AI eLearning Translation vs Professional Translation: How to Choose
AI eLearning translation is fast but uneven. Compare raw AI, AI plus human post-editing and professional translation, with research and tool limits.
Get a Free Quotation
Scoped proposal delivered within 48 hours
AI or professional translation in short
AI eLearning translation is fast and cheap enough for drafts and low-risk internal courses, but its accuracy varies by language and errors add up across a full course. For most training, the safer route is AI translation followed by full human post-editing. Keep professional human translation for compliance, safety and other high-stakes content.
Need your courses translated?
Get a Free Quote →This guide is for L&D, HR and compliance teams that need courses in more than one language. It explains the three routes you can take, what the research says about machine translation of instructional text, what AI tools miss inside an eLearning course, how voiceover fits in, and how to choose a route for each course.
Key takeaways
- There are three routes: raw AI translation, AI translation plus human post-editing, and professional human translation.
- Studies of machine-translated instructions show high sentence-level accuracy in some languages and much lower accuracy in others.
- High sentence accuracy still leaves errors in many full documents, so a whole course needs a human check.
- Built-in AI translation in authoring tools skips some content, such as text inside images and attached PDFs.
- Pick the route by risk: the higher the cost of a misunderstanding, the more human work you need.
What is AI eLearning translation?
AI eLearning translation means using machine translation engines (such as DeepL or Google Cloud) or large language models (such as ChatGPT) to translate the text, captions and narration scripts of a course automatically. It is one of three routes. The other two add human linguists: post-editing the AI output, or translating from scratch.
| Route | What happens | Standard that describes it |
|---|---|---|
| Raw AI translation | A machine translation engine or AI model translates the course text. No linguist checks it, or only a quick look. | Not covered by ISO 18587, which applies to full human post-editing only, or by ISO 17100. |
| AI translation + human post-editing (MTPE) | The engine produces a draft. A qualified linguist then edits every segment against the source until it is accurate and reads naturally. | ISO 18587:2017 covers "full, human post-editing of machine translation output" and the competences post-editors need. |
| Professional human translation | A professional translator translates the content, and a second linguist usually revises it. Translation memory and glossaries help, but no machine draft is used. | ISO 17100:2015 sets requirements for the core processes and resources of a quality translation service. Raw machine output plus post-editing is outside its scope. |
ISO lists both standards as under revision (checked October 2026). The draft that will replace ISO 18587 is titled "Post-editing of non-human translation output". You do not need to buy them to use the ideas. The useful point is simple: "post-edited" should mean a qualified person edited every segment, not a skim.
Translation is only part of the job. Localization adapts the course to the learner's language, culture and other requirements: dates, currency, examples, images, names and legal references. Read more about the eLearning localization process.
How accurate is AI translation for training content?
Accuracy depends heavily on the language pair and on how you measure it. In peer-reviewed studies of machine-translated instructions, Spanish often scored above 90% at sentence level, while languages such as Armenian scored far lower. Even high sentence accuracy left at least one error in many full documents.
We found no published study that compares what learners learn from machine-translated versus human-translated eLearning. The closest evidence comes from healthcare, where researchers tested machine translation of written instructions for patients. These texts are close to training content: short, instructional and high-stakes.
| Study | What was translated | What it found |
|---|---|---|
| 2026, BMJ Quality & Safety (vol. 35, issue 3) | ChatGPT-4 and Google Translate; 50 sets (316 sentences) of emergency department instructions into Spanish, Chinese and Russian | Sentence accuracy for ChatGPT-4: 97% Spanish, 95% Chinese, 89% Russian. Yet 16%, 24% and 56% of its full instruction sets contained at least one inaccuracy. Potential for harm: 1% or less per sentence, up to 6% per set. |
| 2025, npj Digital Medicine (vol. 8, article 629) | ChatGPT-4o; 20 pediatric discharge instructions into Arabic, Armenian, Bengali, simplified Chinese, Somali and Spanish | Raw ChatGPT-4o output did worst in Armenian and Somali. When a professional linguist post-edited it, quality matched or beat professional translation for most languages, and took 7.1 minutes on average versus 16.8. |
| 2021, Journal of General Internal Medicine (vol. 36) | Google Translate; 20 discharge statements into 7 languages (400 translations) | Meaning kept in 82.5% overall: 94% Spanish, 90% Tagalog, 82.5% Korean, 81.7% Chinese, 67.5% Farsi, 55% Armenian. The authors said it "should not be relied on" for patient instructions. |
| 2019, JAMA Internal Medicine (vol. 179) | Google Translate; 100 sets of emergency department instructions (647 sentences) into Spanish and Chinese | 92% of Spanish and 81% of Chinese sentences were accurate. 2% and 8% had potential for significant harm. The authors said it can supplement, not replace, the English instructions. |
What this means for your courses
- 01
The language pair matters more than the tool.
Results for widely used languages such as Spanish were strong. Results for Armenian, Farsi and Somali were weak in the studies that tested them.
- 02
Sentence accuracy hides course-level errors.
In the 2026 study, 97% sentence accuracy still left at least one error in 16% of Spanish instruction sets. A module usually has far more sentences than a set of discharge instructions, so errors are likely in most modules unless a human checks them.
- 03
Post-editing closes most of the gap.
In the 2025 study, machine output edited by a professional linguist matched or beat professional translation for most languages, and took less time on average.
- 04
Newer AI is better, not perfect.
The 2026 study found ChatGPT-4 more accurate than Google Translate in Chinese and Russian, but its authors still called for "professional oversight" for anything that is not low-risk.
These studies used short medical texts, not full courses with quizzes, scenarios and narration. Treat them as a guide to risk, not as a promise for your content. The only reliable test is a sample of your own course, reviewed by a native-speaking subject expert (see the pilot test).
Machine translation vs professional eLearning translation: side by side
Machine translation wins on speed and unit cost. Professional translation wins on accuracy, terminology, tone and cultural fit. AI plus post-editing sits in between: faster than translating from scratch, with a linguist accountable for every sentence.
| Factor | Raw AI translation | AI + human post-editing | Professional human translation |
|---|---|---|---|
| Speed | Minutes for a full course text | Faster than human translation from scratch | Slowest; depends on volume and review rounds |
| Cost driver | Tool subscription or usage credits | Linguist time spent editing, plus tool cost | Linguist time for translation and revision |
| Accuracy | Varies by language pair; errors likely in every long module | Checked sentence by sentence by a linguist | Checked by a translator and a reviser |
| Terminology | Consistent only if the tool supports a glossary | Glossary plus a human who knows the field | Glossary, translation memory and subject knowledge |
| Tone and cultural fit | Often literal; idioms and humor may fail | Fixed by the post-editor | Adapted from the start |
| Content inside images, video, PDFs | Usually not translated by built-in tools | Handled as separate tasks | Handled as separate tasks |
| Voiceover | AI voice from the translated script | AI or human voice from an edited script | Usually human voice from a reviewed script |
| Who is accountable for errors | Nobody, unless you review it | The post-editor and reviewer | The translator and reviser |
| Best for | Drafts, gist, low-risk internal content | Most training content | Compliance, safety, legal, medical, brand-critical courses |
Which route fits your course?
Choose the route by the cost of a misunderstanding. If a wrong sentence could cause an injury, a compliance breach or a lost customer, use professional translation or a full post-edit with a subject expert review. If the content is internal, short-lived and low-risk, raw AI translation with a quick check can be enough.
| Content type | Risk if misunderstood | Recommended route |
|---|---|---|
| Safety procedures, hazard and equipment training | High | Professional translation or full post-editing, plus review by a native-speaking subject expert |
| Compliance, legal, data privacy, anti-bribery | High | Professional translation or full post-editing, plus legal or compliance review in each country |
| Medical, pharma and clinical training | High | Professional translation with subject expert review |
| Customer-facing product training and onboarding | Medium | AI translation + full post-editing |
| Software simulations and system training | Medium | AI translation + full post-editing; recapture screens in the localized software |
| Internal soft skills, policies and updates that change often | Low to medium | AI translation + post-editing, or raw AI with a native speaker check |
| Internal drafts, meeting notes, content you only need to understand | Low | Raw AI translation |
Workplace safety rules raise the stakes. The US Occupational Safety and Health Administration (OSHA) has stated that "an employer must instruct its employees using both a language and vocabulary that the employees can understand."
Not sure which route is right for your courses?
What AI translation tools miss inside an eLearning course
A course is not a text file. Built-in AI translation handles most on-screen text, but vendor documentation lists content it skips: text inside images and videos, attached PDFs, embedded media and some interactive logic. These gaps need a person, or learners see a mix of two languages.
Check these after any AI translation
- 1Text inside images and video. Articulate says text within images or videos is not translated by Articulate Localization. Rebuild these graphics with live text or translate them separately.
- 2Attachments and links. Linked or attached resources, such as PDFs, are not translated by Articulate Localization.
- 3Embedded blocks. In Rise 360, embedded Storyline blocks are not translated; translate them in Storyline 360 first.
- 4Triggers, variables and player labels. When you translate a Storyline 360 course by export, trigger conditions and player text labels must be translated by hand.
- 5Text expansion. W3C cites IBM figures: text up to 10 characters can grow by 200% to 300% in translation, and text over 70 characters by about 130%. Buttons, tabs and labels overflow first.
- 6Right-to-left languages. Arabic and Hebrew are written right to left. Layouts, numbers and English names inside right-to-left sentences need checking by a native reader.
- 7Captions and timing. Captions must match the new narration. WCAG 2.1 success criterion 1.2.2 (Captions, Prerecorded) is Level A.
- 8Language tags. WCAG 2.1 SC 3.1.1 (Level A) requires the default language of each page to be set in code, and SC 3.1.2 (Level AA) covers passages in another language. Without it, a screen reader may read French with English pronunciation.
- 9Quiz logic and feedback. Check that correct answers, feedback text and scoring still match after translation.
- 10Dates, numbers and units. 03/04 means March 4 in the US and 3 April in much of the world.
- 11Software screens. Adobe's localization guide for Captivate Classic says to re-record simulations in the localized version of the application.
For the full list of steps, see the eLearning localization process guide.
Translation software for eLearning: what the authoring tools can do
Several authoring tools now include AI translation. Articulate Localization translates Rise 360 and Storyline 360 courses into 80+ languages, and iSpring Suite has a built-in AI translator. Others, like Adobe Captivate Classic, use an export, translate and import workflow. All still need a human review.
| Tool | Translation options (vendor docs, checked October 2026) | What to know |
|---|---|---|
| Articulate Localization (Rise 360, Storyline 360) | AI translation into 80+ languages; custom glossaries; in-context validation of translations in Review 360. Since March 31, 2026, it also translates Storyline 360 text-to-speech audio and AI narration in Rise 360. | Articulate uses DeepL, Amazon Web Services and Google Cloud, and states that "accuracy can vary depending on the language pairs." It also states that customer data is not used to train its AI translations. |
| Articulate Storyline 360 (export route) | Export course text to XLIFF (1.2 or 2.0) or Word, translate it with any tool or linguist, then import into a copy of the project. | Trigger conditions and player text labels are translated by hand. Captions can be imported as SRT, VTT, SBV or SUB files. |
| iSpring Suite | A built-in AI translator launched in 2025 for 70+ languages; iSpring added 27 more languages in August 2026. XLIFF export and import are also available. | Importing a translation replaces the current text, so iSpring advises making a copy of the course first. |
| Adobe Captivate Classic | Export project captions and closed captions to Word, translate them, then import into a copy of the project. | Software simulations are re-recorded in the localized application, and timing is checked. |
Outside the authoring tools, professional linguists work in translation software that stores translation memory (past approved translations) and termbases. When a course is updated, only new or changed text needs translating again.
What to check in any AI translation tool for training
- Data use. Does the tool store your content or use it to train its models? Read the vendor's own terms, especially for free tools and confidential courses.
- Glossary support. Can you lock product names, job titles and regulated terms? Support can differ by language pair.
- File formats. Does it accept XLIFF or your authoring tool's native files, so layout and interactions stay intact?
- In-context review. Can a reviewer see the translation inside the course, not only in a spreadsheet?
- Updates. Does it detect changed text, so you only retranslate what changed?
- Language coverage. Is each of your languages supported, including regional variants such as Brazilian Portuguese or Canadian French?
AI voiceover or human voiceover?
AI voiceover is now a practical choice for many translated courses, especially content that changes often, because new narration can be generated from the edited script. Human voiceover still suits flagship, emotional or customer-facing courses. In both cases, a native speaker should check pronunciation, pacing and terminology.
Voice over translation software for eLearning works in two steps: it translates the narration script, then generates speech from it. The script is the part that needs human editing. A wrong word in the script becomes a wrong word in the audio, the captions and the transcript.
Accessibility applies in every language. If the course has narration, provide captions, and keep the language of each page set correctly so screen readers pronounce it properly (see what AI tools miss).
A safe AI translation workflow in 7 steps
A safe workflow uses AI where it saves time and people where judgment is needed. Prepare the source, choose a route per module, run the machine translation with a glossary, have a qualified linguist post-edit every segment, review inside the course, test in the LMS, and keep the approved translations for updates.
- 01
Prepare the source course
Fix English errors, remove idioms, replace text in images with live text and leave room for longer words.
- 02
Choose a route per module
Use the risk matrix above. One program can mix routes.
- 03
Build a glossary
List product names, job titles, regulated terms and words that must not be translated.
- 04
Run the machine translation
Use a tool that keeps your file structure (XLIFF or the authoring tool's own feature) and applies the glossary.
- 05
Post-edit every segment
A qualified linguist edits the draft against the source, not just the translated text on its own. ISO 18587 sets requirements for this full human post-editing and for post-editors' competences.
- 06
Review in context and test
A native-speaking subject expert reviews the course on screen. Then check layout, audio, captions, quizzes and tracking in your LMS.
- 07
Keep the approved translations
Store them as translation memory so the next update costs less and stays consistent.
What drives the cost of AI and professional eLearning translation?
The cost depends on how much human work each language needs, not only on the tool. The main drivers are word count, the number of languages, the route chosen, subject complexity, voiceover minutes, text inside graphics and video, review rounds and LMS testing.
Word count and repetition. Repeated text and translation memory reduce the new words to translate.
Languages. Each language adds translation, review and testing; low-resource languages often need more post-editing.
Route. Raw AI, AI plus post-editing and human translation need very different amounts of linguist time.
Subject matter. Regulated, technical and medical content needs specialist linguists and subject expert review.
Media. Voiceover minutes, video, animation and text inside graphics are separate tasks.
Source files. Editable source files (for example Storyline or Captivate files) are faster to localize than published packages.
Review rounds. Each in-country review round adds time.
LMS testing. Every language version needs testing in your LMS.
See what shapes eLearning localization cost, or ask for a quote with your own course details.
Want a route and timeline for your courses?
Get a Free Quote →Common mistakes with AI eLearning translation
- Publishing raw AI output for high-stakes training. A small error in a safety step or a legal definition can cost far more than the translation.
- Judging quality on one language. A tool that works well in Spanish may be weak in Armenian or Somali.
- Reviewing in a spreadsheet only. Many errors only show in context: cut-off buttons, wrong feedback, captions out of sync.
- Skipping the glossary. Without one, the same term can be translated three different ways across a course.
- Forgetting the non-text content. Images, PDFs, video text and embedded blocks stay in the source language.
- Pasting confidential courses into free tools without checking how the vendor stores or uses the text.
- Calling a quick skim "post-editing". Ask who edits, how and against which standard.
How IKHYA translates eLearning
IKHYA is a custom eLearning development company that has delivered eLearning since 2010. We translate and localize courses into 40+ languages with human or AI voiceover, and our clients are in 40+ countries. Our 500th project was a global onboarding program for a FTSE 100 financial services company, localized into 12 languages.
A route for each module
We choose the route module by module. Low-risk and frequently updated modules go through AI translation with a glossary, followed by full post-editing by a native-speaking linguist. Compliance, safety and other high-stakes modules get professional human translation and a review by your in-country subject expert. Every language version is then checked in the course itself: layout, text expansion, audio, captions, quizzes and interactions.
Tools, outputs and LMS testing
We work in Articulate Storyline 360, Articulate Rise 360, Adobe Captivate, Lectora, iSpring Suite, dominKnow ONE and Elucidat, and deliver SCORM 1.2, SCORM 2004, xAPI and HTML5. We test every course in your LMS, for example Moodle, TalentLMS, Docebo or Cornerstone. You receive the editable source files (for example Storyline or Captivate files) for future revisions.
Your free quote
Every free quote includes a recommended conversion route and a timeline for your project. Learn more about our eLearning translation services.
FAQ
For some languages and low-risk content, yes. In recent studies, AI reached 90% or more sentence accuracy in Spanish and Chinese but much lower in languages such as Armenian or Somali, and many full documents still had at least one error. Use human post-editing for training that matters.
Read more: How accurate is AI translation for training content?
MTPE is a workflow where a machine translation engine produces a draft and a qualified linguist edits every segment until it is accurate and fluent. ISO 18587:2017 describes full human post-editing and the competences post-editors need.
You can use them for drafts and low-risk internal content. Check the vendor's terms before pasting confidential material, and have a native-speaking linguist edit the output before learners see it. Pasting text into a chat or web tool also loses your course layout and interactions, so XLIFF or a built-in authoring tool feature is usually easier.
Articulate Localization translates Rise 360 and Storyline 360 courses into 80+ languages, and iSpring Suite has a built-in AI translator for 70+ languages. Adobe Captivate Classic uses an export, translate and import workflow.
Yes. Some tools translate the narration script and generate new speech. For example, since March 31, 2026, Articulate Localization translates Storyline 360 text-to-speech audio and AI narration in Rise 360. A native speaker should still check pronunciation and terms.
Not on its own. Compliance and safety training carries legal and safety risk, so use professional translation or full post-editing plus a review by a compliance or subject expert in each country.
It depends on word count, the number of languages, the route, voiceover and review rounds. AI drafts take minutes; post-editing, in-context review and LMS testing take most of the time. IKHYA's free quote includes a recommended conversion route and a timeline for your project.
Send the editable source files (or the published course if you have no source files), the list of target languages, any glossary or past translations, and whether you need voiceover. Tell us which LMS you use and who will review each language.
Read more: What drives the cost of AI and professional eLearning translation?
IKHYA translates and localizes eLearning into 40+ languages, with human or AI voiceover. Tell us your target languages and we will recommend a route for each one.
AI or professional translation: how to choose
Tell us your courses and languages, and we will recommend the right mix of AI and human translation.
