How to Use AI to Prepare ESL Questions for Speaking Tests

Speaking tests are the assessment I dread most -- not because of grading, but because of the prep. You need enough questions to test 25 students without repeating yourself so much that student #20 already knows all the answers from the hallway chatter. AI makes this manageable.

The Problem with Reusing Speaking Test Questions

We've all done it. Same 10 questions, semester after semester. Students share them. By the third testing period, half the class has memorized answers to questions they haven't been asked yet. You need fresh question banks. Lots of them.

Writing 50 unique speaking questions at a consistent difficulty level takes hours. It's the kind of task that AI handles in minutes.

Generating Question Banks with AI

ChalkLab lets you generate ESL speaking test questions by level and topic. Ask for 20 B1 questions about daily routines, travel, and food -- you'll get questions calibrated for that level with appropriate vocabulary expectations.

The real trick is generating multiple sets. Create three or four parallel versions of the same test. Each set covers the same topics at the same difficulty, but with different specific questions. Now student #20 can't prepare by talking to student #1.

Structuring Test Questions by Difficulty

A good speaking test follows a warm-up to cool-down arc:

• Part 1 -- Personal questions: "What's your favorite meal to cook?" Low stakes, builds confidence
• Part 2 -- Description/narration: "Tell me about a time you got lost." Requires past tense, sequencing
• Part 3 -- Opinion/analysis: "Should schools teach cooking? Why or why not?" Higher cognitive load

AI tools can generate questions for each section separately, which helps maintain that progression across parallel test versions.

Matching Questions to Common Frameworks

If your school uses CEFR levels, you can prompt ChalkLab for A2 or B2 specifically and it'll adjust vocabulary and complexity. For IELTS-style practice, ask for Part 1, 2, and 3 format questions. The output won't perfectly replicate official tests, but it's close enough for classroom practice.

What to Watch For

AI occasionally generates questions that are too similar within a set -- two questions about transportation that basically ask the same thing in different words. Scan each set for overlap. Also check for cultural assumptions -- a question about "your last vacation abroad" doesn't work for students who've never left their country.

Generate more questions than you need. If you want 15, generate 25 and pick the best ones. It's still faster than writing 15 from scratch, and you end up with better quality through selection. For more on AI-generated conversation questions, see my conversation tools roundup.