Why I Stopped Trusting One AI Voice for eLearning
Every instructional designer working with generative AI eventually runs into the same discovery: AI tools don't all write the same way. That difference is a design resource, not a limitation to work around.
I learned this the direct way. I shared a piece of learner-facing content with a colleague for review. Within seconds she said, "oh, this was made with a chatbot." Nothing was factually wrong with it. The structure was fine. But the phrasing gave it away. For example, there were overly balanced sentences and tidy transitions. Moreover, every paragraph wrapped itself up a little too neatly. She wasn't reacting to the content. She was reacting to the voice. The voice read as a tool, not as me.
As an educator and a human being, I don't naturally write like that. I don't reach for em dashes, or string words like "enhance" and "utilize" into every sentence, or organize my thinking into that particular AI cadence. So when I looked at my own content through my end learner's eyes, I saw something uncomfortable: it could read as if I hadn't spent real time on it. If the writing feels generic enough that a learner can tell it took me minutes instead of care, it doesn't just look impersonal — it can feel like they weren't worth the effort. And learners who sense that tend not to want to connect with the content at all.
Was I wrong to use AI? No. AI brings real clarity to how information gets organized and delivered, and that's valuable. But was the final product focused on human speech? Also no. It carried too many AI markers, and the natural rhythm of speech was broken in a way that was obvious to the person reading it. That gap was the signal that I needed to dig deeper — not abandon AI, but get more deliberate about how I used it.
That moment sent me hunting for my own AI voice: one that could hold up to a close read and still sound like a person who understands the learner, not like software that generated a correct answer. It's not enough for a course or a resource to be accurate and well organized. If it sounds like it came out of a chatbot, learners notice, and that recognition creates distance right when you want connection.
The hunt turned up something I didn't expect. I didn't need one AI voice. I needed two. That's what led me to split my instructional design workflow across two different AI tools. This wasn't because one is "better" than the other. Instead, their default writing styles serve different stages of building effective prompts for eLearning content.
Two Tools, Two Registers
One of my tools tends toward a structured, procedural writing style: numbered steps, bolded headers, explicit "Step 1, Step 2" scaffolding. That register can work against you when it shows up in learner-facing eLearning content. However, it's exactly what's useful when architecting the instructions behind an AI tool. For example, ask it to help build a prompt framework. It typically returns something organized, sectioned, and easy to audit.
My other tool reads differently. Its default output tends to be more fluid and conversational. It generally follows natural-language instructions without needing rigid scaffolding to stay on task. For learners, that distinction matters. Content built to sound conversational rather than templated is less likely to carry the AI writing patterns. For instance, formulaic openers and repetitive phrasing can create a quiet barrier between the material and the learner.
The workflow follows each tool's strengths: the structured tool scaffolds the instruction set, and the conversational tool executes against it to produce natural-sounding eLearning content.
Building a Custom Project With the Scaffolding Tool: A Practical Handoff
Writing custom instructions for a persistent AI "project" or workspace is itself an instructional design task. The structured, procedural tool is a genuinely useful drafting partner for it. However, this works only as long as the handoff between tools is deliberate.
- Draft the structure in the scaffolding tool, then remove the meta-commentary. This kind of tool often narrates its own process ("As an AI, I will now outline..."). That framing adds nothing useful to your project instructions and can nudge the output toward the same stilted, AI-sounding phrasing instructional designers are trying to avoid. Cut it before adding the draft to your project.
- Name the voice you want, not just the content structure. A structural outline tells the receiving tool what to cover. It doesn't tell it how the content should sound. Add an explicit instruction inside the project — not a one-off request — asking it to avoid generic AI phrasing and over-signposted lists, and to write the way a knowledgeable colleague would explain something.
- Trim unnecessary formatting. Drafts from the structured tool often carry more headers, bullets, and bold text than the task actually needs. If the goal is natural prose in the final eLearning content, over-formatted input works against that goal. Simplify it before it becomes the project's default output style.
Why AI Prompt Design Matters for Learning Design
For instructional designers and learning experience designers, voice isn't a stylistic footnote. It's an accessibility and engagement factor. Training content that reads as generic or visibly AI-generated creates friction for learners. This is particularly true across multilingual or global teams, where natural phrasing plays a real role in comprehension.
Good learning content should also speak in the learner's own language — not just literally, but in tone and rhythm — so learners recognize themselves in it rather than feeling talked at by a system. When the writing sounds human, it can do more than deliver information. It can function the way a good mentor or coach would: present, attentive, and speaking with the learner rather than at them. That's harder to achieve when every sentence carries the same flattened, generic AI cadence, and it's part of why getting the voice right is worth the extra step in the workflow.
Using one AI tool to architect the instruction logic and another to generate the learner-facing content isn't a workaround. In fact, it's a deliberate application of each tool's actual strengths. This is also a reminder that in AI-assisted learning design, the prompt itself is as much a design artifact as the course it produces.
