If AI can spin up a course, quizzes, and scripts in minutes, is the instructional designer content production or the design of what actually changes behavior?

The shift

Producing instructional artifacts — a course outline, module copy, quiz banks, video scripts, job aids, SCORM packages — goes from scarce (weeks of a designer's authoring time) to abundant (minutes, near-free, at any volume and reading level). What AI does not touch is whether the resulting training changes what a real person does on the job, or whether anyone can prove it did.

The axioms

• An instructional designer's deliverable is the finished learning content: the course, the assessment, the script. Scarce because authoring polished, on-brand, pedagogically-sequenced material took skilled human hours.
• Producing more courses is producing more value, so ID output is measured by throughput — courses shipped, modules built, hours of content in the library. Scarce because content was the expensive bottleneck, so counting it tracked the work.
• A course is worth building only if it's grounded in what learners actually need and can't yet do — needs analysis. Scarce because it requires talking to the business, watching the work, and deciding what the gap really is.
• Applying learning science — how to sequence, when to space and retrieve, what actually transfers to the job — is what makes training work rather than just exist. Scarce because it's specialist judgment about human cognition, not a production skill.
• Someone has to be accountable for whether the training worked: whether behavior changed and a business metric moved, not just whether the course was completed. Scarce because it requires owning an outcome you don't fully control.
• Instructional content must be accurate — factually correct, compliant, and not teaching the wrong thing. Scarce because it required a subject-matter expert's review time.
• The ID has to hold the trust of the business stakeholders who commission and fund the training. Scarce because trust is earned through relationship and delivered results, not produced.

Invalid axioms

1. The ID's deliverable is the finished learning content. Draft courses, quiz banks, scripts, and reading-level variants are now generated in minutes across almost any subject. The craft of authoring was the moat and it's now a commodity. The habit-trap: teams still scope ID work as content-production projects ("we need three courses built"), still hire and price for authoring throughput, and the "full-stack ID" whose value was mostly producing polished material is the most exposed — their scarce skill just became abundant.
2. ID output is measured by courses shipped. Counting content made sense when content was the expensive, rate-limiting step; it was a reasonable proxy for the work. Now that production is nearly free, throughput measures the cheap part and says nothing about the part that matters. The habit-trap: L&D dashboards, headcount justifications, and performance reviews still report modules built and completion rates — a metric that now rewards flooding the LMS with plausible courseware, which AI makes trivially easy to do and impossible to distinguish from real value by that measure alone.

Unchanged axioms

1. Learning-science judgment about what actually changes behavior stays scarce, and is now the load-bearing skill. AI produces content that looks instructionally sound — objectives, chunking, a quiz — without any guarantee it drives transfer to the job. Deciding what the intervention should even be (a course? a job aid? a manager conversation? no training at all, because it's a process problem), how to sequence for retention, and what will survive contact with real work is judgment AI doesn't supply. It generates the artifact; it doesn't decide whether the artifact is the right thing or whether it works. This is where the surviving ID role concentrates.
2. Accountability for whether the training worked stays scarce and can't be offloaded. A model can generate the course but can't be answerable for whether behavior changed or a metric moved. Owning that outcome — defining what "worked" means, standing behind it to the business — is human and unchanged. It gets more load-bearing precisely because the content around it is now free.
3. Needs analysis stays scarce. Deciding what the real gap is requires watching the work, talking to the business, and distinguishing a skill problem from a motivation, tooling, or process problem. AI can draft a needs-analysis template and even synthesize interview notes, but it can't own the judgment call about what's actually worth building — and building the wrong thing faster is now the cheap default.
4. Stakeholder trust stays scarce. The standing to tell a VP "the training you asked for won't fix this" and be believed comes from relationship and a track record, not from output. AI doesn't confer it. If anything it raises the premium on the ID who can be trusted to say no, because saying yes and generating the course is now the frictionless path.
5. Content accuracy still needs an accountable human check, and the burden grew. SME review was always the correctness gate. It didn't get cheaper — and AI-generated material is confidently wrong in ways human-authored drafts usually weren't, so the volume of output needing verification went up while the cost of producing it went to zero. Verification, not authoring, is now the expensive step.

New axioms

1. A flood of plausible courseware that looks fine and doesn't change behavior. When anyone can generate a competent-looking course in minutes, the LMS fills with material that passes a glance and a completion check but drives no transfer. The scarce act is no longer producing training; it's filtering the abundant supply down to what actually works — and nobody's job is currently defined as that filter.
2. Proving efficacy becomes the scarce act, and the field isn't set up to do it. Once content is free, the only defensible thing an ID owns is evidence that a specific intervention moved a real metric — error rates, ramp time, conversion, safety incidents. Most L&D functions have never measured past completion and satisfaction scores, because when content was the bottleneck, shipping it was the deliverable. Building the data fluency and measurement discipline to prove behavior change is now the constraint, and it's a skill set the profession largely doesn't have yet.
3. Verifying AI-generated instructional content is accurate and effective, at the volume it's now produced. The bottleneck moves from authoring to checking — is this factually right, compliant, on-strategy, and actually teaching the correct thing? Confidently-wrong content generated at scale means verification load rises exactly as production cost falls, and there's no established practice for doing that review at the speed the generation now happens.

Where it breaks

"We measure ID by courses shipped" (INVALID) collides with "the LMS is now flooded with plausible courseware that doesn't change behavior" (NEW). A throughput metric doesn't just fail to catch the flood — it actively rewards it, because generating more modules is now free and scores well on exactly the number being watched. Functions still counting output are incentivizing the precise failure mode the new abundance creates, and calling it productivity.

The second collision: "the ID's deliverable is finished content" (INVALID) collides with "proving efficacy is now the scarce act" (NEW). The full-stack ID whose value was authoring is being asked to survive on a completely different skill — data fluency and outcome measurement — that the content-production era never required them to build. The role isn't shrinking so much as swapping its center of gravity, and the person optimized for the old center is the one most exposed. How fast this bites depends on how quickly generation quality and agentic authoring tools improve, which is moving fast — the calibration here is that the content-commodity call is already true for straightforward corporate training and hardening for the rest.

ai-axioms.vercel.app/axioms/education/instructional-design-ai-axioms