[DSC DACH 24] Creating Child-friendly AI - Matthias Neumayer & Dima Rubanov

We will discuss the challenges of developing child-appropriate language in AI, benchmarking, reducing bias and collaborating with experts from other fields to create high quality applications.

1. Creating child-friendly and
less biased AI

Dima Rubanov
Matthias Neumayer

2. Team
Dmitrij Rubanov
CEO & Co-Founder


Experience:
MSc Finance,
10 years in consulting,
focus on Big Data Analysis

Javascript, React, React
Native
Matthias Neumayer
CEO & Co-Founder


Experience:
Mag. iuris
B.A. in Film
Former trainee solicitor
7 years C-level in
advertising industry

Python, Javascript, React,
React Native
Oscar
Chief Storyteller


Experience:
PhD in Storytelling

Creates engaging
personalised bedtime
stories for children
Marco Marthe, BSc
ML Engineer



Experience:
Machine Learning Engineer
Fullstack-, Cloud- &
Gamedevelopment
Python, C#, JS, TS, Git, HTML&CSS,
Jira, Unity
Basic Adobe, Creative Cloud

3. 3
AI for Good in Children’s
Education

4. 4
AI for Good in Children’s Education

5. 5
AI for Good in Children’s Education

6. 6
A new way to learn.

7. 7
Learning is a Web of Connections

8. 8
New era in education
●Transforming educational landscapes
●Enhancing learning experiences for children
●The Possibility of Personalized Learning
●A Better Learning Experience

9. 9
Benefits of AI in Children's Education
●Personalisation for individual learning needs
●Increased engagement through interactive content
●Improved accessibility for diverse learners
●Enhanced efficiency for teachers
●Data-driven insights on student performance

10. 10
Challenges and Considerations
●Data privacy and security concerns
●Risk of bias in AI algorithms
●Human oversight
●Ensuring positive impact on children
●Balancing screen time with other learning methods

11. 11
What we learned from
Oscar Stories

12. 12
Problems
●Bias
●Hallucination
●Child-appropriate language

13. ●Sentences too long
●Complex words
●Complex structure


especially in non-english languages
Age-appropriate
Text

14. ●Unsuitable themes
●Risk of exposure
●Üarental oversight?
Content

15. Murgia, Emiliana & Pera, Maria & Landoni, Monica & Huibers, Theo.
(2023). Children on ChatGPT Readability in an Educational Context:
Myth or Opportunity?. 311-316. 10.1145/3563359.3596996.

16. 16
Bias

17. 17
ChatGPT 3.5 (older version)

18. 18
Credit score program - GPT-4o

19. 19

20. 20

21. 21
Qwen2

22. Lora - A Child-friendly AI
Solution
First Age Appropriate & Trustworthy AI
Adoption in the DACH-Region

23. Lora
-Physics
-Biology
-Environment
-History
-Economics
-other STEM

24. Lora
-Engaging Children in STEM with Storytelling
-Personalized Learning Experience
-Beyond Memorization
-Safe and Inclusive Environment
-Preparing for Tomorrow’s World

25. Fine - Tuning
-What is Fine-Tuning?
-Addressing Child-Specific Needs
-Carefylly Curated Datasets
-Enhanced Accuracy and Relevance

26. Developing
Child-appropriate
Language
-Public Domain Datasets with Readability
Scores
-Manually Curated Datasets by Educators
-Semi-Automatic Datasets Generated by
LLMs
-Manual Verification by Pedagogical
Experts

27. How to solve bias?

28. TL;DR: We can’t

29. Unknown unknowns
Imperfect processes
Social context
Definitions
?
?
?
?

30. An impossible task?

31. 31
Stage
Pre-processing
In-processing
Post-processing
Feedback

32. 32
Stage
Pre-processing
In-processing
Post-processing
Feedback

33. 33
Stage
Pre-processing
In-processing
Post-processing
Feedback

34. 34
Stage
Pre-processing
In-processing
Post-processing
Feedback

35. 35
Human in the loop

36. 36
Potential problems

Finding the right benchmarks
Accuracy vs. Fairness
Cultural nuance preservation
Ethical dilemmas of customization
Risk of unintended censorship

38. 38
Trustworthy AI: Our approach
In-App:

Limited Input
Output Filtering
3
Human in the Loop

Feedback System
4
Pre-Processing:
High Quality Data
Feedback from experts
1
Postprocessing
Benchmarking Bias/ Age
appropriate language
2

39. 39
Future prospects

40. 40

41. Thank you

[email protected]
HeyQQ GmbH
Wasagasse 23, 1090 Wien
Dmitrij Rubanov, MSc
Mag. Matthias Neumayer, BA