An AI grading system for French-language exams used in Canadian immigration applications produced confident—but incorrect—scores for dozens of learners after their microphones failed. The system returned silence as a caption credit instead of an error.
The incident was one of five recurring failure modes the company identified after years of using AI to evaluate spoken and written French for permanent-residence candidates. Each failure generated a numerical score rather than a crash, and each went unnoticed until learners acted on the results.
The model graded silence as beginner French
When learners’ microphones captured no audio, the speech-to-text model returned the French caption line “Sous-titrage Société Radio-Canada,” a phrase it had frequently encountered in training data paired with quiet segments. The grading model treated that phrase as a speaking sample and assigned the lowest level on the scale.
No component checked whether there was actual content to grade. The solution involved adding a pre-grading filter that rejects known transcription artifacts, prompts, and transcripts below a minimum length. Learners now receive a message to verify their setup and retry rather than a confident wrong score.
Related: Software Rollouts Often Fail to Deliver
Evaluating an AI assessment vendor should include asking what a recording of silence produces. The response can be telling.
Question banks developed hidden patterns
A review of the company’s multiple-choice banks revealed that 97% of correct answers appeared in only two of the four positions. While no learner complained, test-savvy candidates could exploit the pattern to improve scores without understanding the material.
Position bias is one example of statistical tendencies that emerge in item banks. Distractors consistently shorter than correct answers or grammar questions where the longest option always wins are common issues. AI-generated items make this worse, because a model asked for a thousand questions will reproduce the same structural habits repeatedly.
The company now shuffles options in a fixed way for each question and conducts automated reviews on every new batch of generated items before learners see them.
Related: Business leaders weigh AI investment options carefully
Correct answers for the wrong reasons
Reading and listening comprehension questions often allowed learners to score points by matching words without grasping the meaning. In one case, the passage stated, “the meeting was moved to Thursday because the director was travelling,” and the correct answer repeated the phrase word for word.
A review found over a hundred questions where the right option could be assembled directly from the source text. The fix introduced a “liftability gate” that flags items exceeding a paraphrasing threshold. The correct answer must now require understanding the concept, expressed in different words.
Accurate feedback that had no impact
Early AI feedback was technically correct but ineffective. It informed learners of their level and listed error categories in encouraging language. Learners read it once and repeated the same mistakes the next day.
What proved useful was specificity tied to evidence. Instead of generic advice like “work on verb tenses,” the feedback now includes the learner’s own sentence alongside a corrected version, with the rule that distinguishes them. Vocabulary suggestions shifted from “vary your word choice” to pointing out the three instances where a repeated word weakened an argument.
Related: Raccoon Gang Joins Canvas LMS as Partner
The company adopted a new standard: every piece of feedback must include something the learner can challenge. If feedback could apply to any submission, it provides no real insight. This requirement reshaped how prompts were written more than any model improvement.
These failures are not limited to language testing. Any team using AI to score assessments—whether for certification, skills, or practice—will likely encounter similar issues. The important step is asking practical questions early: What does silence produce? What patterns exist in the question bank? Can answers be matched directly from the source? What happens to a submission when grading fails?
A learner who loses a practice exam an hour before their real test will not care whether the model it came from was accurate or not. These questions require recognizing that the model is the least likely part to fail and applying scrutiny to everything surrounding it.
