are. Design specifies the objectives, the structure, and the assessment strategy. Development builds the actual content and media. Implementation delivers it. Evaluation measures whether it worked.

Why ADDIE matters in the AI era: AI is overwhelmingly a Development-phase accelerant. It drafts content, media, and items at high speed. But the phases that determine whether the course is any good, Analysis (what gap, what learner) and Evaluation (did it change behavior), are exactly the phases AI cannot own, because they require deciding what is true about the organization and what counts as success. A designer who lets AI rush the build while skipping Analysis is producing a fast, confident answer to a question nobody verified was the right question. ADDIE's value in 2026 is that it names the human-owned bookends around the AI-accelerated middle.

SAM: The Iterative Alternative

SAM, the Successive Approximation Model, was developed by Michael Allen of Allen Interactions as a more iterative, prototype-driven alternative to a strictly linear ADDIE. Instead of moving once through analyze, design, develop, SAM cycles through rapid rounds of design, prototype, and review, getting a rough version in front of stakeholders early and refining it across short iterations. The insight is that learning needs are often discovered by building something imperfect and reacting to it, not by specifying everything up front.

SAM and AI are natural partners, which is also the trap. AI makes prototyping nearly free, so a team can spin up version after version in an afternoon. That is genuinely powerful for ideation. The danger is mistaking a fast iteration loop for a verified one: cycling quickly through prototypes that nobody checked against a source of truth just produces polished guesses faster. The discipline is to keep the verification gate, the SME sign-off, and the accessibility check inside the loop, not to drop them because the iterations are quick. SAM with AI is a way to design faster, never a way to verify less.

The Alignment Discipline: Backward Design and Constructive Alignment

If you learn one framework from this lesson, learn this one, because it is the question the AI module almost always fails. Backward design means you start from the end: first decide what the learner must be able to do, then design the assessment that would prove they can do it, and only then build the content that gets them there. Constructive alignment is the principle that the objective, the assessment, and the learning activities must all point at the same thing. If the objective says "configure the firewall rule," the assessment must require configuring a firewall rule, and the content must teach configuring a firewall rule. When those three drift apart, the course is broken even if every screen looks perfect.

This is the single most common failure in AI-generated learning, and it is invisible to anyone who does not know the framework. An AI tool will generate a module on, say, harassment prevention, with a polished narrative and a ten-item multiple-choice quiz. Every screen reads well. But the objective (if one was even stated) was "recognize and respond appropriately to harassment," and the quiz asks learners to recall the definition of a protected class and match policy terms to their descriptions. The quiz tests recall of vocabulary; the objective demanded recognition and response in a real situation. The course is misaligned. It will certify people as competent at remembering definitions while teaching nothing about what to do when it happens, and the gap will surface only when an incident does. Backward design is the antidote: the assessment is designed from the objective before the content exists, so the AI cannot quietly substitute an easier thing to measure.

A beautiful module that tests the wrong thing is not a course. It is a confident lie about what your people can do, shipped at scale.

Bloom's, Gagne, Mayer, and Merrill: The Design Toolkit

Four named frameworks give you the precision tools to specify difficulty, sequence a lesson, govern media, and structure problem-centered learning. Each one also gives you a specific question to ask of an AI draft.

Bloom's Revised Taxonomy

Bloom's Taxonomy is a hierarchy of cognitive levels, revised in 2001 by Anderson and Krathwohl into a verb-based ladder: Remember, Understand, Apply, Analyze, Evaluate, and Create. It tells you the cognitive altitude an objective is operating at. "List the four steps" is Remember. "Decide which procedure applies to this situation" is Analyze or Evaluate. Why you care: the most common AI alignment failure is writing an objective at a high level ("evaluate the risk") and then generating an assessment at a low level ("recall the definition of risk"). Bloom's gives you the language to catch the mismatch. You read the objective's verb, you read the assessment's verb, and if they are not at the same level, the course is misaligned no matter how good it looks. AI will draft objectives and items at any level you specify, which is exactly why you must specify, and verify, the level.

Gagne's Nine Events of Instruction

Gagne's Nine Events of Instruction, developed by Robert Gagne, is a sequence for structuring a single lesson so it matches how attention and memory actually work: gain attention, inform the learner of the objective, stimulate recall of prior knowledge, present the content, provide guidance, elicit performance (practice), provide feedback, assess performance, and enhance retention and transfer. (You will see the name written both as Gagne and as Gagne with an accent; the framework is the same.) Why you care: AI-generated modules routinely skip the events that make learning stick, especially elicit performance and provide feedback. A model loves to present content and then jump straight to a final quiz, dumping information without the practice-and-feedback loop in the middle. Gagne's sequence is your checklist for whether a draft is a lesson or just a narrated document with a test bolted on.

Mayer's Multimedia Principles

Mayer's Cognitive Theory of Multimedia Learning is a set of twelve evidence-based principles for how to combine words and visuals without overloading the learner. The most operationally important for AI work are the coherence principle (cut anything that does not serve the objective, including decorative extras), the signaling principle (highlight what matters), and the redundancy principle (do not narrate identical on-screen text word for word, because reading and hearing the same words at once overloads working memory). Why you care: AI-generated media violates these constantly. AI video tools love to generate a slide of dense text and then have an avatar read it aloud verbatim, a textbook redundancy violation that measurably hurts learning. AI loves decorative stock imagery that adds cognitive load and serves no objective, a coherence violation. Mayer's principles turn "this video feels off" into a specific, nameable defect you can fix.

Merrill's First Principles of Instruction

Merrill's First Principles of Instruction, published by M. David Merrill in 2002, distills decades of design research into five conditions that promote learning: it is problem-centered, it activates prior knowledge, it demonstrates the skill, it gives learners a chance to apply it, and it integrates the new skill into the learner's world. Why you care: AI defaults to information-centered content, walls of facts to be remembered. Merrill's principles push you to demand problem-centered design: does this module put the learner in a realistic task, or does it just tell them things? An AI draft can be re-prompted toward problem-centered structure, but only by a designer who knows to ask for it.

Action Mapping and the 70-20-10 Heuristic

Two more frameworks address a question AI never asks on its own: should this content exist at all, and where does formal training even fit?

Cathy Moore's Action Mapping starts from the on-the-job action, not the content. The method asks: what do people need to do differently, what is currently stopping them, and what practice would close that gap. Content is added only when a real performance gap genuinely requires knowledge. Why this matters enormously in the AI era: because AI makes content nearly free to produce, the temptation is to generate everything, to dump every fact the SME mentioned into screens because it costs almost nothing to do so. Action mapping is the discipline that resists the dump. It forces the question "will building this actually change what someone does on the job," and most of the time the honest answer is that a job aid or a single practice scenario would beat a forty-screen course. When production is cheap, the scarce discipline is restraint, and action mapping is restraint formalized.

70-20-10 is a heuristic, popularized through the Center for Creative Leadership, suggesting that workplace capability comes roughly 70% from experience and on-the-job challenge, 20% from social learning and others, and 10% from formal training. Treat it as a rough framing, not a measured law; the exact ratios are not precise science. Why you care: it situates the formal course, the thing AI is so good at generating, as a small slice of how people actually become capable. An AI tool can flood you with formal content, the 10%, while the 70% and 20% (real practice, coaching, working alongside experts) remain where most capability is actually built. The heuristic keeps the AI-accelerated formal course in proportion and reminds you that producing more courses is not the same as building more capability.

A Worked Example: An AI Draft Against the Frameworks

Watch the frameworks turn a vague "this module feels weak" into a precise, defensible critique. A designer receives an AI-generated module on a new expense-reporting policy and runs it through the frameworks.

Before the frameworks, the designer could only say the module felt off. After the frameworks, the critique is specific and actionable: this should probably be a job aid, not a course; if it stays a course, write the objective first, redesign the quiz to require applying the policy to a realistic expense scenario, add a guided-practice step with feedback, strip the redundant narration, and cut the decorative video. The AI did the production. The frameworks did the thinking, and the thinking is the part that made the module actually teach something. None of those seven checks came from the tool. Every one came from the designer's knowledge of how learning works.

This is the entire argument of the lesson in one example. The AI is fast at the development phase and useless at every judgment that determines whether the development was worth doing. The frameworks are how a learning professional supplies that judgment, fast and defensibly, on top of AI's speed. A designer who knows them can take an AI draft from "looks finished" to "actually works" in an hour. A designer who does not know them ships the draft and learns its flaws from an incident report or an audit.

Key Takeaways

• AI collapsed the cost of producing content, which moved the value of the profession to the judgment the frameworks encode: what to build, whether it is correct, and whether it works. The frameworks make a designer worth more in the AI era, not less.
• ADDIE (Analysis, Design, Development, Implementation, Evaluation; origin around 1975 at Florida State for the US Army) names the human-owned bookends (Analysis and Evaluation) around the AI-accelerated middle (Development).
• SAM (Michael Allen's Successive Approximation Model) is iterative and pairs naturally with AI prototyping, but fast iteration is not verified iteration; keep the verification gate inside the loop.
• Backward design and constructive alignment are the discipline AI most often fails: design the assessment from the objective first, and keep objective, assessment, and content pointing at the same thing.
• Bloom's revised taxonomy (2001, Anderson and Krathwohl) calibrates cognitive level so you can catch an objective and an assessment sitting at different altitudes; Gagne's nine events check whether a draft is a real lesson with practice and feedback or just narrated content plus a quiz.
• Mayer's twelve multimedia principles (coherence, signaling, redundancy and others) name the cognitive-load defects AI media generates by default, like narrating on-screen text verbatim over decorative video.
• Merrill's First Principles (2002) push for problem-centered design over AI's default information dump; Cathy Moore's action mapping is the restraint that asks whether a course should exist at all, which matters most precisely because AI makes content cheap.
• 70-20-10 is a heuristic, not a law: it keeps the AI-generated formal course (the 10%) in proportion to the experience and social learning where most capability is actually built.