Teaching and learning with generative AI

Fengchun Miao
Chief, Unit for Technology and AI in Education
Future of Learning and Innovation Division, Education Sector, UNESCO
 UNESCO’s main achievements in digital education
Setting global standards on digital education
Beijing Consensus on AI and Recommendation on Recommendation on Open
Education the Ethics of AI Educational Resources

Supporting 80+ countries’ digital education policies and 5 guidance

AI and Education: Guidelines for Guidance for AI competency Guidelines Education
Guidance for ICT in Education generative AI in frameworks: on OER and
Policy-makers Policies education and research teachers & students Policies Blockchain

Steering digital futures of education (Digital open school projects)

Global convening (e.g., Digital Learning Week, Gateway)

Worldwide grass-rooted case studies

▪ 80+ examples on AI & ED: ▪ 24 winning projects of UNESCO Prize for ▪ OER: Policy, costs,
Compendium (2019); (2020) ICT in Education and transformation
–14 frontrunner
▪ First global mapping of ▪ Case studies on distance during COVID-19:
projects on OER
government-endorsed AI Finland, Korea, Saudi Arabia
curricula (2022)
▪ 14 best practices in mobile learning
 UNESCO Work on AI and Education
(100+ ministers and 10 000+ participants from 120+ countries)

2019 Conference 2020 Forum 2021 Forum 2022 Forum 2023 2024
(Beijing Consensus) (AI competencies) (AI as a (AI for teachers)
common good)

Beijing Guidance for Ministerial

AI & Education Seminars for
Consensus policy-makers Meeting on
policy 50+ countries
generative AI Digital
Recommendation Survey on AI in Learning
Watching AI Generative AI
Week 2024
on the Ethics of AI education Digital
trends
Learning
AI competencies Mapping K-12 AI Report on “K- To be launched Week 2023 5 regional
curricula 12 AI curricula” in June 2024
for school students seminars on
AI competency
AI competencies To be launched for 100+
Consultations Expert meetings in June 2024 countries
for teachers

Guidance for generative AI in education and research

A publication on “AI and Futures of Learning”

 Main publications on AI and Education
Ranking in downloads and consultations in all UNESCO publications
Top 20 for 2 years Top 10 for 2 years Top 30 for 2 years No.1 for 4 months

6 UN languages 6 UN languages + Korean 6 UN languages + Will be in 10+ languages

Portuguese

Top 40 for 2 years

 Teaching and learning with generative AI:
policy contexts

1. Regulating
ethics of AI in
education

2. Promoting the AI and education: 4. Exploiting the

design & use of AI for Toward digital use of AI to support
inclusion and equity humanism futures of learning

3. Building AI
competencies for
teachers and
students
Guidance for generative AI in education
and research: Key messages
 1. A human-AI collaboration model
for examining AI in education
AI: from moving targets Education: knowledge
to proportionality evolution
Production
E.g. ChatGPT-3 and storage Individuals
▪ Data sources: Crawling of data
webpages (61.75%), Social 1. Evolution of
media (8.86%); Libraries(15.9%), knowledge
wikipedia (3.49%) production
Access to
Human
and control
interaction
▪ Languages: EN (92.647%)， of data Human- 2. Projection
FR(1.819%), DE+ES +IT+PT+ machine social values
NL (2.60%), …CN (0.099%) Data and collaboration
algorithm
▪ Unexplainable methods based
Sovereign 3. AI job skills;
states
to generate outputs decision- challenging
making learning
▪ Generative AI doesn’t outcomes
understand semantics Human-AI
Environment
and the real world interfaces and
and AI ecosystems
devices
 Generative AI is an AI insides,
insides of human thinking processes

Transformer
across symbolic Human agency and
representations of human motivation
thinking

Human-
Supercharger accountable
of literature reviews discovery and
Human- interaction
machine
collaboration Human protecting
Discriminator linguistic and
of computational skill cultural diversity
performance
New threat for the security of
digital learning infrastructure
Blinded
to values and the real Greater harm for environment
world and ecosystems
 Teaching and learning with …

ICTs Externalized info. processors

Predicative AI Embodied Semi-internalized data advisor

ethics of AI in decision loops
+
Generative AI AI society Internalized supercharger
citizenship for thinking → embodied
cognitive tools
…… ……
2. Eight controversies around
generative AI
 Controversy 1:
Data deprivation worsening digital poverty
A taxonomy for reducing AI divides

Funding and
Digital
technology
literacy for all
capabilities of
Internet citizens to Affordability Domestic
Accessibility Capabilities training
penetration ensure the of the Local governance
of AI chips of mining locally
needed for local internet or availability of of the
and data from relevant
data relevance and data AI developers development
computing Internet models or
production cultural subscription and use of AI
localizing
diversity of
foundation
data
models

→ Define and govern the entire life cycle of the data-based productivity
 Controversy 2:
Outpacing national regulatory adaptation
Tiered shortage of domestic regulations on AI particularly GeAI
Reflections on implications for
curriculum and assessment
Capacities for proper use of Singapore No response or N/A
generative AI in education
EU, US, China,
Regulations on generative AI Japan?
No response or N/A

Adjusting copyrights laws EU, US, No response or N/A

(labeling AI-generated content) China

Ethics of AI (including in 20+ countries No response or N/A

education)

National strategies on AI 70 countries No response or N/A

No response or
General data protection laws 137 countries N/A
 Controversy 3:
Use of content without consent

▪ Violating GDPR: New York Times lawsuit on using its

content without consent
▪ Violating right to be forgotten (GenAI can’t unlearn)
Controversy 4: Unexplainable AI models
used to generate outputs
The Foundation Model Transparency Index (Stanford University)
Examples of foundation models Llama2 ChatGPT-4 PaLM 2
Indicators （%） （%） （%）
Data 40 20 20
Data labor 29 14 0
Compute 57 14 14
Methods for model’s training 75 50 75
Model basics 100 50 67
Model access 100 67 33
Capabilities 60 10 80
Risks 57 57 29
Model mitigations 60 60 40
Distribution 71 57 71
Usage policies 40 80 60
feedback 33 33 33
Impact 14 14% 0
Average scores 57 47 39

→ Is the large-language model a category mistake?

 Controversy 5:
AI-generated content polluting the internet

▪ AIGC: Disinformation, misinformation, hate speeches

▪ Data polluted by AI lead to collapse of GenAI models
 Controversy 6:
Lack of understanding of the real world

GitHub’s Hallucination Leaderboard : the actual errors are worse

Model Accuracy Hallucination Rate
GPT 4 97.0 % 3.0 %
GPT 4 Turbo 97.0 % 3.0 %
GPT 3.5 Turbo 96.5 % 3.5 %
Llama 2 70B 94.9 % 5.1 %
Llama 2 7B 94.4 % 5.6 %
Llama 2 13B 94.1 % 5.9 %
Anthropic Claude 2 91.5 % 8.5 %
Google Palm 2 87.9 % 12.1 %
Google Palm 2 Chat 72.8 % 27.2 %
 Controversy 7:
Threatening plural knowledge construction
▪ Data-poor populations have limited digital presence online →
Their voices are not represented in the data → GPT models can
further marginalize already disadvantaged
▪ LLMs may propagate race-based medicine
▪ ChatGPT replicates gender bias in recommendation letters
▪ Boston Consulting Group: The diversity of ideas from GPT-4 is 41%
lower human ideas.
 Controversy 8:
Generating deeper deepfakes

Nearly zero-cost generation The total deepfake videos

and sharing of deeper by October 2023 presents
deepfakes especially: taking a 550% increase over 2019
<25 mins and nearly $0 to with 98% being deepfake
pornography and 99% of
create a 60-second deepfake
targeting are women.
pornographic video using just
one clear face image.

https://www.homesecurityheroes.com/state-of-deepfakes/#key-findings
3. Proportionality of Generative AI
for teaching and learning
3.1 Safety of digital infrastructure and value projection

▪ Mandate the use of local data in training models

▪ Local EdGPT
3.2 Age appropriateness of chat-based learning
for children

Age limitation for independent conversations with GenAI

platforms (13 years or 16 years)
3.3 Pre-cooked content for learning
3.3.1 Content hallucinations lead to knowledge illusions
3.3.2 Undermining thinking and intellectual development
3.3.3 Inert innovation and intelligence enfeeblement
4. Strategies
 4.1 Regulate generative AI in education

▪ Risk categorization of AI for education: draw the red line

▪ Accountabilities of all key duty-bearers
▪ Validating AI for education to ensure ethical principles,
linguistic and curriculum relevance, cultural and age
appropriateness
▪ Promoting teacher and student agency, human motivation,
and intellectual development
4.2 Build AI competencies for teachers and students
UNESCO AI competency framework for teachers (under development)
▪ the primary users of AI in education
▪ the designers and facilitators of students’ learning with AI
▪ ultimate safeguards for ethical uses across learning settings
▪ role models of lifelong learners about AI
Aspects Progression
Acquisition Deepening Creation
Human-centred Benefit-risk analysis Human accountability AI society
Minimum
Mindset responsibility
Standard
Ethical principles
for teachers Safe and responsible Co-creating AI ethical
Ethics of AI uses rules
AI Foundations & Basic AI technique and Application skills Creating with AI
What AI foundations and application skills should all teachers learn?
Applications applications
AI Pedagogy AI-assisted teaching AI-pedagogy integration AI for pedagogical
transformation
AI for Professional AI enabling lifelong AI to enhance AI for professional
Development professional learning organizational learning transformation
 UNESCO AI competency framework for school students
(under development)

▪ the responsible users of AI

▪ the active co-creators of AI
▪ the leaders of design and development of next generations of AI
Progression Levels
Competency Aspects
Understand Apply Create

Human-centred mindset
Human Agency Human Accountability AI Society Citizenship

Ethics of AI
What ethics
Embodied Ethicsof AI should weand
Safe help studentsUse
Responsible to learn
Ethicsand practice?
by Design

AI techniques and
AI Foundations Application Skills Creating AI Tools
applications Minimum standard:
*AI system design AI literacy
Problem Scoping Architecture Design Iteration and
Feedback Loops
 To launch
▪ AI competency framework for school teachers
▪ AI competency framework for school students
 4.3 Guide instructional design on uses of GenAI
A “Human-agent and pedagogically proper interactions” template

Required
Appropriate
Potential Appropriate human
GenAI tools
but domains of Expected Requirements pedagogical Possible
and
unproven knowledge outcomes for the users methodologies risks
comparative
uses or problems and example
advantages
prompts
AI Advisors
for
augmented
research
outlines

……
 4.4 Trade-offs and tests of use modes of
Generative AI in teaching and learning

Trade-off: Why should we use? When to use?

• Local(ized) EdGPT for digital textbooks (eg. Republic of Korea)

• Transformers for curriculum co-design (eg. Oak National Academy)
• Formative discriminators for computational skills (e.g. auto-
coaching learning of coding )
• Superchargers for higher-order thinking or inquiry-based learning

Use petrol engines as indicators to test the electric motors?

What sort of tests can lead to the leap innovation in pedagogy?
Does the leap innovation in pedagogy really exist?
What sort of human-digital/AI social contract
should we build?
 A fundamental ask:
Long-term implications of GenAI for education?

Post-AI digital humanity and learning:

Co-existing with human by design

Human-AI-hybrid intensive learning?

Thanks!

[email protected]
https://www.unesco.org/en/digital-education