Checking for understanding: 12 techniques that work online

Why don’t “Any questions?” and “Does that make sense?” work?

Checking for understanding is the habit of collecting evidence, during the lesson, of what each learner has actually understood. The two most common checks collect almost none. “Any questions?” asks a learner to diagnose their own confusion and admit it out loud. The ones who are most lost often cannot say what they don’t know. Online, with cameras off and a quiet chat, silence tells you even less.

“Does that make sense?” is worse, because it invites a yes. Learners mistake familiarity for understanding. In Roediger and Karpicke’s 2006 study in Psychological Science, students who spent four study periods rereading a passage were more confident they would remember it than students who read it once and then took three recall tests. A week later, the rereaders recalled 40% of the passage and the recall group 61%.

Black and Wiliam, reviewing classroom practice in 1998, described a related trap: a teacher who answers their own question after two or three seconds leaves no time to think, so questions drift toward the simple and the “discussion” happens with the few quick students. The fix is the same in both cases. Ask a question whose answer shows the understanding, and get an answer from every learner.

What does the research say about checking for understanding?

The case for checking rests on two bodies of evidence. The first is formative assessment. Black and Wiliam’s 1998 review, summarized for teachers as “Inside the Black Box,” reported effect sizes of 0.4 to 0.7 in the formative-assessment studies it reviewed. Later work was more cautious. A 2011 meta-analysis by Kingston and Nash reviewed more than 300 studies, found only 13 with enough information to calculate effects, and estimated a weighted mean effect of 0.20, higher in English language arts (0.32) than in math (0.17) or science (0.09). Computer-based formative systems came in at 0.28. The honest reading: checking helps, and how much depends on what you do with the answer.

The second is cognitive science on how people remember. Dunlosky and colleagues’ 2013 review of ten study techniques rated practice testing and distributed practice as high utility, self-explanation as moderate, and rereading and highlighting as low. Several of the techniques below are checks and learning events at once: answering the question is itself part of how the idea sticks.

Checks while you teach (techniques 1–4)

The examples use one topic throughout, a first lesson on photosynthesis, so you can see how the checks fit together.

1. Hinge questions with misconception-mapped options

A multiple-choice question asked at the point where your next move depends on the answer. Dylan Wiliam’s rule of thumb: students should need no more than two minutes to answer, and you should be able to read every response in 30 seconds or less. The aim is that students with the right thinking and students with the wrong thinking give different answers.

Example: “Most of the dry mass of a large tree comes from: (A) minerals in the soil, (B) water, (C) carbon dioxide from the air, (D) sunlight.” The answer is C; A, B and D each match a common misconception. Pitfall: distractors nobody would pick, which make the question a formality. AI tutor: store the misconception behind each wrong option in the lesson plan, so a wrong choice triggers the hint written for that mistake.

2. Predict, then reveal

Ask for a prediction before showing the result, then ask the learner to explain any gap. Science teachers know it as predict–observe–explain, described by White and Gunstone in Probing Understanding.

Example: “A plant is sealed in a jar in the dark for two days. Will the carbon dioxide in the jar go up, down or stay the same?” (Up: respiration continues, photosynthesis stops.) Pitfall: revealing the answer without asking why the prediction was off, which is where the learning happens. AI tutor: record the prediction, show the result on the board, then ask for the explanation and score it.

3. Explain it back

The learner explains the idea in their own words. In Chi and colleagues’ 1994 study, eighth graders who were simply prompted to explain each line of a text on the circulatory system to themselves gained more from pretest to posttest than a control group who read it twice. It was a small study (14 prompted students, 10 controls); Dunlosky’s review, looking across the wider research, rates self-explanation as moderately useful.

Example: “In one sentence, why does a plant need light to make glucose?” A good answer says light supplies the energy and that energy ends up stored in glucose. Pitfall: accepting a parroted phrase (“because of photosynthesis”). AI tutor: score the answer against a short rubric of must-have ideas and ask a follow-up about whichever one is missing.

4. No-opt-out follow-ups

When a learner says “I don’t know,” the exchange isn’t finished until that same learner gives the right answer. Doug Lemov named this “No Opt Out” in Teach Like a Champion.

Example: The learner can’t name the gas plants take in. Narrow it: “We breathe it out. What is it?” Then return to the original question and have them answer it in full. Pitfall: moving on to someone else, which teaches that “I don’t know” ends the question. AI tutor: nudge after silence, give a smaller hint, then re-ask the original question.

Checks after a concept (techniques 5–8)

5. Find the error

Show a plausible answer containing one planted mistake and ask the learner to find and fix it. It tests whether they can apply the idea, not just recognize it.

Example: “A student writes: ‘Plants photosynthesize during the day and respire at night.’ What’s wrong?” (Plants respire all the time; in daylight, photosynthesis usually outpaces it.) Pitfall: leaving the wrong version as the last thing on screen. AI tutor: check that the learner both spots the error and states the corrected version, then write the correction on the board.

6. Confidence ratings

After answering, the learner rates their confidence: guessing, fairly sure or certain. Confidence alone is not evidence, as the rereading result above shows, but paired with the answer it sorts learners usefully. Confident and wrong is a misconception to fix first; unsure and right needs practice, not reteaching.

Example: after the tree-mass question, “How sure were you?” Pitfall: asking for confidence instead of an answer. AI tutor: store the rating with the answer and put confident-wrong concepts first in the revisit queue.

7. Apply it to a new case

Change the surface of the problem and keep the principle. A learner who has only memorized the example will stall.

Example: “Pondweed under a lamp gives off bubbles. What happens to the bubbles if you move the lamp farther away, and why?” (Fewer bubbles: less light, slower photosynthesis, less oxygen.) Pitfall: a “new” case that is the same example with new numbers. AI tutor: keep transfer questions and their expected answers in the reviewed lesson plan rather than improvising them, so the science stays right.

8. Show your work

Ask for the steps, not just the result. In a live online class, everyone writes on a shared whiteboard, or types in the chat and hits enter at the same moment so nobody copies.

Example: “Write the word equation for photosynthesis.” (Carbon dioxide + water → glucose + oxygen, using light energy.) Pitfall: marking only the final line, which hides where the reasoning broke. AI tutor: ask for the steps one at a time by voice or text; handwritten working on paper is still easier for a human to read.

Checks that come back later (techniques 9–12)

9. Exit tickets

Two or three quick prompts at the end of a lesson, read before the next one. Include one question on the hardest idea and one asking what is still unclear.

Example: “Name the two inputs of photosynthesis. Where in the cell does it happen? What’s one thing you’re unsure about?” Pitfall: collecting tickets nobody reads. AI tutor: turn the answers into a per-learner summary for the teacher and open the next session with the weakest point.

10. Low-stakes retrieval quizzes

Short recall questions, ungraded, with the answer given straight after. In Roediger and Karpicke’s experiment, rereading won on a test five minutes later (83% vs 71%), but after a week the recall group was well ahead (61% vs 40%).

Example: start the next session with “Without looking back: why did the bubbles slow down when the lamp moved?” Pitfall: treating it as a grade, which makes learners anxious and tempts them to look things up. AI tutor: ask aloud, score, give the answer, and log which ideas were forgotten.

11. Faded worked examples

Show a fully worked example, then remove steps one at a time until the learner does the whole problem. Renkl and Atkinson describe fading as a way to raise the demands gradually so learners aren’t overwhelmed; in “backward” fading, the last step is the first to go.

Example: worked: “45 bubbles in 3 minutes is 15 bubbles per minute.” Faded: “60 bubbles in 5 minutes. You do the last step.” Then: “Compare the rates at 20 cm and 40 cm from the lamp.” Pitfall: fading too fast, or never fading at all. AI tutor: ask for the next step before showing it on the board.

12. Spaced revisits of weak concepts

Come back to what was shaky after a gap: later in the lesson, at the start of the next one, a week on. Dunlosky’s review rates distributed practice as high utility.

Example: the learner missed the sealed-jar prediction, so revisit it at the end with a variant (the same jar in bright light) and again next session. Pitfall: revisiting everything equally, or straight away, which isn’t spacing. AI tutor: flag weak concepts per learner and schedule them into later recaps.

All 12 checking for understanding strategies at a glance

You don’t need all twelve in every lesson. A hinge question, an explain-it-back and a retrieval question next time cover most of the value. The rest are for when those show a problem.

If you are evaluating AI tools for any of this, the quickest test is whether the tool asks the learner questions or only answers them. A chatbot that explains well but never checks gives you the same false signal as “Does that make sense?” We compare the two designs in Tutezy vs ChatGPT.

What does a 15-minute lesson with checks look like?

Here is how the photosynthesis lesson above might run one to one, online. Every step ends with the learner producing something, and every answer decides what happens next.

1. 1

   Warm-up retrieval (2 min)

   Two recall questions on last lesson’s weakest point. Wrong answers get a quick reteach before anything new.
2. 2

   Inputs and outputs (4 min)

   Teach the word equation, then the tree-mass hinge question. A wrong option triggers the hint for that misconception.
3. 3

   Why light matters (3 min)

   Predict what happens when the lamp moves away, reveal, then “explain it back” in one sentence.
4. 4

   Measuring the rate (3 min)

   A worked bubble-count example, then a faded one where the learner does the last step.
5. 5

   Find the error (2 min)

   The “respire only at night” statement, with a confidence rating. Confident-wrong goes to the top of the revisit list.
6. 6

   Exit ticket (1 min)

   Three prompts. Weak concepts become the next session’s warm-up, which closes the loop.

This is the same loop that made tutoring effective in the studies behind Bloom’s 2 sigma problem: check, correct, don’t build on a gap, come back to it. For self-paced courses, the same checks placed between videos are one of the fixes in our guide to improving online course completion rates.