Declarative Narrative (ExLENT/ITEST/DR K-12)

Need / problem / solution

Even though a lot of the public discourse around AI is hyperbolic and poorly-informed, it is clear that AI is becoming more capable at doing knowledge work. AI technologies are changing our ideas about the nature of knowledge, intelligence, and human dignity. AI technologies and epistemic shifts are also likely to impact social structures, labor markets and individual career prospects, as work that people were once paid to do is instead done more cheaply by computers. If we believe K12 education should prepare youth for social, policital, and economic participation, then there is a strong case for including AI in the K12 curriculum in some way.

But if we did want youth to graduate high school knowing something about AI, what we would them to know? And how might we teach it? These two fundamental questions run parallel to questions about computational thinking (CT) which were the focus of two National Research Council workshops [@nationalresearchcouncil2010, national2011report] a decade ago, and which remain largely unanswered. Framing learning goals and pedagogies for AI is probably more difficult than for CT, and depends on the latter. As a result of this lack of clarity, we are seeing a predictable frenzy of ed-tech products, curricula, and services, which do not explain what learning outcomes they are aiming at or how they purport to teach them.

Following @kafai2020, I think we can distinguish AI learning goals into cognitive, situated, and critical framings. Whereas the cognitive framing is predominant in K12 computer science [@mcgill2019], very little cognitive AI is currently proposed, probably because of the multiple advanced prerequisites, including linear algebra, statistics, calculus, algorithms, and programming. (This is not to say cognitive approaches to AI are impossible in K12; it’s worth keeping in mind Bruner’s claim that “Any subject can be taught effectively in some intellectually honest form to any child at any stage of development” [@bruner1960].) Situated approaches to AI tend to follow the success of Scratch, encouraging learning through creating and sharing personally-meaningful projects. Critical approaches to AI focus on critiquing the processes encoded in AI tools and their products. For example, Joy Buolamwini’s film “Coded Bias” explores the reasons why computer vision systems have tended to be much worse in recognizing dark-skinned people’s faces, and how AI tools such as computer vision reproduce racial discourses.

The problem I see is that these framings are disconnected from each other. The situated approaches encourage students to make things with AI without transforming the tools, or even learning much about how they work. The critical approaches focus on sociopolitical crititiques which are not grounded in technical understandings or evidence, leaving students with no basis for distinguishing strong arguments from weak arguments, or accurate evidence from inaccurate. Much like an iPad, the interface feels magican and its innards are hermetically sealed. What’s missing from current AI education proposals is the connections between the screen and its underlying infrastructure, what Tom Liam Lynch calls “sub-screenic literacies” [@lynch2019], in which students learn to read not just what’s on the screen but the systems underlying them.

In my dissertation, I developed Unfold Studio, a web application for interactive storytelling. One essential learning mechanism in Unfold Studio is traversing the stury’s narrative interface. The narrative structure has excellent potential for engaging with situated identities and communities, as well as critical questions about power and privilege. But Unfold Studio’s narrative interface is an unstable, leaky abstraction, a “half-baked microworld” [@kynigos2007]. Users are encouraged to look behind the curtain at how the code works, to mess with it, and then to create their own [@lee2011].

Could we extend interactive storytelling to teach AI in a similar way, simultaneously learning something about the tool’s inner workings, while using it to create personally-meaningful artifacts for an authentic audience, and gaining critical consciousness of the tools and the worlds in which they are deployed? I imagine a system I am calling “Declarative Narrative” in which users collaboratively author narratives with a large language model such as ChatGPT, by delaratively describing plots, settings, and characters with which the player will interact. The author’s job would be more like a film director (or, more accurately, a video game director), who designs audience’s experience and describes how it should unfold, and then leaves the implementation details to others.

I am planning an NSF submission this fall related to these ideas. I would like to discuss together the premises above, as well as any related precedents and ideas you have about how the system might work.

DECLARATIVE NARRATIVE

Building off of Will Wright’s talk:

• computers model structure and process

• How to engage with AI. I’ve been uncomfortable with it because I don’t know what the speaker things they see.

• Difference between simulation and game design

  • Problems come up in game design, not in simulation. The simulation is just enacting algorithmic processes.

• The metaphor of the ball: The ball doesn’t make a game.

• ChatGPT just killed off open-ended interactive storytelling (Discuss Twisty Little Passages, etc.)

  • It’s not fun to make lifelike discourse anymore. (The bitter lesson…)

• But it could make explicit, narrowly-scripted narrative more valuable than ever before, for making places within discourse spaces.

  • For pursuing explicit learning goals with respect to AI. But what would those goals be?

• How might we think about user interfaces for AI?

  • Traversable abstractions.

• Microworlds for AI?

  • What is a microworld? A playground where you learn the rules through play. But when there is a closed abstraction, you can play without learning much. For example, an iPad.
    • Half-baked microworlds. (XX)
    • Traversable abstractions. (me)
    • “Sub-screenic literacies” (Lynch)

I’m concerned that “AI Literacy”-based learning will be unable to traverse the abstraction.

For school more broadly: How might we make failure fun?

RETTL Proposal

• Find English teachers to work with (Corey?)

(declarative narrative!)

• Create characters?
• Create settings?
• Interaction of needs, characters, settings I manage state!

Notes from Lab Meeting 2023-02-16

Finn: This didn’t go in the dirction I expected it to go in teaching about machine learning. Grace: AI technologies will affect social structures, labor markets, etc. But what is the purpose of teaching AI? How can AI prepare the young for social, economic, and political participation? If you are learning computational thinking, you are learning how to solve a real problem.

Do we want people to know how to use AI? Do we want people to know how to build AI?

In the process of creating the narratives, perhaps that is a little microcosm of a lot of the math that we’re looking at.

How much is the goal the techniques? How much is it the models?

ChrisP: Are we looking for CT in AI?

ChrisH: Not sure if this makes sense, but I kind of feel an analogy to critical data science education. purposes and aims vs. methods and techniques. Data Feminism. What’s the power in gathering, etc. Finn: Maybe these questions belong in the classroom.

Isaak: Math Solver tools