A safety-first framework for using AI in instructional design workflows
Everyone in learning and development (L&D) is now under pressure to create 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’s a catch. Large-scale language models (LLMs) like ChatGPT and Claude are not knowledge engines. They are predictive engines. They don’t care about the truth. They care what the next word will be. In a creative writing class, that’s a feature. Whether it’s compliance training, safety protocols, or technical onboarding, this is your responsibility.
When AI “hallucinates” (i.e. confidently states facts that are not true), confusion ensues. People get hurt when learners follow safety procedures based on illusions. When managers follow illusory HR policies, companies get sued. This guide details the safety-first workflow. Think of this as treating the AI less like an expert and more like an untrustworthy intern whose work needs to be checked line by line.
In this article…
Why AI is useful and where it impacts L&D
AI is amazing for structural tasks. You can also find key points in a messy transcript. Passive voice can be changed to active voice. Brainstorm 10 ideas for roleplay in seconds. But if you strive for accuracy without guardrails, you will fail.
failure mode
Phantom Quote: You’re looking for research on adult learning. AI provides perfect APA citations for studies that don’t exist. Context collapse: Upload your 2019 policy. The AI will use it, but because the 2019 text was longer, it will ignore the 2023 update mentioned in the prompt. The “average” trap: AI is trained on the internet. If you request a leadership course, you will be given general and average advice that may be inconsistent with your specific company culture. Bias Amplification: If unchecked, AI will often default to gendered language (e.g., “he” for doctor and “she” for assistant) based on past training data.
6-Step “Safe AI” Workflow
To use AI safely, you need to change the way you write prompts. Never ask open-ended questions like “Create a course on fire safety.” It’s a gamble. Instead, use the “source pack” approach.
Step 1: Define your goals and “to-do” list
Start with the end in mind. What are your specific performance goals?
Target audience: Senior sales managers. Objective: Apply the new “Consulting End” matrix. Limitation: Don’t use generic sales advice you can find online. Please use only our internal terminology.
Step 2: Create the source pack (boundary)
This is the most important step for safety. Collect PDFs, transcripts, and slide decks containing the truth.
Clean up your data. If you’re uploading a transcript, please remove the chat first. Prompt strategy: Explicitly tell the AI to: “Please use only the source text provided for your answer. If your answer is not in the text, please state ‘I don’t know.’ Don’t use outside knowledge. ”
Step 3: Generate drafts instead of final content
This tool is used to build skeleton, not muscle.
Ask for a summary based on the source pack. Ask for three different analogies to explain complex terms in the text. Ask them to condense a 10-page technical manual into a 1-page work aid.
Step 4: Fact checking and evidence tags
Before we can fix the flow, we need to fix the facts.
Traceability: If an AI makes a claim, can you find statements in the source pack to support it? Numbers and dates: AIs are notoriously bad at math and timelines. Please check all numbers manually. Links: Click all URLs. AI often generates invalid or invented links.
Step 5: Instructional Design QA
Once the facts are clear, let’s turn to the science of learning.
Cognitive load: Did the AI dump a wall of text? Break it. Bloom’s Taxonomy: Are quiz questions just testing memory (low level) or application (high level)? AI defaults to memory questions because they are easier to generate. Tone: Does it look like a robot? Infuses human warmth and empathy.
Step 6: Pilot and Iterate
Don’t announce it to the entire company. Send the module to 5 users. Watch them receive it. If you get stuck with an AI-generated explanation, it’s not clear enough. Please rewrite it yourself.
QA checklist
Perform these six checks before publishing your AI-generated content.
precision and sourcing
“Ctrl+F” test: Are all factual claims found in the original source document? Hallucination check: Ensure external statistics, dates, and regulations were not invented by AI. Link validation: Click all hyperlinks. Try to lead to live, relevant pages rather than dead ends.
Adjustment to goals
Fluff Filter: Did the AI add any “nice to know” history or background information? If it doesn’t support your learning objectives, remove it. Action-oriented: Does the content teach the learner how to perform the task, or only about the task? Audience match: Is the level of complexity appropriate? (e.g., don’t tell a software engineer “what a browser is”).
Validity of evaluation
Distraction Check: Is the wrong answer plausible in a multiple-choice question? AI often creates clearly stupid distractions that make quizzes too easy. Answer Key: Based on your policy, is the answer absolutely correct? Feedback: Did the AI generate helpful feedback as to why the answer is incorrect?
Cognitive load and clarity
Brevity: Are the paragraphs short (3-4 sentences)? AI tends to be redundant. Active voice: Did the AI use a passive voice (e.g. “The form must be signed”)? Change this to active voice (e.g. “Sign the form”). Formatting: Are lists used instead of dense blocks of text?
Accessibility basics
Alt text: If the AI suggests an image, is the description functional and understandable for a screen reader? Reading level: Is the language simple enough? (Aim for a grade 8 reading level for general compliance). Contrast: If AI generates a slide layout, is the text legible against the background?
Tone, inclusion, policy, and compliance
Bias Scan: Check pronouns and roles. Has AI made the manager a “he” and the assistant a “she”? Brand voice: Does it sound like a robot or a human? Add warmth and empathy as needed. Safety and legality: Make sure there are no absolute promises (e.g., “If you follow this, you’ll never get hurt”) that could create liability.
Two mini examples
Example A: SME transcript
Context: There is a 45-minute recording of a product manager rambling about a new software feature. Bad way: Paste everything and say “write the script”. Result: AI includes product manager complaints against engineering teams, and critical login steps are missed. Safe method: Clean: Manually remove complaints from your record. Prompt: “Act as a technical writer. List the step-by-step login process based solely on the attached transcript. Format it as a numbered list.” QA: Validate the steps against a real software sandbox. I realized that the AI forgot to click “Save”. Add it manually.
Example B: Quiz generator
Context: Code of Conduct course quiz required. Bad method: “Write 5 difficult questions about ethics” Result: AI asks philosophical questions such as “What is the nature of truth?” It has nothing to do with company policy. Safe Method: Prompt: “Use the attached ‘Gifts and Entertainment Policy’ PDF to write three scenario-based multiple-choice questions. Learners must decide whether they can accept a gift. For all correct answers, quote the specific clause from the PDF.” QA: Check the clauses. Make the scenario feel realistic rather than cartoonish.
Measurement: Did it work?
Creation speed is a vanity metric. Effectiveness must be measured.
what to track
Quiz reliability: See the analysis. If 100% of your learners answered question 3 correctly, it is too easy. If 0% is correct, the AI wrote a confusing question or the content doesn’t cover it. Confidence score: Ask learners, “How confident are you in applying this skill?” If confidence is low, the AI-generated content may be too abstract.
A/B testing
If you want to prove that this works, please run a test.
Group A: Take old legacy courses created by humans. Group B: Take the new AI-assisted (and human-verified) course.
Compare time and proficiency. If Group B learns the same amount in half the time, the workflow is successful.
At the end
AI is a tool like a calculator or a spell checker. You cannot submit a financial report without double-checking the calculator entries. Don’t publish your training without double-checking the AI output. The goal is not to let AI do the work. The goal is to have AI do the grunt work (summarizing, formatting, drafting) so you can focus on high-value work like strategy, context, and relationships.
