1 of 32

AI as Unreliable Narrators in (CS) Education

Benjamin Xie (he/they)

Embedded Ethics Postdoctoral Fellow

Stanford University

benjixie@stanford.edu · benji.phd · @benjixie

2 of 32

Overview

  1. Who I am / What I do
  2. Perspectives on AI (in Education)
  3. AI in K-12 Education
  4. AI & Computing Education
  5. Closing Thoughts on Learning

3 of 32

Activity (3 min): �On the AI Iceberg�benji.phd/kf

4 of 32

I’m a CSEd + HCI Researcher

Computing Education (CSEd)

  • code comprehension, metacognition
  • psychometrics

HCI research

  • participatory design
  • interactive systems

Previously: MIT, UW, Code.org�Currently: Stanford Human-Centered AI Institute (HAI), Ethics Center

5 of 32

I design critical & equitable human-data interactions

Critical: connect data to social structures & history

Equitable: More equal distributions of inputs & outputs

Human-data interactions: learning across boundaries

P(reported non-binary) gets item correct)

P(reported female) gets item correct

benjixie.com/las21

P(reported non-binary) gets item correct)

P(reported male) gets item correct

6 of 32

Assimilating to {unfamiliar, dominant} norms

7 of 32

AI & Tyranny of the Majority

8 of 32

For talk, AI refers to data-driven Gen AI tools

Many kinds of AI tools (“reasoners”): such as predictors, classifiers, recommenders, planners, generators

Focus of this talk on Generative AI tools built upon Large Language Models (e.g. ChatGPT for text generation, DALLE-2 for image generation)

9 of 32

AI is built by select few to maximize profit

AI tools tend to perpetuate dominant social norms.

  • Hidden human biases, exploitation, e.g. data annotators (Casper et al. 2023)
  • Homogenization of language & culture (Bender et al. 2021, Blodgett & O’Connor 2017, Cao et al. 2023)
  • Stereotypes getting more subtle, still harmful, e.g. disabled communities (Gadiraju et al. 2023)
  • “Progress” measured by benchmarks (which AI trains on)�(Raji 2021, Blili-Hamelin & Hancox-Li 2023)

Sasha Luccioni

10 of 32

Which images are more “racy”?

11 of 32

Example: “Racy” Images & Shadowbanning

AI tools rate photos of women as more sexually suggestive

  • “Racy” reflect values of human annotators (underpaid men in Asia, Africa)
  • overfitting on bras, pregnant bellies, exercise

Impacts:

  • Shadow-banning images w/ high racy score
  • limitation of expression (e.g. pregnant women, pole dancers)

12 of 32

“AI in education is a public problem” -Ben Williamson

  • hype >> proven benefits
  • oracle/guru >> unreliable narrator
  • opacity >> responsibility
  • standardization >> cultural-responsiveness
  • business needs >> learner needs

13 of 32

Equity (in CS Edu)

Def’n: equal & adequate distribution of inputs (resources, opportunities) or outputs (growth, outcomes) �Levinson, Gergon, & Brighouse 2022

Equity in CS Edu:

  1. Mitigating preparatory privilege �(Margolis & Fisher 2002)
  2. Contextualizing student data
  3. Pluralizing CS

▲resources

◼️opportunities

⬟growth

⬢outcomes

more challenging

ECEP Alliance

14 of 32

EdTech & Inequities

"I believe that the motion picture is destined to revolutionize our educational system" -Thomas Edison, 1922

Same tools used differently by different schools.

(Ed)Tech tends to perpetuate and/or exacerbate existing social inequities

15 of 32

AI in K-12 education

16 of 32

Many framings of AI

Model-Centric AI: improve model performance while treating data as static, objective

Data-Centric AI: improve model & data iteratively (Jarrahi, Memarian, & Guha 2023)

Human-Centered AI: social and ethical implications of AI on individuals, communities, societies

17 of 32

Ex: Learning facial recognition…

Model-Centric AI: Different techniques

Data-Centric AI: Imbalances in data result in performance differences by gender, skin tone

Human-Centered AI: Why Black communities do not want facial recognition in policing

Joy Buolamwini / Gender Shades. 2018

Pew Research 2021

India Times, 2017

18 of 32

“Dual Exploration” w/ AI & Disciplinary Learning

19 of 32

Co-Design w/ Cross-Disciplinary Teachers

Study w/ 8 U.S. high school STEM & humanities teachers teachers. Most from Title I schools which served Hispanic and Latinx, Black and African-American students.

Goal: Understand teachers’ perspectives on AI, opportunities & limitations in using Gen AI tools in class

20 of 32

Perceived Dichotomy: Technical vs Ethical

you can sort of think about [AI] is a universal system....In my engineering class, we do a whole unit on universal systems... So I guess it's in context of what you're trying to bring [AI] into. So for [English Language Arts]... the learning objectives are a little different. I like this lesson, because... they're learning the technical things... And less about, like, `how to make an argument' or `what ethics are.'”

- Engineering teacher serving Latinx students

21 of 32

Opportunities: AI as augmentative, empowering

Empowering: Enable students with less art skills to create and share art

Augmentative: ChatGPT to convey students’ ideas in more professional tone

21

“How can this algorithm can assist [students] in writing a more professional business plan?”

- Econ teacher serving Black/African- American students

“could have like a competition or something, and hang their pieces in the hallway. So everyone can say, `hey? I created this,' because not everyone's an artist. But I look at [DALL·E] as a tool to help them create art.

- CS (web development, augmented reality) teacher serving white students

22 of 32

Engaging with Limitations of AI

Limitations to reflect on affordances

  1. Consider tradeoffs of using AI (or not)
  2. Identify improvements to AI tools

Alchemy of prompt engineering as hindrance?

  1. Students bothered by imperfections in generated images?
  2. Generating art seems easy, but prompt engineering is hard

22

�“the discussion I want to have is is [AI generated art] even close? And then to hit ‘ways in which AI algorithms would need to improve in order to …become stronger…’ [For example,] hands have got to be better. Anatomy's got to be better. This is not going to be a good resource until X, Y, and Z.”�- animation & digital media teacher

“I think I would like [students] to get started with… drawing it themselves first, so they don't get bogged down with trying to find the perfect wording and then get really frustrated and be like oh, `I can't create what I want to create, because I don't have the words for it,'” �- art teacher serving Hispanic students

23 of 32

Co-designed, Cross-disciplinary, Critical AIEd

Bridging technical and ethical/social perspectives of AI

  • Future: Discipline-specific pedagogy, teacher preparation�

“Dual exploration:” AI learning to support disciplinary learning (and vice-versa)

  • More cross-disciplinary co-design!�

Designing to make limitations of AI tools part of learning experiences

  • Tools to make GenAI tools more suitable for educational contexts

24 of 32

AI, CSEd, and Equity

25 of 32

Many Types of Programmers

CS courses often target professional programmers

Many program as part but not all of careers (end-user programmers, Nardi 1993)

Many more want to be engage in computing discourse (conversational programmers, Chilana, Singh, & Guo 2016)

26 of 32

Ethical & Socials Lens to CSEd

Pluralizing CS by teaching about computing through technical, social, and ethical lenses

At Stanford, I embed ethics into CS courses

  • representing names, gender, people in code
  • algorithms relying on commensurable factors ($, time)

ethics content

students’ prior knowledge,

beliefs, expectations

prior CS knowledge

My approach to embedding ethics in CS:

27 of 32

GenAI tools changing what to teach

Tools can generate syntactically correct code instantly

=> Less emphasis on code writing?

Greater emphasis on reading, comprehending, debugging, verifying correctness?

28 of 32

GenAI tools risk over-reliance

Learners have greater risk of over-reliance on AI tools �(Perry et al. 2022, Jalil et al. 2023, Wang et al. 2023)

Existing tools do not convey uncertainty, afford customizability, reflect dominant social norms

=> must also teach self-regulation to align AI usage with goals

CS expertise

risk

29 of 32

Pluralizing CSEd & AI

Many different programmers, different skills to teach, ways to incorporate AI

Mitigate risks of over-reliance on AI tools by teaching right skills (self-regulation, debugging)

conversational

end-user

professional

types of programmers

Model-Centric AI: improve model performance while treating data as static, objective

Data-Centric AI: improve model & data iteratively

Human-Centered AI: social and ethical implications of AI on individuals, communities, societies

different framings of AI

30 of 32

Closing Thoughts

31 of 32

Learning isn’t individual!

Learning process is one in which learners can realize their human dignity and potential (to participate in existing systems, challenge oppressions) without enduring a process of dehumanization.

Learning computing provides learners with an opportunity to engage with powerful tools that they can use to unseat individual and collective social oppressions.

32 of 32

AI as Unreliable Narrators in (CS) Education

Takeaways:

  1. Co-design curriculum, PD, tools to realize more equitable use of GenAI, sociotechnical understanding
  2. Embrace different framings of AI to pluralize engagement in tech communities

Ongoing and future work

  1. Planning workshop on “AI & Advocacy"
  2. Understanding AI vs student performance on CS assessments
  3. Building youth advocates capacity to use data to support env. advocacy

Recommended Reading:

  1. AI in education is a public problem, Ben WIlliamson 2024
  2. The mounting human and environmental costs of generative AI, Sasha Luccioni
  3. Inspiring Action: Identifying the Social Sector AI Opportunity Gap, Stanford HAI & Project Evident
  4. AI and the Everything in the Whole Wide World Benchmark, Raji, Inioluwa Deborah, Emily M. Bender, Amandalynne Paullada, Emily Denton, and Alex Hanna. 2021.

conversational

end-user

professional

types of programmers

Model-Centric AI: improve model performance while treating data as static, objective

Data-Centric AI: improve model & data iteratively

Human-Centered AI: social and ethical implications of AI on individuals, communities, societies

different framings of AI

Benjamin Xie (he/they) Stanford University

benjixie@stanford.edu · benji.phd · @benjixie