The problem teachers face when generating discussion questions is time. The promise of AI-generated questions is compelling: unlimited, instant, customised questions on any topic at any level. No more spending 20 minutes writing nine questions before a lesson. Type a prompt, get your questions, teach.
In practice, AI-generated questions are sometimes excellent and sometimes deeply mediocre. Understanding the difference - and knowing when to trust them and when to rewrite or replace them - is the practical skill that matters. And the comparison isn't just AI versus teacher-written: it's also AI versus purpose-built question banks, which operate differently from both.
YapYapGo is a classroom speaking practice tool for ESL and EFL teachers with a built-in question bank of over 36,000 questions written specifically for classroom pair speaking. Here's an honest comparison.What AI does well with discussion questions
Speed and volume. No argument here. ChatGPT, Claude, or Gemini can produce 20 discussion questions on any topic in under a minute. This is dramatically faster than teacher writing time. Topic specificity. AI handles niche, specific, or unusual topics better than pre-written question banks. "Generate 10 discussion questions about deep-sea mining" or "questions about the role of the grandmother in Japanese family life" produces workable questions on topics most banks don't cover. Level adaptation. With a clear prompt specifying CEFR level and what that means, AI produces reasonably level-appropriate questions. The adaptation is not always perfect, but with prompt refinement it improves. Cultural customisation. AI can be instructed to generate questions relevant to a specific cultural context or student background. "Generate questions about technology for adults in Vietnam where smartphone use is near-universal" produces different questions from a generic prompt. Varied formats. AI readily produces questions in specific formats: debate motions, this-or-that choices, ranking tasks, role play scenarios - whatever the format, AI can generate it on request.What AI does poorly
Consistent CEFR calibration. AI frequently misjudges what B1 looks like versus B2, or C1 versus B2. Questions labelled "B2" are often either too simple or too abstract. Without manual checking, you may be giving students questions that are pitched wrong. Avoiding repetition across sessions. AI has no memory of what your class has discussed before. Every session, you're generating fresh questions without any filter for what's been covered. Students may end up discussing similar themes repeatedly without you noticing. Classroom safety filtering. AI doesn't know your specific students. A question that's fine for adult professionals may be completely inappropriate for teenagers, or vice versa. Age-group filtering requires deliberate prompting and then checking. The "obvious correct answer" problem. A surprising proportion of AI-generated discussion questions have an implied correct answer that most people would give. Questions with obvious right answers produce short, shallow conversations. A question bank designed for pair speaking is engineered to avoid this; AI is not. Grammatical idiosyncrasy. AI occasionally produces questions with slightly unnatural phrasing that experienced teachers recognise immediately but which students find confusing. This is rare but worth checking.What purpose-built question banks do well
Consistent classroom safety. Banks designed for ESL contexts have been filtered for age-appropriateness, cultural sensitivity, and classroom suitability. This filter is built in rather than requiring deliberate prompting. CEFR calibration by design. Well-built question banks are written at specific levels by people who know what B1 discussion looks like in a classroom context. This calibration is more reliable than AI's level labelling. Non-repetition tracking. A question bank paired with session tracking ensures classes never discuss the same question twice. This is something AI cannot provide without external memory systems. The discussibility design. Questions written for pair speaking are specifically designed to produce genuine disagreement, extended reasoning, and continued conversation. This is harder to prompt AI to reliably produce.Tool tip: YapYapGo uses a purpose-built question bank of 36,000+ questions filtered by age group and CEFR level, with session tracking to prevent repetition. The AI Discussion mode does something different from the main question bank - it generates AI questions in real time during class on teacher-specified topics, which is the genuine strength of AI: specific, novel, customised prompts that don't exist in any pre-written bank.
The hybrid approach
The most effective approach uses both:
Use a purpose-built bank for the core of most lessons. Reliable CEFR calibration, age-appropriate filtering, non-repetition tracking, and discussibility design are structural advantages that matter most for the main speaking blocks. Use AI for unusual topics, custom scenarios, and specific formats. When the pre-built bank doesn't cover your specific topic (deep-sea mining, local current events, a topic from your students' specific professional context), AI is the right tool. When you need a debate motion, a ranking task, or a role play scenario, AI generates these faster than any bank. Always check AI output before using it. Read every AI-generated question before presenting it to a class. The check should take 30 seconds. Look for: obviously correct answers, cultural insensitivity, CEFR mismatch, and unusual phrasing.For 30 specific prompt templates that produce better AI discussion questions, see our post on ChatGPT prompts for ESL speaking activities. A classroom countdown timer, random student picker, and conversation topic generator keep the logistics of any discussion activity - AI-generated or bank-based - clean and efficient.
Sources:
- Nation, I.S.P. (2001). Learning Vocabulary in Another Language. Cambridge University Press. - Criteria for effective communicative task design.
- Ur, P. (1981). Discussions That Work. Cambridge University Press. - What makes a discussion question genuinely generative vs superficially engaging.
- Godwin-Jones, R. (2019). Tools and Trends in Self-Paced Language Learning. Language Learning and Technology. - AI and language teaching: practical strengths and limitations.