A Safety-First Framework For Using AI In Your Instructional Design Workflow
Everyone in Learning and Development (L&D) is currently under pressure to produce more content, faster. Artificial Intelligence seems like the obvious answer. It promises to cut development time in half and instantly generate scenarios, quizzes, and summaries. But there is a catch. Large Language Models (LLMs) like ChatGPT or Claude are not knowledge engines; they are prediction engines. They do not care about the truth. They care about what the next word should likely be. In a creative writing class, that is a feature. In compliance training, safety protocols, or technical onboarding, it is a liability.
When an AI “hallucinates” (i.e., it confidently states a fact that isn’t true), it creates a mess. If a learner follows a hallucinated safety step, people get hurt. If a manager follows a hallucinated HR policy, the company gets sued. This guide details a safety-first workflow; think of it as treating AI not as an expert, but as an unreliable intern who needs their work checked line by line.
In this article…
Why AI Helps, And Where It Breaks In L&D
AI is incredible at structural tasks. It can take a messy transcript and find the main points. It can rewrite passive voice into active voice. It can brainstorm ten ideas for a role-play in seconds. But it fails when you ask it to be accurate without guardrails.
The Failure Modes
- The phantom citation: You ask for research on adult learning. The AI gives you a perfect APA citation for a study that does not exist.
- The context collapse: You upload a policy from 2019. The AI uses it, ignoring the 2023 update you mentioned in the prompt because the 2019 text was longer.
- The “average” trap: AI is trained on the internet. If you ask for a leadership course, it gives you generic, average advice that might contradict your specific company culture.
- Bias amplification: Unless checked, AI often defaults to gendered language (e.g., doctors are “he,” assistants are “she”) based on historical training data.
The “Safe AI” Workflow In 6 Steps
To use AI safely, you must change how you write your prompts. Never ask open-ended questions like “Write a course on fire safety.” That is gambling. Instead, use a “source pack” approach.
Step 1: Define The Goal And The “Do Not” List
Start with the end in mind. What is the specific performance objective?
- Who: Senior Sales Managers.
- Goal: Apply the new “Consultative Closing” matrix.
- Constraint: Do not use general sales advice found online. Use only our internal terminology.
Step 2: Create The Source Pack (The Boundary)
This is the most important step for safety. Gather the PDFs, transcripts, and slide decks that contain the truth.
- Clean the data. If you upload a transcript, delete the small talk first.
- Prompt strategy: Explicitly tell the AI: “Use ONLY the provided source text to answer. If the answer is not in the text, state ‘I do not know.’ Do not use outside knowledge.”
Step 3: Generate Drafts, Not Final Content
Use the tool to build the skeleton, not the muscle.
- Ask for an outline based on the source pack.
- Ask for three different analogies to explain a complex term from the text.
- Ask it to summarize a 10-page technical manual into a one-page job aid.
Step 4: The Fact-Check And Evidence Tag
Before you fix the flow, you must fix the facts.
- Traceability: If the AI makes a claim, can you find the sentence in your source pack that supports it?
- Numbers and dates: AI is notoriously bad at math and timelines. Check every single number manually.
- Links: Click every URL. AI often generates dead or invented links.
Step 5: The Instructional Design QA
Once the facts are clean, look at the learning science.
- Cognitive load: Did the AI dump a wall of text? Break it up.
- Bloom’s taxonomy: Are the quiz questions just testing memory (low level), or are they testing application (high level)? AI defaults to memory questions because they are easier to generate.
- Tone: Does it sound like a robot? Inject human warmth and empathy.
Step 6: Pilot And Iterate
Don’t launch to the whole company. Send the module to five users. Watch them take it. If they get stuck on an AI-generated explanation, it’s not clear enough. Rewrite it yourself.
The QA Checklist
Before you publish any AI-generated content, run it through this six-point inspection.
Accuracy And Sourcing
- The “Ctrl+F” test: Can every factual claim be found in your original source documents?
- Hallucination check: Verify that no external statistics, dates, or regulations were invented by the AI.
- Link validation: Click every hyperlink. Ensure they lead to live, relevant pages, not dead ends.
Alignment To Objectives
- The fluff filter: Did the AI add “nice to know” history or background info? If it doesn’t support the learning goal, delete it.
- Action-oriented: Does the content teach the learner how to do the task, or just about the task?
- Audience match: Is the complexity level right? (e.g., Don’t explain “what is a browser” to software engineers).
Assessment Validity
- Distractor check: In multiple-choice questions, are the wrong answers plausible? AI often writes obviously silly distractors that make the quiz too easy.
- Answer key: Is the correct answer indisputably correct based on your policy?
- Feedback: Did the AI generate helpful feedback for why an answer is wrong?
Cognitive Load And Clarity
- Brevity: Are paragraphs short (3-4 sentences)? AI tends to be verbose.
- Active voice: Did the AI use passive voice (e.g., “The form should be signed”)? Change it to active (e.g., “Sign the form”).
- Formatting: Are lists used instead of dense blocks of text?
Accessibility Basics
- Alt text: If the AI has suggested images, are the descriptions functional and descriptive for screen readers?
- Reading level: Is the language simple enough? (Aim for Grade 8 reading level for general compliance).
- Contrast: If the AI has generated slide layouts, is the text legible against the background?
Tone, Inclusivity, Policy, And Compliance
- Bias scan: Check pronouns and roles. Did the AI make the manager “he” and the assistant “she”?
- Brand voice: Does it sound like a robot or a human? Add warmth and empathy where needed.
- Safety and legal: Ensure no absolute promises are made (e.g., “Follow this, and you will never be injured”) that could create liability.
Two Mini Examples
Example A: The SME Transcript
- Context: You have a rambling 45-minute recording of a Product Manager explaining a new software feature.
- The bad way: You paste it all in and say, “Write a script.”
- The result: The AI includes the Product Manager’s complaints about the engineering team and misses the critical login step.
- The safe way:
- Clean: Delete the complaints from the transcript manually.
- Prompt: “Act as a technical writer. Based only on the attached transcript, list the step-by-step login process. Format as a numbered list.”
- QA: You verify the steps against the actual software sandbox. You realize the AI missed “Click Save.” You add it manually.
Example B: The Quiz Generator
- Context: You need a quiz for a Code of Conduct course.
- The bad way: “Write 5 hard questions about ethics.”
- The result: The AI asks philosophical questions like “What is the nature of truth?” which is irrelevant to company policy.
- The safe way:
- Prompt: “Using the attached ‘Gifts and Hospitality Policy’ PDF, write 3 scenario-based multiple-choice questions. The learner must decide if they can accept a gift. For every correct answer, quote the specific clause from the PDF.”
- QA: You check the clauses. You ensure the scenarios feel real, not cartoonish.
Measurement: Did It Work?
Speed of creation is a vanity metric. You need to measure effectiveness.
What To Track
- Quiz reliability: Look at the analytics. If 100% of learners get Question 3 correct, it’s too easy. If 0% get it right, the AI wrote a confusing question, or the content didn’t cover it.
- Confidence scores: Ask learners, “How confident are you in applying this skill?” If confidence is low, the AI-generated content might have been too abstract.
The A/B Test
If you want to prove this works, run a test.
- Group A: Takes the old, human-written legacy course.
- Group B: Takes the new AI-assisted (and human-verified) course.
Compare the time to proficiency. If Group B learns the same amount in half the time, your workflow is a success.
Closing
AI is a tool, like a calculator or a spell-checker. You wouldn’t publish a financial report without double-checking the calculator’s inputs. Do not publish training without double-checking the AI’s output. The goal isn’t to let AI do the work. The goal is to let AI do the drudgery (the summarizing, the formatting, the drafting) so you can focus on the high-value work: strategy, context, and human connection.