Edtech corporations' artificial intelligence (AI) competition is shifting from content generation to verification. AI can quickly create test items and textbooks, but even with accurate source material, answers or explanations can be wrong, or the output may not match learners' levels. Corporations are working to raise content reliability by applying education-expert review standards to AI.
On the 20th, according to the Edtech industry, more corporations are using verified source data to generate educational content. The aim is to reduce the chance that AI will create items based on inaccurate information and to raise the reliability of the output.
Daekyo(019680) and Megazone Cloud founded a joint venture, Definition, which in its AI question-generation service "Problem G" creates passages and items based on prompts owned by teachers or achievement standards announced by the Ministry of Education.
VISNAG Education used past exam questions as data for its workbook-generation function, which composes a workbook that meets conditions when difficulty, grade, correct-answer rate, point values, and other factors are set. It lowers the chance of errors by selecting and combining questions from verified items that appeared on exams and match student conditions.
Even when using source data, it is hard to eliminate errors completely. In May, the international academic journal Postgraduate Medical Journal, published by Oxford University Press, carried a systematic review titled "Validity of AI-generated multiple-choice questions in medical education," which found that the error rate of AI-generated multiple-choice items ranged from under 1% to as high as 45%, depending on the study. The researchers said Generative AI is useful for drafting items, but not at a level that can be used without human verification.
In response, corporations are advancing their verification systems alongside the use of source data. Weverse Brain, which runs the foreign-language learning service "Speaking Max," built an "AI content factory" that creates six-stage draft learning content—such as listening items and repeat-after-me sentences—when source materials like lecture videos, curricula, and existing items are input. It added a process to review the AI-produced output. It reflected human-applied verification standards—such as notation appropriate to the learner's level and example-sentence placement by instructional sequence—into an AI-based review system.
Other corporations are verifying both the accuracy and learning suitability of generated educational content. They also check whether the generated content can be used in actual classes, and release content only after reviewing not just the correctness of answers but also the learner level, curriculum, item sequencing, and explanation methods.
As the center of gravity in Edtech competition moves from generation to verification, making verification processes more efficient is expected to emerge as a task. Even if review systems are precisely designed, humans must make the final call, so the more educational content is produced using AI, the heavier the burden on review staff may become. Another variable is how much expense corporations can bear for the verification process.
Jo Se-won, CEO of Weverse Brain, said, "The molecular-level convergence of educational design expertise for review and AI technological capability for generation will determine future competitiveness."