The AI That Drafts Your Course, Not Replaces You

Every course platform is adding an AI button right now. Most of them are vague about what it actually does, evasive about whose API key it uses, and worryingly silent about what happens to the content it generates before a human sees it. That silence is the actual risk. An AI feature that can publish content without a person reviewing it first isn’t a productivity tool, it’s a liability with a friendly icon.

We’re going to look at exactly how Learnomy’s AI Authoring module is built, using its real settings screen and the actual policy decisions baked into the code, because the interesting part of an AI feature isn’t that it exists, every plugin has one now, it’s the specific guardrails around it.

This module was not enabled by default on the install we’re working from, it’s off unless a site owner deliberately turns it on, and its generation wizard won’t do anything at all until an AI provider is actually connected. Both of those facts turn out to matter more than any generation demo would: they’re the actual proof that this was built with restraint, not just speed, and restraint is precisely the quality that’s hardest to fake and easiest to skip when a team is racing to ship an AI feature before a competitor does.

Why AI in a course tool is a different risk category than AI in a chatbot

It’s worth being explicit about why the stakes here are higher than they might first appear. An AI feature bolted onto a customer-support widget that gives a slightly wrong answer produces an annoyed customer who asks a human instead. An AI feature bolted onto a course platform that generates a slightly wrong quiz question, and that question ships to students without review, produces something worse: a student memorizes an incorrect fact because it appeared in a graded assessment from a course they paid for and trusted. The failure doesn’t just annoy someone, it teaches them something false, under the authority of a credential-bearing course.

That’s the actual reason a draft-first rule matters more here than in most other AI feature categories, and it’s why evaluating an AI authoring tool by “does it write well” is asking the wrong question. The right question is “what happens to what it writes before a student ever sees it,” and that’s a question about workflow and defaults, not about model quality. A better model with no review step is still worse than a mediocre model that never skips review, because the review step is where a subject-matter expert catches the plausible-sounding wrong answer a model has no way of knowing is wrong.

The two questions that actually matter

Before evaluating any AI authoring feature, there are exactly two questions worth asking, and everything else is a distraction. First: whose AI account is this actually using, and who’s billed? Second: can AI-generated content become visible to a real student without a human looking at it first? A vendor that can’t answer both of these plainly, in writing, on a settings screen, hasn’t actually thought through what they built, they’ve bolted a chat interface onto a content editor and called it a feature.

Whose AI key is it, actually

Here’s the real AI Connection card, exactly as it renders in the Learnomy settings:

Read the description line carefully: “AI authoring uses the provider you connect once in WordPress core under Settings > Connectors. The key is stored by WordPress and shared across every plugin. Learnomy never stores it.”

This is a genuinely different architecture from most plugin-level AI features, and it matters more than it sounds like it should. Learnomy doesn’t ask you to paste an API key into its own settings screen, where it would be one more secret sitting in one more plugin’s database table, one more thing to rotate if that specific plugin ever has a vulnerability. It reads from WordPress’s own core AI Connector system, the same connection every other AI-aware plugin on your site can share. Connect Anthropic, Google, or OpenAI once, in WordPress’s own Settings, and every plugin built to use it, Learnomy included, draws from that single connection.

The practical result: on this install, exactly as shown, all three providers read Not connected, and the card says so honestly, “Not connected yet. Connect a provider to enable AI authoring.”, with a direct button to go set that up, rather than pretending the feature works when it structurally can’t yet. No AI action anywhere in Learnomy will do anything until that connection exists, which is precisely the point: the feature has a real, checkable off-switch that isn’t a Learnomy setting at all, it’s a WordPress-level decision you make once.

The safety model, in the site owner’s own words

Scroll down on that same settings page and you get the actual policy controls, not marketing copy, the literal checkboxes a site owner sets:

Enable AI authoring, off by default, with the description stating plainly: “When off, no AI actions appear for anyone.” Not “hidden from students” or “limited functionality”, no AI action appears, full stop, for any role, anywhere in the product, until this is deliberately turned on.

Who can use AI, Administrators, always, with no way to revoke it (they’re the ones who turned the feature on in the first place). Instructors, as a separate checkbox a site owner can leave unchecked, with a clarifying note: “Instructors author on the front-end builder; admins author in wp-admin.” That’s the multi-instructor system covered earlier in this series intersecting with AI authoring directly, a site owner running a marketplace of guest instructors can choose whether those instructors get AI assistance at all, independent of whether admins do.

Draft first, checked, and described as “Strongly recommended. AI never auto-publishes.” This is the single most important policy on the whole page, and it’s worth restating in plain terms: no matter who generates content with this feature, admin or instructor, that content lands as a draft. It does not go live. It does not reach a student. A human has to open it, read it, and decide to publish it, every single time. There is no configuration path that lets AI content skip that step.

What it actually generates, and where

The AI wizard shows up as a small “sparkles” trigger button in three specific, real places in the product, based directly on the actual button code: a “Generate with AI” button inside an existing course’s curriculum builder, a “Create with AI” button on the frontend “Create New Course” page for instructors who’d rather not start from a blank form, and a “Suggest questions” button inside the quiz editor, scoped to a specific quiz and topic.

Each of these opens the same kind of wizard: propose content, let the person review it, let them remove what they don’t want, and only then commit it, as a draft. The quiz-question flow is explicit about this in its own interface copy: “Remove any you do not want. The rest are created as draft questions and added to this quiz.” And after the fact, it’s equally explicit about what didn’t change: “New questions are added as drafts. Existing questions are never changed or removed.” An instructor using AI to pad out a question bank can’t accidentally have the AI silently rewrite a question that’s already been used in a graded attempt, new content only ever adds, it never edits or deletes anything that existed before the AI touched the quiz.

Usage limits, because someone is paying for every generation

Because this connects to an external AI provider billed at the account level, the settings page includes real usage controls, not just a feature toggle:

“The AI provider is billed to you, so set how much Instructors can use. Administrators are never limited.” Below that, a concrete number: Generations per Instructor / month, defaulting to 50, with 0 meaning unlimited. That default number is a deliberate choice, generous enough for an instructor to genuinely use the tool across several courses in a month, capped enough that one instructor generating hundreds of drafts overnight doesn’t produce a surprise on your AI provider’s bill at the end of the month.

There’s also a Model selector, with guidance built directly into the interface rather than left to a support article: “Authoring is text-only, so a lower-cost model is usually the right choice. A Haiku-class model is cheapest, a Sonnet-class model balances cost and quality. Avoid Opus-class models for routine content.” This is a genuinely useful default for a site owner who has no reason to know the cost tradeoffs between AI model tiers, the setting itself teaches you which one to pick and why, rather than presenting an opaque dropdown of model names with no guidance.

Why “drafts only” is the correct design, not a limitation

It would be easy to read the draft-first requirement as AI Authoring being deliberately hobbled, why generate content at all if it always needs a manual review step before anyone sees it? The answer is that the review step is exactly where the value of a human instructor still lives.

An AI model can propose a reasonable-sounding quiz question about a topic it has no way of knowing your specific course actually covered in the way the question assumes. It can draft a course outline that sounds structurally sound but skips a prerequisite your actual curriculum depends on. None of that makes the AI useless, it makes the first draft a genuinely useful starting point that still needs a person who actually knows the subject to check it before a paying student sees it. Auto-publishing would optimize for speed at the exact point where speed causes the most damage: the moment unreviewed content reaches someone who’s trusting your course to actually teach them something correctly.

This is also why the per-instructor usage cap and the draft-only rule work together rather than independently. A cap alone doesn’t stop bad content from reaching students, it just limits how much of it there could be. A draft-only rule alone doesn’t stop cost from spiraling, it just contains where the content ends up. Together, they mean: a bounded amount of AI-assisted drafting happens, billed predictably, and every single piece of it passes through a human before it becomes something a real student can act on.

Three ways this actually gets used

The solo creator staring at a blank course outline. For someone building their first course alone, the hardest part is rarely the subject matter, it’s structuring it. “Create with AI” on the frontend course-creation page exists specifically for this moment: propose an outline, get something concrete to react to instead of a blank form, and then edit it into shape using your own actual expertise. The AI’s job here isn’t to know your subject better than you do, it’s to turn a blank page into a first draft you can disagree with, which is a much easier starting point than nothing at all.

The instructor who’s great at teaching and slow at writing quiz questions. Writing a genuinely good multiple-choice question, one with plausible wrong answers that actually test understanding rather than trick the reader, takes real effort per question, and a course with a dozen lessons might need dozens of them. “Suggest questions,” scoped to a specific quiz and topic, produces a batch to react to instead of stare down one at a time. The instructor still decides which ones are good enough to keep, but they’re editing and selecting rather than generating from nothing, for every single question.

The academy owner managing instructors who have wildly different writing speeds. In a multi-instructor marketplace, some guest instructors are naturally fast writers and some are subject experts who dread the writing part of course creation. Enabling AI Authoring selectively for instructors, with a shared monthly cap, means the academy owner can offer this as real support to the instructors who need it, without opening an unlimited AI bill or handing out a feature nobody’s tracking the cost of. The “Instructors” checkbox next to “Administrators (always)” is exactly the lever for this, on for the marketplace as a whole, with the cap as the actual cost control.

Common questions before you enable this

Does the AI ever see or use my students’ data? Based on what the settings describe, AI Authoring is scoped to content generation, course outlines, quiz questions, using the topic and context an instructor provides at generation time. It’s not described as a system that reads existing student submissions, grades, or personal information to generate anything. If that distinction matters for your compliance posture, it’s worth confirming directly against your specific AI provider’s own data-handling terms, since the provider, not Learnomy, is what actually processes the request.

What happens if I disable the connected AI provider after content has already been generated? Disconnecting the provider stops new generations; it has no reason to affect drafts or published content that already exists, since those are stored as ordinary Learnomy content once created; the AI provider is only in the loop at generation time, not afterward.

Can an instructor generate unlimited content by making many small requests instead of one big one? The monthly cap is described as “Generations per Instructor,” which reads as counting each generation action, not the size of what came back. A dozen small requests would count as a dozen generations against the same 50-per-month default, so this doesn’t look like a loophole in the cap’s design, worth confirming with a real test if a specific instructor’s usage pattern looks unusual.

Is the generated content flagged as AI-written anywhere a student would see? Nothing in the settings screen mentions a visible AI disclosure on published content, which lines up with the framing here: once an instructor has reviewed, edited, and published a draft, it’s their content, the same as anything else they wrote by hand. The “draft first” rule is the control point, what a student ultimately sees already passed through a human’s judgment, regardless of what fraction of the first pass came from a model.

Do I need a paid plan with a specific AI provider to use this at all? The settings screen lists Anthropic, Google, and OpenAI as connectable providers via WordPress’s own core Connectors system, each with a direct “Get API key” link. Which one you pick, and what it costs, is a decision made at the WordPress-core level and with that provider directly, Learnomy’s role is consuming whichever one you’ve already connected, not brokering the relationship or adding its own markup on top.

What a careless version of this feature would look like

It’s worth spelling out the alternative, because “AI-assisted authoring” as a phrase covers a huge range of actual implementations, from genuinely careful to genuinely reckless, and the phrase alone tells you nothing about which one you’re getting.

A careless version stores its own API key in its own settings screen, meaning you’re now managing a second AI credential specifically for this one plugin, with no visibility into whether it’s being reused, logged, or rate-limited sensibly. A careless version defaults to on, so a site upgrading the plugin discovers AI-generated content has already started appearing somewhere, without anyone having made an active decision to enable it. A careless version publishes what it generates directly, because “one click and it’s done” demos better than “one click and now review a draft”, treating the review step as friction to be removed rather than the actual safety mechanism it is. A careless version has no usage cap at all, because caps require the vendor to think through billing edge cases, and it’s easier to ship without them and let the site owner discover the problem on their AI provider’s invoice at the end of the month.

None of those choices are hypothetical failure modes invented for this article, they’re the default outcomes of optimizing an AI feature for “looks impressive in a two-minute demo” over “safe to hand to instructors I don’t personally supervise.” Every one of them is also the exact opposite of a specific, verified behavior on the real settings screen walked through above: off by default, using a connection you already control, draft-only with no override, and a real numeric cap with sane guidance on model selection. The contrast isn’t subtle once you know what to actually check.

Reading the actual code comments, because they tell you what the team was worried about

One more piece of evidence worth surfacing, because it’s the kind of detail that never appears in a features page but tells you exactly what a team was thinking about when they built this. The extension’s own internal description of itself, pulled directly from its metadata, reads: “AI-assisted course, lesson, and question authoring on top of WordPress core AI Connectors. Admin + Instructor, draft-first.” That’s four words doing a lot of work, “draft-first” isn’t a marketing phrase, it’s the module’s own one-line self-description of its core safety property, meaning the team building it considered that property important enough to be the defining characteristic of the whole feature, not an afterthought bolted on after an early version shipped without it.

Similarly, the quiz-question generation flow’s own interface copy, “New questions are added as drafts. Existing questions are never changed or removed”, reads like a sentence written by someone who had already imagined the specific bad outcome (an AI silently rewriting a question a class had already been graded against) and wrote the rule to close that exact gap. That’s a different kind of engineering than a generic “AI review recommended” disclaimer copied across a dozen different generation features without much thought about what could specifically go wrong in each one.

What to check before you turn this on for your own instructors

If you’re evaluating whether to enable AI Authoring, or any AI writing feature in any course platform, these are the specific things worth verifying directly, based on exactly what this settings screen makes checkable:

• Confirm which AI account is actually being billed. If a plugin asks you to paste a separate API key into its own settings rather than using a connection you already control elsewhere, that’s a second secret to manage and a second bill to track, worth knowing before you enable anything.
• Confirm draft-only is enforced, not just recommended. A checkbox labeled “recommended” that a user could uncheck is a much weaker guarantee than a rule the code enforces unconditionally. Ask directly, or read the description text closely, Learnomy’s copy (“AI never auto-publishes”) describes a hard rule, not a preference.
• Decide who gets access before you announce the feature. If you’re running a multi-instructor marketplace, deciding up front whether guest instructors get AI assistance, and whether that’s a perk you use to attract better instructors, or a cost you’d rather control tightly at first, is a five-minute settings decision that’s much easier to make before instructors are already asking for it.
• Set the usage cap deliberately, not at the default. Fifty generations a month per instructor might be exactly right, or it might be far more or less than what your specific catalog needs, the number only means something once you know how many instructors you have and how actively they’re building.

That checklist applies to any platform, not just this one. Our comparison of the best LMS platforms with AI features puts the same questions to every major plugin’s AI implementation.

The honest pitch

An AI authoring feature earns trust the same way every other feature in this series has: by being checkable, not just claimed. “AI-assisted” on a features page could mean anything from a genuinely careful draft-first workflow to a chatbot that publishes whatever it generates the moment you click a button. The difference between those two is not a detail, it’s the entire question of whether the feature is safe to hand to people you don’t personally supervise.

What we found on this install is the careful version: a connection that uses infrastructure you already control, a policy that’s off by default and has to be deliberately turned on, a hard rule against auto-publishing that the interface states in plain language, and usage limits that acknowledge someone is paying for every generation. None of that is exciting the way a slick generation demo is exciting. It’s the boring, correct engineering that makes the exciting part actually safe to hand to real instructors, on a real site, with real students on the other end of whatever gets published.

Why this design choice is the actual signal to watch for

Step back from AI Authoring specifically and consider what this module reveals about how the rest of the plugin was likely built. A team that defaults a new, trendy feature to off, routes billing through infrastructure the site owner already controls rather than adding a new key to manage, and hard-codes a no-auto-publish rule into the interface copy rather than leaving it as a suggestion, is a team optimizing for the site owner’s actual risk exposure over the speed of a product announcement. That’s a meaningfully different set of priorities than shipping a flashy “AI-powered” badge on a pricing page and working out the safety details later, in response to the first incident.

The test that actually matters here isn’t whether the AI generates good quiz questions, most current AI models can do that reasonably well, and that capability is table stakes at this point, not a differentiator. The test is what happens at the edges: what’s the default state, who’s billed, what happens to content nobody reviewed, what happens when the cap is hit, what happens if the connection breaks mid-generation. Every one of those edge questions has a specific, checkable answer on the settings screen walked through above, and every answer favors caution over convenience. That’s the actual case for trusting a feature like this with real instructors and real students, not the quality of a single generated paragraph, but the discipline visible in how the feature behaves when something isn’t a clean, happy-path request.