As parents, many of us have welcomed AI tools like ChatGPT into our children’s study routines. On the surface, an AI that offers 24/7 writing feedback or instant math explanations feels like having an endlessly patient, top-tier private tutor at home.
However, when our engineering and learning design team—alongside professional exam instructors—subjected AI grading systems to rigorous stress tests for high-stakes qualifications (such as IELTS, IGCSE, and A-Levels), we uncovered a concerning reality: AI hallucinations and unreliable scoring are far more prevalent than most people realize.
If left unchecked, relying blindly on unconstrained AI can give students a false sense of security and actively hinder their academic progress. Here is what we discovered during our testing, how we solved it with deterministic engineering, and how you can ensure AI remains a safe, effective tool for your child.
To test how accurately AI evaluates student writing, we intentionally submitted essays embedded with common student mistakes—such as spelling errors, missing punctuation, run-on sentences, and subtle grammatical traps. The results revealed three major vulnerabilities:
When we submitted the exact same essay twice without changing a single character, the raw AI awarded a top-tier score of 85.0 (C2 Expert) on the first run, but dropped to 20.0 on the second. Because generative AI models rely on probabilistic sampling, unconstrained AI can produce drastically different results for identical work, confusing students about their true skill level.
Standard LLMs are trained to be polite and encouraging. In our testing, when an essay was filled with repeated errors—such as confusing "there" and "their", subject-verb mismatches, and run-on sentences—the AI caught only a fraction of the mistakes and handed out an inflated score. It prioritized making the user feel good over providing strict, accurate feedback.
When reviewing student vocabulary, raw AI frequently pushes students to replace natural, clear phrasing (e.g., "positive emotional state") with overly complex alternatives (e.g., "complex and subjective emotion"). In many cases, these recommendations alter the intended meaning of the sentence, teaching students flawed writing habits that lower their actual exam scores.
Generative AI models do not "know" facts or official exam rubrics; they predict the next most likely word based on patterns in their training data.
When an AI model is asked to grade an essay or explain a complex mark scheme without strict external constraints, it often fills in knowledge gaps by fabricating plausible-sounding criteria out of thin air. In education, this is called hallucination. Relying on hallucinated feedback for official Cambridge or Edexcel exams can be disastrous when students face real human examiners.
Does this mean AI has no place in education? Absolutely not. It simply means AI should never operate without strict code-based supervision.
To ensure students receive 100% reliable, objective feedback, MOOBON Axiom uses a deterministic multi-layer engineering pipeline that constrains the AI:
If your child uses AI for daily self-study, here are two practical guidelines to keep in mind:
AI is an extraordinary tool for high-density practice, but only when it is held to uncompromising standards of accuracy. By engineering strict guardrails around generative models, we ensure that students get the honest, consistent, and rigorous feedback they need to build genuine mastery.
Learning isn't about getting artificial praise—it's about actively engaging your brain to reach your full potential.