Creating Arabic and English courses with AI: a practical, responsible workflow

Start from your material, not from a blank prompt

AI is fast at producing course-shaped text. The risk is that it fills gaps with confident, generic content that is not your policy, your product or your regulator’s rules. The fix is grounding: give the model the source material, such as the manual, the policy, the recorded session or the existing deck, and ask it to build only from that. Anything the source doesn’t cover should come back as a question for you, not as an invented paragraph.

Before you upload, decide what the course is for. Write two or three outcomes in plain words: what should a learner be able to do afterwards? That single paragraph will shape the outline more than any prompt trick.

Review the outline before any lesson is written

The outline is the cheapest place to fix a course. Ask for modules, lessons and one line per lesson on what it teaches and where in the source it comes from. Then have a subject expert check three things: is anything missing that learners must know, is anything included that doesn’t belong, and is the order right for someone new? Only when the outline is approved should lessons be drafted. Fixing structure after forty lessons exist is slow and expensive.

Build a terminology glossary first

Bilingual courses drift when the same English term becomes three different Arabic words across lessons. Before drafting, agree a short glossary: each key term in English, the Arabic you will use, and terms you deliberately keep in English, such as product names. Give the glossary to the model with every request, and to every reviewer.

• Decide on register: Modern Standard Arabic for most training, with a lighter Gulf-friendly tone where it suits the audience
• Fix how you write numbers, dates and units, and use them the same way everywhere
• Keep regulator and programme names exactly as the official body writes them, in both languages

Write two versions, not a translation

Word-for-word translation produces Arabic that is technically correct and hard to read. Ask for a natural Arabic version of the same lesson, with the same outcomes, examples and quiz logic, but sentences built the way Arabic readers expect. Keep the two versions linked, so when one changes you know the other needs review. Quizzes need extra care: check that the correct answer is still correct in both languages, and that a distractor hasn’t become accidentally right in translation.

Watch for right-to-left pitfalls

Arabic text that looks fine in a document can break on screen. Most problems come from mixed-direction content: an English term, a number or a code inside an Arabic sentence. The W3C’s guidance is a good reference for whoever builds your pages: set direction on the page with markup rather than CSS, use logical start and end alignment instead of left and right, and isolate text of unknown direction so it doesn’t disturb the sentence around it.

• Punctuation that jumps to the wrong end of a line next to an English word
• Numbered steps, arrows and progress bars that still point left to right
• Screenshots and diagrams that show only the English interface
• Fonts without proper Arabic support, and text in images that can’t be edited or read aloud

Assessments and practice need extra care

AI writes plausible quiz questions quickly, and that is exactly the problem: a question can look right and test nothing useful. Ask for questions tied to a stated outcome, with the correct answer and a one-line explanation that points back to the source. Prefer scenarios over recall: “a customer asks for X; what do you do?” teaches more than “which article covers X?” Check that wrong options are believably wrong, not obviously silly, and that no question gives away another’s answer. For anything that leads to a certificate, have a person approve the final question bank in both languages.

A review workflow that holds up

AI shifts the work from writing to reviewing, so the review has to be real. A workable sequence: a subject expert approves the outline; the model drafts lessons from the source; the expert checks every factual claim against the source; an Arabic reviewer reads the Arabic version as a learner would, not side by side with the English; someone checks quizzes in both languages; and a final pass happens on the actual screen, on a phone, in both directions. Record who approved what and when. Frameworks such as NIST’s AI Risk Management Framework and its Generative AI Profile are useful if you want to formalise this.

How Classbase helps

Classbase’s course builder is designed around this workflow: it drafts from the files you upload, asks you to approve the outline before writing lessons, keeps Arabic and English versions of the same course linked, and shows every change before it lands. Your reviewers still make the final call.