As we build out CyberScholar (an all-new version of CGScholar, with embedded AI), we've been pondering design principles for AI in education. Here's our first draft - feeback welcome!

1. Cyber-Social Relations

AI and humans are always in dialectical relation, two fundamentally different kinds of intelligence interacting. There should be no AI without human moderation. For instance, AI feedback must be accompanied by metacognitive self review and reflection on the differences between the knowledge creator’s human perspective and AI perspectives.

2. Learner Differences

AI education systems should be tuned to learner differences, changing the register of their interactions with learners according to learner responses, and over time developing a finely calibrated understanding of the learner’s interest, needs, and potential learning trajectories.

3. Deep Thinking

Teaching with AI should be oriented to complex epistemic performance—not the factoids, yes/no answers, or the low-level memory horizons of select response tests. Learning requires the construction of holistic knowledge artifacts. For instance: a science project report, literary interpretation, historical explanation, makerspace design, worked math problem, etc.

3, Multimodality

Student knowledge representations should demonstrate capacities for transposition of meanings between text, image, space, object, body, sound, and speech. For instance, all visual or video media must be accompanied by textual labels and/or exegesis both to reveal thinking and for AI analysis

4. Collaboration

AI collaboration should be complemented by human-human collaboration, such as peer feedback and/or teacher presence. In the case of asynchronous learning, this will require a social threading architecture where students interact with people who have recently reached the same point in a learning progression or soon will.

6. Hybrid Delivery

In-person, online, synchronous, and asynchronous delivery formats should all be based on the same principles and grounded in shared architectures, supporting “hyflex” movement across delivery modes.

7. Co-Design

Teachers, learners, and the pedagogical and collective intelligence curated in AI become co-designers (always!) of new knowledge.

8. Model Declaration

With each student-created artifact, systems should automatically declare both to the student and the teacher: a) the selected or curated specialized Knowledge Base for RAG/CAG; b) the selected or created the Rubric Agents; c) Foundation Model.

9. AI Transparency

With each student-created artifact, systems should automatically declare how much and in what ways the learner used AI to extend (or reduce) their epistemic effort.

10. Assess-as-You Go

Granular analytics means that all assessment is formative, and summative assessment is no more than retrospective progress view on data that was in the first instance formative. Assessment is embedded to the extent that there is no temporal or artifactual distinction between instruction and assessment.

11. Knowing what Learners Know

The evidence of having learned is not a cursory B+ or 72%, but knowledge artifacts. AI analytics can analyze these for zones of proximal knowledge where further learning might most productively be situated. Published to a student portfolio, education institutions and employers will get a fuller picture of a learner’s capabilities than transcripts or college application letters.

Watch an an overview of CyberScholar alpha v. 0.5.1 here: