1 of 69

Avishkar Bhoopchand

Google DeepMind, Deep Learning Indaba�13 February 2025�MENA ML - Doha, Qatar

AI For Education

The challenges and opportunities

2 of 69

Talk Outline

Introduction
The Scale and Complexity of the Problem
Evaluation
Technical / Research Challenges in AI for Education Systems
Ethics and safety considerations

3 of 69

Introduction

4 of 69

5 of 69

6 of 69

7 of 69

8 of 69

9 of 69

10 of 69

11 of 69

UN SDG 4

Ensure inclusive and equitable quality education and promote lifelong learning opportunities for all.

12 of 69

Bloom, B. 1984. The 2-sigma Problem: The Search for Methods of Group Instruction as Effective and one-to-one Tutoring

13 of 69

The most effective teaching method is also not possible to scale

14 of 69

Or is it..?

15 of 69

Enter Generative AI

Students are early adopters

Survey usage stats:

86% use genAI
54% weekly;
24% daily

Digital Education Council Global AI Student Survey 2024

16 of 69

So is ChatGPT enough…?

LLMs convey information

They are tuned to be helpful.

This is not the same as learning

Creates a false sense of mastery

Lehmann M et al. 2024. AI Meets the Classroom: When Does ChatGPT Harm Learning?

17 of 69

The Scale & Complexity of the Problem

18 of 69

Education for all

1.5 billion students worldwide each with different:

Prior knowledge
Learning speeds
Native languages
Cultural contexts
Learning preferences
Access to technology
Support systems at home

19 of 69

Not just for Amira in her maths class…

The elementary student learning to read

The high school student tackling physics

The adult learning a new language

The vocational student mastering a trade

20 of 69

Complexity of 1:1 Tutoring

21 of 69

Immediate Feedback

22 of 69

Immediate Feedback

Infers students knowledge

23 of 69

Immediate Feedback

Infers students knowledge

Adjust explanations to suit the student

24 of 69

Immediate Feedback

Infers students knowledge

Adjust explanations to suit the student

Builds Rapport and Trust

25 of 69

Immediate Feedback

Infers students knowledge

Adjust explanations to suit the student

Builds Rapport and Trust

Adjust to emotional & cognitive state

26 of 69

Pedagogy Principles

Encourage Active Learning

Deepen Metacognition

Manage Cognitive Load

Motivate & Stimulate Curiosity

Adapt to learner’s goals & needs

Jurenka et al. 2024. Towards Responsible Development of Generative AI for Education: An Evaluation-Driven Approach

Now unfortunately there is some disagreement and debate in the teaching and learning sciences community about what exactly constitutes good pedagogy, which by the way is the method and practice of teaching, and you may have seen this yourself where different teachers have different styles and there’s not necessarily one style that works better than another, it depends on the learner and on the context.

But as part of the work my team at Google DeepMind did was to survey the literature, speak to many teachers and learning sciences experts and distill their knowledge into 5 general principles of good pedagogy that can be applied to both the design of materials and to the act of teaching and tutoring.

These principles are:

Encourage active learning: active learning means that the learner should manipulate information through discussion, practice, and creation, instead of passively absorbing information
Manage cognitive load: information should be presented in multiple modalities, be well structured, and segmented into manageable chunks
Deepen metacognition: metacognition means “thinking about thinking”, so learners should reflect on the thought process they followed in solving problems. This enables them to generalise their skills beyond a single context
Motivate & stimulate curiosity: which leads to self-efficacy and lifelong learning
Adapt to the learners goals and needs: by assessing the current state and the goals, and making a plan to bridge the gap

We also break these principes down further into more specifics, and you can have a look at the paper referenced here if you’re interested in learning more.

27 of 69

An AI Tutor needs to…

Adhere to pedagogical principles

Model the full space of human knowledge

Understand how knowledge builds upon itself

Adapt to individual learning trajectories

Maintain engagement over months and years

Do all this while being robust, fair, and explainable

28 of 69

AI is set to disrupt Education

29 of 69

Haven’t we heard this all before?

MOOCs were also set to disrupt education:

Democratise access to quality education globally
Self-paced interactive learning would outperform traditional education
Free / low-cost would disrupt traditional institutions

30 of 69

The Reality…

MOOCs have very low completion rates: < 10%

Those who succeed are typically already educated, often with post-graduate degrees and affluent

They are already skilled at self-directed learning

Sustainable business models have proved elusive

31 of 69

The Lessons

Technology often amplifies rather than bridge existing educational divides�We need to design explicitly for struggling learners

Learning infrastructure is as important as content�Design with community, feedback, accountability in mind

Tools for independent learning work best for those who need them the least�Self-learning is a complex metacognitive skill that needs to be developed

Education is deeply embedded in social and cultural contexts�One-size-fits-all solutions typically fail

Technology alone is not the answer

32 of 69

Sociotechnical systems

Stakeholders

Community Impact

Power Dynamics

Institutional practices

Cultural Norms

Policy & Governance

In order to succeed and to avoid falling into these traps, we need to take a sociotechnical approach. Unfortunately I missed Shakir’s talk on Sunday, but I imagine he would have spoken about this.

In developing AI solutions for education, we have to engage all the different stakeholders - learners, teachers, parents, administrators, government, academics

We have to engage with established institutional practices

We have to consider the power dynamics that might be at play, for example between a teacher and a student, between schools and teachers, between governments and schools

We have to consider the community and what impact any technology we introduce may have on them

We have to consider cultural and social norms, how these vary between communities and how they might affect the systems we design. For example, in many countries one may find that cultural norms about not questioning elders might make students less likely to ask questions in class

And we also need to consider the impact of policy and governance

We need to take a participatory approach and include all of these players in the design of our systems.

33 of 69

Demo - LearnLM

34 of 69

Jurenka et al. 2024. Towards Responsible Development of Generative AI for Education: An Evaluation-Driven Approach

LearnLM Team, Google. 2024. LearnLM: Improving Gemini for Learning

35 of 69

Evaluation

36 of 69

Which outcome is better?

🧑🏽‍🎓

Gets perfect test scores but can’t apply concepts to a new problem

👨🏻‍🎓

Scores lower but develops a deep understanding

🧑🏽‍🎓

Learns slowly, but develops strong study habits

👨🏻‍🎓

Makes quick progress, but becomes dependent on ChatGPT

37 of 69

What to measure?

Test scores, completion rates, engagement time

But this is not the full picture

Need to consider:

The time effect
Multiple goals
Generalisation
Ethical considerations

We need: standard protocols, longitudinal studies, metrics for understanding, measurements of secondary effects

Of course pragmatically, we may want to start with what’s easy to measure. Things we already have quantitative metrics for - test scores, completion rates and engagement time.

But this doesn’t give us the full picture. We also need to consider

The time effect of learning - is learning persistent? Is it transferable?

That there are multiple goals: academic performance, conceptual understanding and emotional well-being.

Generalisation - does it work in different subjects? Cultures? Students?

And we also need to consider the ethical considerations around running evaluations in education. For example what are the ethics of running A/B tests on students? We can and should be learning from other domains that have deeply considered such issues like in healthcare.

So some of the research challenges here relating to evaluation are how to build and assess standard protocols for evaluation, how to perform longitudinal studies, how to measure understanding and how to measure secondary effects of our interventions.

38 of 69

Types of Evaluation

Extrinsic

Measures actual educational outcomes and impact

Can take a long time
Expensive
Hard to attribute

Better for strategic decisions

Intrinsic

Measure performance on specific tasks or capabilities

Quick to measure
Clear metrics
Relatively cheap
But may not correlate with real impact

Good for rapid iteration

39 of 69

1 year

1 month

1 week

1 day

1 hour

Automatic

Side by side (conversation)

(Simulated) learner ratings

Side by side (turn level)

2 stage pedagogy ratings

Effectiveness studies

Qualitative studies

User Feedback

Longitudinal Studies

Strategise

Iterate

40 of 69

1 year

1 month

1 week

1 day

1 hour

Automatic

Side by side (conversation)

(Simulated) learner ratings

Side by side (turn level)

2 stage pedagogy ratings

Effectiveness studies

Qualitative studies

User Feedback

Longitudinal Studies

Strategise

Iterate

41 of 69

Automatic Evaluations

What are they?

A Function: model output => metric
e.g. Language model rates responses of another language model

Why are they needed?

Human evaluations are slow and expensive
Allows for fast iteration … are our ideas working?
Compare many models
Quantify and track progress

Challenges

Lack of public benchmarks (none for tutoring!)
Evaluating good pedagogy is hard for humans, let alone for models

42 of 69

Automatic Evaluations

Break down the task

Carefully craft scenarios

Give the model context

Scenario-based supporting info

43 of 69

1 year

1 month

1 week

1 day

1 hour

Automatic

Side by side (conversation)

(Simulated) learner ratings

Side by side (turn level)

2 stage pedagogy ratings

Effectiveness studies

Qualitative studies

User Feedback

Longitudinal Studies

Strategise

Iterate

44 of 69

LearnLM Team, Google. 2024. LearnLM: Improving Gemini for Learning

45 of 69

1 year

1 month

1 week

1 day

1 hour

Automatic

Side by side (conversation)

(Simulated) learner ratings

Side by side (turn level)

2 stage pedagogy ratings

Effectiveness studies

Qualitative studies

User Feedback

Longitudinal Studies

Strategise

Iterate

46 of 69

TutorCopilot: A Randomised Controlled Trial

A “copilot” for online tutors

Size: 900 tutors; 1800 students

Metric: Exit tickets

Result: Significant increase in scores in treatment group

Larger effect for less effective, and less �experienced tutors

(But still far from 2-sigma!)

Wang R et al., 2024. Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise

Recently there was a paper by Rose Wang evaluating a project they called Tutor Copilot. This was the first randomised controlled trial of a human-AI intervention in live tutoring.

In this study, they took a model they had already built to act as an AI tutor and partnered with an online tutoring service. They gave a treatment group of tutors access to the AI Tutor to help them formulate responses to the students they were talking to. So in this study, it was a human-AI collaboration where the human tutor still had ultimate control, the human tutor would request a suggestion from the AI Tutor, the AI Tutor would give multiple options to the tutor, and the tutor could pick, edit the response or just discard the suggestions entirely when responding to the student.

The study involved 900 tutors and 1800 students

The metric they measured was scores in exit tickets, which are short tests given to the students to allow them to complete a lesson and pass on to the next one

They found a statistically significant improvement in the exit ticket scores for students who were tutored by tutors who had access to the Copilot. And more-so, this effect was larger for less effective and less experienced tutors. This was a truly amazing result, but i’ll note that the effects here are still in single digit percentages, and far from the 2-sigma that Bloom promised.

So this was the effectiveness evaluation.

47 of 69

TutorCopilot: A Randomised Controlled Trial

Qualitative analysis: higher quality strategies in treatment group

User feedback

Helpful especially for breaking down complex concepts
Sometimes not appropriate for student’s grade-level

They also performed a qualitative analysis of the strategies used by the tutors. They found that higher quality strategies were used by the tutors in the treatment group, such as prompting the student to explain or ask a question to guide the students thinking. While less effective strategies were used in the control group like giving away answers.

They also collected user feedback, another form of evaluation and found that tutors said that the copilot was helpful for breaking down complex concepts, but sometimes the responses weren’t appropriate for the students grade level.

So this study utilised 3 different evaluation approaches - effectiveness studies, qualitative studies and user feedback. These allowed them to answer different, but important questions.

Effectiveness answers “does it work”

Qualitative answers “how does it work”

User feedback answers “do our users want to use it”

48 of 69

Taxonomy of Choices (Intrinsic Evals)

Real Learners

Single-turn

Unguided

Advanced Learner

Human

Learner

Single-Turn

Pointwise

Role-playing experts

Multi-turn

Scenario Guided

Novice Learner

Automatic

Educator

Conversation-level

Pairwise (side by side)

Data Collection

Ratings

49 of 69

Technical & Research Challenges in AI for Education

50 of 69

The Challenges from an AI perspective

Multi-modal interaction (visual, verbal, written)

Real-time state estimation (student understanding)

Dynamic task selection (what to teach next)

Natural language generation (explanations)

Causal reasoning (why did the student make this mistake?)

Long-term planning (curriculum design)

All while handling partial observability and delayed feedback

51 of 69

Knowledge Representation & Tracing

Questions:

How do we model the relationship between concepts?
How do we handle (and encourage) multiple valid solution paths?
How do we model forgetting?

Possible directions:

Knowledge graph embeddings
Diverse reasoning traces

52 of 69

Theory of Mind

Questions:

How do we infer the knowledge state of the learner?
What misconceptions might explain a wrong answer?

Possible directions:

Multi-Agent methods
Causal models of misconceptions

53 of 69

Personalisation

Questions:

How do we handle the cold start problem?
How do we balance exploration vs exploitation?
How do we infer learning styles and preferences?
How do we infer the emotional state of the learner from behaviours?

Possible directions:

Agentic methods and clever prompting strategies
Meta-learning for fast adaptation

54 of 69

Content Generation

Questions:

How do we ensure factual accuracy?
How do we control the difficulty level?
How do we make content culturally relevant and appropriate?
How do we communicate with the learner in the language they’re most comfortable learning in?

Possible directions:

Grounded content generation
Recognition and handling of uncertainty
Factuality verification
Context-aware methods for multicultural and multilingual understanding

55 of 69

Interaction

Questions:

What modalities beyond text are most effective for pedagogy?
What is the role of interactivity in driving learning outcomes?

Possible directions:

Multi-modal models
Code generation
Human Computer Interaction and UX Research
Tool use

56 of 69

Optimising for Learning Outcomes

Questions:

What does AI pedagogy look like?
Can we move beyond behaviour cloning of human tutors?

Possible directions:

Multi-turn RL
Learner simulators

57 of 69

Ethical and Safety Considerations

58 of 69

Privacy and Data Protection

Data needed for personalized and effective experience

Education data can be sensitive; may involve work with minors

Case Study:

InBloom - non-profit to manage & warehouse student records.
Collected 400+ data points incl. family relationships, records about abuse & violence etc.
Legislation created to prevent state schools sharing data with 3rd party aggregators

New York Times. 2014. InBloom Student Data Repository to Close

On the topic of privacy and data protection. We of course need to collect a lot of data about learners in order to deliver a truly personalised and effective experience

But this kind of data can be highly personal and sensitive in its nature

As a case study, in the early 2010s, a non-profit called InBloom was creating a data warehouse of student records that could be shared between schools and education providers. �Their data warehouse was designed to be as comprehensive as possible, collecting over 400 data points about each student�Some of this data included highly sensitive information like about family relationships and structures, records about abuse and violence and other such highly personal information.

Eventually in 2014 after numerous concerns raised about the type of data they were collecting, some legislation was added that prevented state schools from sharing information with 3rd party data aggregators, leading to InBloom shutting down.

The lesson here is that as system designers we have a responsibility as the custodians of data, to know what is being collected, why it’s being collected, and making sure that the data is being handled responsibly and in accordance with relevant policies like GDPR and especially rules around working with data relating to minors.

59 of 69

Equity & Access

Widening the educational gap and digital divide

LLMs tend to be Western-centric and don’t perform as well in the languages and contexts of underrepresented regions

E.g. access to online learning during COVID-19 pandemic lockdowns

60 of 69

Dependency

Risk of children developing emotional dependency on AI tutors

Striking a balance between being helpful and making students dependent on the technology

Anthropomorphism and self-disclosure

When it comes to dependency, there is a risk of children developing emotional attachments to AI tutors, especially if they are interacting with them regularly. We’ve already seen such cases play out in a bad way in the space of language models.

We also saw earlier how helpful assistants can lead to a false sense of mastery where students start outsourcing their thinking to an AI. So we need to strike a balance between being helpful and making students dependent on the technology.

Finally a topic that’s studied by some of my colleagues in the context of our AI tutor is around the question of anthropomorphism. A specific example of this is around self-disclosures. Self-disclosures are sharing of personal information like your thoughts, dreams, fears or experiences. We might see this play out in a human student-tutor relationship where for example a student might indicate that they’re struggling to understand a topic. A human tutor might then make a self-disclosure by sharing some personal information about themselves and say something like “i also struggled a lot with this concept, but eventually i understood it, don’t worry, it will click soon”. For a human tutor this is a very effective technique because it builds rapport and trust with the student. So by blindly copying human tutoring conversations we might inadvertently train our AIs to make similar statements. Or perhaps we have a fancy RL process that figures out that self-disclosures like this are effective strategies. But for an AI to do this is disingenuous and ethically questionable, it can be seen as manipulation because an AI of course doesn’t have such experiences.

61 of 69

Safety

Safety considerations look different

Discussion of otherwise sensitive topics may be allowed in educational settings for development of critical thinking skills, learning about history etc

Conflict between good pedagogical practice and safety

Eg. tune an AI tutor to be encouraging of questions
But what about:�Student: <harmful question>�Tutor: That’s a great question! I’m glad you’re thinking about this

62 of 69

Conclusion

63 of 69

Summary

Global challenge in quality education for all

64 of 69

Summary

Global challenge in quality education for all

Hugh opportunity to leverage AI for social impact

65 of 69

Summary

Global challenge in quality education for all

Hugh opportunity to leverage AI for social impact

Current GenAI is not suitable for this task

66 of 69

Summary

Global challenge in quality education for all

Hugh opportunity to leverage AI for social impact

Current GenAI is not suitable for this task

Problem is more difficult than it sounds!

67 of 69

Summary

Global challenge in quality education for all

Hugh opportunity to leverage AI for social impact

Current GenAI is not suitable for this task

Problem is more difficult than it sounds!

Many exciting research opportunities

68 of 69

Summary

Global challenge in quality education for all

Hugh opportunity to leverage AI for social impact

Current GenAI is not suitable for this task

Problem is more difficult than it sounds!

Many exciting research opportunities

Join us in this journey!

69 of 69

The End

Thank You�—

شكراً