At BlueDot Impact, we run online courses in [AI alignment](https://aisafetyfundamentals.com/alignment/), [AI governance](https://aisafetyfundamentals.com/governance/), and [biosecurity](https://biosecurityfundamentals.com/pandemics/). We’ve trained 3,500 people in 110 countries, including staff in governments (UK AI Safety Institute, US CDC, UK Health Security Agency) and the private sector (OpenAI, Anthropic, Google DeepMind).

Over the years, we've experimented with many learning activities. There’s one we developed that consistently gets high ratings and seems to be particularly effective: the teach-swap-explain. This article explains the teach-swap-explain activity, how to use it, and why it’s so effective - drawing from scientifically backed science of learning principles.

## What is teach-swap-explain?

Before the session, participants choose something to research. On our courses, this is often a particular case study or scientific paper. Here’s an example exercise from [session 2 of our AI alignment course](https://course.aisafetyfundamentals.com/alignment?session=2):

> **Misalignment case study**
> In this exercise, you'll explore a random case study and analyse it with the frameworks you've learnt about from the resources. In the session, you'll have to teach one of your fellow participants about your case study so make sure you know it well! To start, pick one of the following case studies:
> 
> - **_[Racial bias in health algorithms](https://www.ftc.gov/system/files/documents/public_events/1548288/privacycon-2020-ziad_obermeyer.pdf)_**
> - **_[Keys and Chests environment (section 3.3)](https://ar5iv.labs.arxiv.org/html/2105.14111)_**
> - **_[The evolved radio (section IV. B and C)](https://people.duke.edu/~ng46/topics/evolved-radio.pdf)_**
> - **_[Pneumonia chest radiographs ("Discussion" section)](https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1002683#sec026)_**
> - **_[OpenAI emergent tool use ("Surprising behaviors" section)](https://openai.com/research/emergent-tool-use#surprisingbehaviors)_**
> - **_[CycleGAN steganography ("Hidden Information" section)](https://ar5iv.labs.arxiv.org/html/1712.02950)_**
> - **_[Skin lesion classification ("Nonstandardized and standardized input in AI classification" section)](https://www.sciencedirect.com/science/article/pii/S0022202X18322930)_**
> - **_[Radiographic COVID-19 detection](https://www.nature.com/articles/s42256-021-00338-7)_**
> - **_[Atari games (section 5)](https://ar5iv.labs.arxiv.org/html/1802.08842)_**
> 
> Tell your cohort (in Slack) what case study you've selected - ideally so you can cover a range of case studies. Then answer the following:\*
> 
> 1. _What do you think the human-level intended outcome for the system was?_
> 2. _What was the goal specified to the model?_
> 3. _What did the model actually do?_
> 4. _Is this outer or inner misalignment?_
> 5. _Only if outer: What flaw in the reward signal is being exploited? Why does this achieve the goal specified, but not the human-level outcome?_
> 6. _Only if inner: What change in distribution occurred between training and deployment? (anthropomorphising the policy here) What learned goal does the model appear to pursue when deployed?_
> 7. _Does this failure have any traits that are similar to a scheming or sycophantic model? (the answer will be no for some case studies!)_

Then in the live session, participants are paired up with someone who studied a different case study. One person teaches their prepared material to the other for 7 minutes. The learner asks questions throughout. Then they swap roles for another 7 minutes. Finally, everyone returns to the main group where a few participants explain what they just learned (not their original material) to everyone.

If you're thinking this sounds suspiciously straightforward to be one of our best activities – you're right. The magic lies in how it leverages multiple principles of effective learning simultaneously, while creating perfect conditions for engagement.

## Why is it so effective?

### Science of learning principles

The teach-swap-explain structure activates several [key principles from the science of learning](https://www.upword.ai/blog-posts/five-principals-of-effective-learning).

**Deep processing:** When you teach someone else, you can't just passively absorb information – you have to actively process it. The science tells us that the more mental processing we do with information, the better we retain it. The activity forces this processing for both the teacher and learner:

- The teacher has to think about how to structure their explanation, generate examples and answer lots of live questions. In general, being able to explain something to others and field their follow-up questions requires much more understanding, rather than surface level memorization.
- The learner is pushed to be a very active listener who asks lots of relevant questions, given they have the goal of being able to explain it back to the group later.

**Deliberate practice:** Learning science shows that focusing on the challenging aspects of a task is crucial for improvement. The teach-swap-explain format naturally does this:

- The teacher is pushed by the learner’s questions to explain parts of the concept the learner hasn’t immediately grasped, which are usually the most tricky parts. These questions can also often highlight where the teacher is less clear themselves, which forces the teacher to clarify their own thoughts or go back to the resources to better understand them. The teacher also gets immediate feedback when explanations aren't landing or are contradictory from the learner.
- The learner gets to focus on the areas they are most confused by, and ask questions to clarify their understanding in these areas.

**Associations:** Learning sticks better when we connect new information to existing knowledge. We generally design the teach-swap-explain prompts to strengthen these associations. In the example above you can see we’re not just getting people to learn what happened in the case study - we’re making them process it and associate it with the learning materials for that week on distinguishing inner vs outer alignment.

**Chunking:** People can take in about 3-4 key points (or ‘chunks’) of information at one time. Given each teaching session is just 7 minutes, there is usually only time for about this many points to be made - preventing overloading participants and having them forget the materials.

**Dual coding:** We learn best when we combine both verbal and visual content. This doesn’t always happen, however we do encourage participants to leverage this by either:

- Screen sharing to show a relevant diagram or image as they explain the concept
- Using the Zoom whiteboard feature to sketch out a diagram as they talk through the idea (which also boosts the teacher’s deep processing!)

### Motivation theory

Our courses are voluntary, and because of the high-quality of our participants they often lead busy lives doing important roles! This means we need to compete with their other priorities.

Self-determination theory (SDT) identifies three core psychological needs that drive intrinsic motivation: autonomy (freedom to choose what to do), competence (experience mastery), and relatedness (connect with and experience caring for others).

<Embed url="https://storage.k8s.bluedot.org/website-assets/migrated/a2bbff91-8a4b-44d1-a23d-ab852751362f.svg" />Teach-swap-explain provides some of all three:

- **Autonomy:** Participants choose which case study they explore in the exercise, and get to choose how to explain it (as a teacher), and what to explore about it by asking questions (as a learner)
- **Competence:** The teacher demonstrates their competence when they’re explaining the concept that they know better than the learner.
- **Relatedness:** The live activity has participants interacting with each other 1:1, where they have the full focus of each other. The teacher and learner both have to work on the shared goal of understanding each other's case studies, and are therefore incentivized to be interested in each other's case study. To do well, participants need to empathize with the other person to understand where they are getting confused.

### Other engagement benefits

Beyond learning theory, there are some simple practical advantages.

The activity motivates engaging with the exercise, and doing it well (no hiding on mute here!). Because you know that someone else will quiz you deeply to understand it themselves, you need to have a good understanding - and will be found out if not.

The activity also ensures everyone gets roughly equal speaking time, preventing the common problem of a few voices dominating group discussions. And people aren’t silent for long periods of time - there’s a constant back and forth of questions.

## Try it yourself!

Want to give it a go yourself? Find a colleague interested in teaching techniques and try this exercise:

Before meeting, each person researches different ideas and how they map onto teach-swap-explain:

- Person A: [Adult learning theory](https://www.growthengineering.co.uk/adult-learning-theory/)
- Person B: [Introducing desirable difficulties](https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/04/EBjork_RBjork_2011.pdf)

Now run a teach-swap-explain session:

- Person A teaches B how teach-swap-explain aligns with their researched idea (7 minutes)
- Person B teaches A the same, for their idea (7 minutes)
- Share your most surprising insights from each other's learnings with another colleague C!


Teach-swap-explain: a learning activity for course designers to create highly effective learning experiences

Adam Jones • December 12, 2024