> Original data from 137,000+ AI prompts and 26,000+ courses: who creates courses with AI, how they brief it, what they build and how learners perform.

---

# The State of AI Course Creation 2026

What 137,000+ AI prompts and 26,000+ courses reveal about who creates courses with AI, how they brief it and what learners get out of it.

_Updated 9 November 2026_

- **37,284** course creators analysed
- **21** words in the typical AI brief
- **54%** write in a language other than English
- **26%** build job skills, the top category
## Who creates courses with AI

Course creators, teachers, corporate learning teams and training providers all use AI to create courses, for very different reasons. To understand who they are and what they build, we analysed the first prompt of 37,284 people.

### Four groups create most courses with AI, each with a different goal

Among people creating courses for others, course creators and coaches account for 31%, teachers and lecturers 28%, corporate learning and HR teams 24%, and training providers and consultants 16%. A further one in five people whose role is clear use the same tools to build study material for their own learning.

Each group arrives with a different goal. Corporate teams build job skills, compliance and onboarding training. Teachers build lessons for their subjects. Creators build courses to sell.

_Who is creating the course, among people building for others_

| Category | Value |
| --- | --- |
| Course creators and coaches | 31% |
| Teachers and lecturers | 28% |
| Corporate L&D and HR | 24% |
| Training providers | 16% |

Base: 2,043 learning-related first prompts where the creator is identifiable and is building for others, from 10,000 classified prompts, May to Nov 2026

**What this means for training teams:** Design your process for the group you are in. A corporate team needs review and sign-off steps; a creator needs a clear promise to the buyer; a teacher needs alignment to the curriculum. The AI draft is the same starting point for all of them, but the finishing work is not.

### Job skills, not compliance, are what most people build with AI

Job and technical skills are the largest category at 26% of learning-related prompts, followed by academic subjects at 18% and personal development at 9%. Compliance and safety training accounts for 6%, and onboarding for under 2%.

The pattern suggests AI is being used to fill the long tail of practical, role-specific skills that rarely got a formal course before, rather than only to speed up the mandatory training every organisation already has.

_What the course is for_

| Category | Value |
| --- | --- |
| Job and technical skills | 25.7% |
| Academic subjects | 18.1% |
| Personal development | 9.3% |
| Business and earning income | 7.3% |
| Compliance and safety | 6.3% |
| Soft skills and leadership | 4.1% |
| Languages | 4% |
| Exam preparation | 2.3% |
| Onboarding | 1.5% |

Base: AI-classified random sample of 7,490 learning-related first prompts (from 10,000 classified), May to Nov 2026. 21% had no identifiable purpose

**What this means for training teams:** Look beyond the compliance catalogue. The highest-value use of AI course creation is often the specialist know-how that lives in a few experts' heads: a system, a procedure, a product. Those are often the courses there was never the time or budget to build.

### Education and technology lead, but every sector is represented

K-12 education (12%) and technology (11%) are the largest sectors, followed by business and entrepreneurship (7%), higher education (6%) and healthcare (5%). Finance, construction and trades, manufacturing and the public sector each account for 2 to 3%.

The spread matters more than the ranking. The same tools are being used for a primary school science unit, a hydraulic maintenance certification and a pharmacy dermocosmetics course.

_Sector of the course_

| Category | Value |
| --- | --- |
| K-12 education | 11.9% |
| Technology and IT | 10.8% |
| Business and entrepreneurship | 6.8% |
| Cross-industry | 5.8% |
| Higher education | 5.6% |
| Healthcare | 5.2% |
| Beauty, wellness and fitness | 3.9% |
| Finance and insurance | 3.4% |
| Professional services | 3.4% |
| Creative arts and media | 3.2% |

Base: AI-classified random sample of 7,490 learning-related first prompts (from 10,000 classified), May to Nov 2026. 22% had no identifiable sector

**What this means for training teams:** Do not wait for an industry-specific tool. If your sector is specialised, build from your own documents and standards so the course reflects your field, rather than relying on the AI's general knowledge.

### Each language is a different market

Portuguese-language prompts lean towards independent creators: course creators and coaches account for 17% of Portuguese prompts, more than twice their share in English, and courses about earning income (14%) and personal development (15%) are about twice as common as in English. English has the most corporate mix, with learning teams and training providers making up 13% of prompts.

_Share of learning prompts by creator type, by language_

| Category | Course creators and coaches | Corporate L&D and training providers |
| --- | --- | --- |
| English | 7.8% | 13.1% |
| Portuguese | 17.3% | 8.6% |
| Spanish | 4.9% | 7.3% |

Base: 3,014 English, 1,464 Portuguese and 1,415 Spanish learning-related first prompts from 10,000 classified, May to Nov 2026

**What this means for training teams:** If you train or sell across languages, adapt the offer as well as the words. Who creates courses, and why, differs by language, so a course that works in one market may need a different angle in another.

## How people brief the AI

The first prompt works as the brief for the whole course. Many first prompts are short, and the design detail they leave out is the clearest opportunity this report found.

### The typical brief is 21 words long

The median first prompt is 21 words, about the length of a sentence. 44% of first prompts are under 100 characters, while 20% run past 500 characters, often because the person pasted in their own outline or content.

Most people then keep talking to the AI: 76% send more than one prompt before their course is generated, using the conversation to fill in what the first brief left out.

_Length of the first prompt_

| Category | Value |
| --- | --- |
| Under 100 characters | 44% |
| 100 to 500 characters | 36% |
| Over 500 characters | 20% |

Base: first prompts from 37,284 people, May to Nov 2026

**What this means for training teams:** Write the brief before you open the tool. A few sentences covering who the learners are, what they should be able to do afterwards and how long they have will save several rounds of correction.

### Nearly half of briefs leave out every key design decision

We checked each learning-related prompt for five decisions an instructional designer would ask about: the audience, their level, the length of the course, the learning outcomes and how learning will be assessed. 47% of prompts mention none of them. Only 13% mention three or more.

The audience is the detail people include most often (31%). Learning outcomes (17%) and assessment (16%), the two that most shape a course, appear in about one prompt in six.

_What first prompts specify_

| Category | Value |
| --- | --- |
| The audience | 31.2% |
| Tone or style | 24.9% |
| The level | 19.2% |
| Learning outcomes | 16.8% |
| Course length | 16.1% |
| Assessment | 16.1% |
| Number of modules | 14.1% |

Base: AI-classified random sample of 5,619 learning-related first prompts, May to Nov 2026

**What this means for training teams:** Brief the AI the way you would brief an instructional designer. At minimum, state who the course is for, what they should be able to do by the end and how you will check. Those three lines change the structure of what the AI produces, not just the wording.

### Corporate teams and training providers write the most complete briefs

Corporate learning teams and training providers leave out every key design decision in only one brief in five, against nearly half overall. Corporate teams name the audience most often (70%), and training providers (45%) and corporate teams (44%) are the most likely to give the AI their own source material. Course creators state learning outcomes most often (44%) and are the least likely to provide source material (17%).

_Share of prompts that include each element, by builder_

| Category | Names the audience | Gives source material | States outcomes |
| --- | --- | --- | --- |
| Corporate L&D | 70% | 44% | 27% |
| Training providers | 59% | 45% | 35% |
| Teachers | 60% | 37% | 20% |
| Course creators | 46% | 17% | 44% |

Base: 1,289 learning-related first prompts from identifiable creators building for others, from 7,599 classified prompts, May to Nov 2026

**What this means for training teams:** Borrow habits from the groups that brief most completely. If you are a creator, add your own material: your method, examples and stories are what make a paid course worth buying. If you are in a corporate team, add outcomes and assessment to the audience and documents you already provide.

## How courses are built

AI has removed the blank page from course design, but it has not removed the design decisions. The data shows clear norms emerging for course length and structure, and a measurable difference between courses built from existing material and courses built from a prompt alone.

### Most AI-built courses settle at six to nine modules

The median course has 8 modules and about 28 pages. Half of all courses fall between 6 and 9 modules, and only one in ten runs past 12.

That range is long enough to cover a topic and short enough for a learner to finish in a few sittings.

_Share of courses by number of modules_

| Category | Value |
| --- | --- |
| 1 to 3 | 10% |
| 4 to 6 | 26% |
| 7 to 9 | 39% |
| 10 to 12 | 15% |
| 13 or more | 10% |

Base: 26,266 course generations by 13,688 creators, Aug to Oct 2026

**What this means for training teams:** Use 6 to 9 modules as a starting point. If an outline runs past 12 modules, check whether it would work better as two courses, which also gives learners a finish line sooner.

### One in three courses starts from an existing document, and those courses go deeper

A third of creators begin by uploading material they already have: a policy, a slide deck, a manual or a PDF. The other two thirds type a prompt and let the AI work from general knowledge.

The difference shows up in the outline. Courses built from a document have a median of 8 modules, against 6 for prompt-only courses. Source material gives the AI specific content to organise, so it covers more ground and stays closer to how the organisation actually works.

_How courses start_

| Category | Value |
| --- | --- |
| From a typed prompt only | 67% |
| From an uploaded document | 33% |

Base: 26,266 course generations, Aug to Oct 2026

**What this means for training teams:** If the knowledge already exists in a document, start there. A prompt is fine for general topics, but company-specific training should be built from the source so the course reflects your own procedures, not generic ones.

## After the first draft

The first draft now arrives in minutes. What people do next shows where the design effort has moved: from writing content to enriching, checking and shaping it.

### New users generate their first course a median of five minutes after signing up

The middle half of new users produce a first course between 3 and 10 minutes after creating an account, and 40% of all new signups build a course within their first week.

Speed to a first draft is now measured in minutes. The time that remains, and where quality is won or lost, is in reviewing that draft against what learners actually need.

_Minutes from signup to first course_

| Category | Value |
| --- | --- |
| Fastest quarter | 3 |
| Median | 5 |
| Slowest quarter | 10 |

Base: 11,566 first courses from 28,542 new signups, Aug to Oct 2026

**What this means for training teams:** Spend the time saved on drafting on review: checking accuracy, tightening examples and testing the course with a few real learners.

### After the draft, people add to it far more than they cut

When creators ask the AI to change their course, the most common requests are to add new content (17%), add images (14%), add a quiz (10%) or add examples and scenarios (8%). Requests to shorten (1.5%) or simplify (2%) are rare.

12% of requests are questions rather than instructions, such as how to do something in the course or whether the AI can change a section. Editing has become a conversation.

_What people ask the AI to change after the first draft_

| Category | Value |
| --- | --- |
| Add new content | 17% |
| Add images | 14% |
| Ask the AI a question | 12% |
| Add a quiz | 9.5% |
| Rewrite a section | 9% |
| Change format or layout | 9% |
| Add examples or scenarios | 8% |
| Add interactive blocks | 6% |
| Restructure | 6% |
| Fix errors | 4% |
| Simplify | 2% |
| Shorten | 1.5% |

Base: 800 editor AI requests from 2,260 people (max 3 per person), Apr to Nov 2026. A request can include several changes

**What this means for training teams:** Before adding more to a draft, check whether a section could be shorter or simpler. Few people ask the AI for either, though both often help learners.

### Video is the content people add most once the draft is done

When creators edit their AI draft, 48% add a video, 32% an image and 28% a quiz. Interactive elements such as assignments (16%) and flip cards (14%) are used by a smaller but committed group.

The pattern suggests creators treat the AI draft as the structure and text, then add the media and practice that make it engaging.

_Share of editors who added each type of content_

| Category | Value |
| --- | --- |
| Video | 48% |
| Image | 32% |
| Quiz | 28% |
| Notes and text | 23% |
| Assignment | 16% |
| Flip cards | 14% |

Base: 2,860 editors, Aug to Oct 2026

**What this means for training teams:** Plan for practice, not just content. A quiz or assignment in every module gives learners a way to check their understanding, and gives you evidence that the training worked.

## The languages of course creation

English is still the largest single language, but nearly half of all courses are now created directly in another language rather than translated afterwards.

### 45% of courses are built in a language other than English

Portuguese (15%) and Spanish (14%) together account for nearly three in ten courses, and Brazilian Portuguese alone represents 13% of everything built. French, Arabic and German follow, and 18 languages each produced at least 30 courses in the quarter.

This matters because training written in the learner's working language reduces the cost and delay of translating an English master course.

_Share of courses by course language_

| Category | Value |
| --- | --- |
| English | 55% |
| Portuguese | 15% |
| Spanish | 14% |
| French | 6% |
| Arabic | 4% |
| German | 1.3% |
| Italian | 1% |
| Chinese | 0.8% |
| Dutch | 0.7% |
| Russian | 0.6% |

Base: 25,607 courses with a recorded language, Aug to Oct 2026

**What this means for training teams:** If your workforce or customers work in more than one language, build each version natively rather than translating a single English course. It is now as fast to create in Portuguese or Spanish as in English.

## How learners perform

Learner results vary far more between courses than course design does. These figures use course-level medians, so a handful of very large training portals do not dominate the picture.

### Learners pass three in four quiz attempts in the typical course

The median course has a quiz pass rate of 76%. A course at the 25th percentile passes 62% of attempts, and one at the 75th percentile passes 91%.

Both ends of that range are worth investigating. Outside mastery-based courses, a pass rate above 90% can mean questions are too easy to tell you much, while a rate well below 60% usually points to unclear content or ambiguous questions rather than weak learners.

_Quiz pass rate by course_

| Category | Value |
| --- | --- |
| 25th percentile | 62% |
| Median | 76% |
| 75th percentile | 91% |

Base: 131,527 graded attempts across 242 courses with 10 or more graded learners, Aug to Oct 2026

**What this means for training teams:** Aim for a first-attempt pass rate between 70% and 90%. Review any question that almost everyone gets right or most people get wrong; it is usually the question, not the learner, that needs fixing.

### The median course is finished by 46% of the learners who start it

Completion is the most variable number in this report. The top quarter of courses are finished by 78% of learners or more, while the bottom quarter are finished by 6% or fewer.

A gap that wide is rarely explained by content alone. Whether the course is required, whether managers follow up, and how long it takes to finish all shape the result.

_Course completion rate by course_

| Category | Value |
| --- | --- |
| 25th percentile | 6% |
| Median | 46% |
| 75th percentile | 78% |

Base: 6,558 learners in 214 courses with 10 or more learners, Aug to Oct 2026

**What this means for training teams:** Benchmark completion against courses with a similar audience, not against all courses. For optional training, shorter courses, clear deadlines and a reminder from a manager do more for completion than extra content.

### Nearly a third of course building and learning happens on a phone

32% of creators generated a course on a mobile device, and 29% of learners who completed a course did so on a phone. Mobile is no longer an edge case for either side of training.

_Device used_

| Category | Phone | Computer or tablet |
| --- | --- | --- |
| Creators building | 32% | 68% |
| Learners completing | 29% | 71% |

Base: 13,688 creators and 11,782 learners who completed a course, Aug to Oct 2026

**What this means for training teams:** Check every course on a phone before you publish it. Short pages, readable images and tap-friendly quizzes matter for roughly a third of your audience.

## Where courses go

A course is only useful once it reaches learners. Export choices show how AI-built training moves into the systems organisations already run.

### PDF leads exports, and SCORM 2004 has overtaken SCORM 1.2

Over half of users who export a course take a PDF (53%). Among those loading courses into another learning management system, SCORM 2004 (16%) is now used slightly more than the older SCORM 1.2 (14%), while LTI remains small at 5%. Users can export in more than one format, so shares add up to more than 100%.

_Share of exporting users by format_

| Category | Value |
| --- | --- |
| PDF | 53% |
| HTML download | 21% |
| iFrame embed | 20% |
| SCORM 2004 | 16% |
| SCORM 1.2 | 14% |
| Zapier | 14% |
| LTI | 5% |

Base: 1,317 users who exported a course, Aug to Oct 2026

**What this means for training teams:** If your LMS supports SCORM 2004, use it: it reports completion and success separately and supports richer tracking than 1.2. Keep a PDF version for learners who need an offline or printable reference.
## Go deeper into the data

- [How people brief AI to build a course: 37,284 prompts](https://www.coursebox.ai/blog/how-people-brief-ai-to-build-a-course)
- [L&D teams vs course creators: how each briefs AI](https://www.coursebox.ai/blog/l-and-d-teams-vs-course-creators-ai-briefs)
- [How Brazil creates courses with AI: 6,821 people](https://www.coursebox.ai/blog/how-brazil-creates-courses-with-ai)
- [What people change after AI drafts a course](https://www.coursebox.ai/blog/what-people-change-after-ai-drafts-a-course)
- [How hard are AI course quizzes? 131,527 attempts](https://www.coursebox.ai/blog/how-hard-are-online-course-quizzes)
## Methodology and data

Aggregate statistics on how courses are created with AI on Coursebox in 2026: who builds courses, what they build, how they brief the AI, course length, source material, course language, editing requests, learner quiz pass rates and completion, device use and export formats. Based on product data from May to November 2026.

- Data period: 2026-05-14 to 2026-11-09
- Sample: 137,126 AI prompts from 37,284 people; 26,266 courses; 131,527 graded quiz attempts
- Last updated: 2026-11-09
- Licence: CC BY 4.0

**Data sources:** Anonymous product data from Coursebox, an AI course creator and LMS. Prompt analysis uses 137,126 course-creation prompts from 37,284 people (14 May to 9 Nov 2026) and 37,908 editor AI requests from 2,260 people (Apr to Nov 2026). Course, learner and export figures use product events from 1 Aug to 31 Oct 2026.

**How prompts were classified:** Length, language, device and file use are measured on every person's first prompt. Sector, purpose, builder type, intended learner and which design decisions a prompt specifies come from an AI classification of a stratified random sample of 10,000 first prompts written in people's own words (repeated template text removed) and 800 editor requests (at most three per person). Two AI models were used; 304 prompts were classified by both, and the disagreements were checked by hand against the classification rules. Design-decision figures use the 7,599 prompts labelled by the model that matched the hand check most closely. A manual check of 40 random classifications from each model found about 92% correct on the main categories.

**Exclusions:** Bot and automated traffic, Coursebox staff and contractor accounts, test accounts and local development traffic are removed from every figure.

**Definitions:** Learning-related prompts are first prompts that ask for learning material (a course, lesson, study aid, assessment or teaching material): 7,490 of the 10,000 sampled. Modules are the top-level sections of a generated course. Pass rate is the share of graded quiz attempts marked as passed. Completion rate is the share of learners who opened a course and then completed it, excluding the course creator. Course-level figures are medians across courses.

**Privacy and minimum sample:** Only aggregate counts, shares and medians are published. No prompt text, names or company details appear in this report; example courses are paraphrased. Any group smaller than 30 is left out. The underlying data is not published.

**Limitations:** The data describes Coursebox users, not every course in the world. Builder type is identifiable in about a third of prompts. Sample shares carry a margin of about plus or minus 1 to 2 points at 95% confidence, more for subgroups. Prompt text is only available from May 2026. Course generation tracking had a gap from 23 to 27 Oct 2026.
## Questions about the report

Yes. The report is published under CC BY 4.0, so you can quote, chart or republish any figure as long as you credit Coursebox and link to this page. Each finding has its own link you can cite directly. Can I use these statistics in my article or report? Among people creating courses for others, course creators and coaches account for 31%, teachers and lecturers 28%, corporate learning and HR teams 24% and training providers 16%. Job and technical skills are the most common subject, at 26% of learning-related prompts. Who creates courses with AI? Treat it as a design brief. State who the course is for and their level, what learners should be able to do by the end, how long the course should be and how learning will be checked. 47% of prompts in this report include none of those, and adding them changes the structure of the course, not just the wording. How should I brief AI to create a course? In this report the median AI-built course has 8 modules and about 28 pages, and half of all courses have 6 to 9 modules. Courses built from an uploaded document run longer than prompt-only courses: 8 modules against 6. How long is the average online course? No. The report only publishes aggregate counts, shares and medians. No prompt text, course content, names or company details are published, example courses are paraphrased, and any group smaller than 30 is left out. Is any customer or learner data exposed?
## Build a course from your own material

Upload a document or type a topic and Coursebox drafts the full course, with quizzes, in minutes.

- Free plan, no time limit
- Unlimited learners
- SCORM and LTI export