This is the Vol. 21 capstone. You have now seen applied product categories from chatbots through regulated verticals (healthcare, finance, legal). Education AI closes the volume on a different integrity axis: tutoring that teaches versus assessment that must remain valid. The stack reuses chat, research/citation, and sometimes coding assistants—but the product claim and mode switch change everything.
Vol. 20 privacy (especially minors), fairness, and transparency apply. This lecture is educational engineering, not a surveillance kit and not instructions to cheat. After the recap, Vol. 22 leaves product patterns and enters the ecosystem: providers, APIs, and tools—starting with OpenAI.
Learning Objectives
By the end of this lesson, students should be able to:
- Split tutor mode (hints, Socratic questions, similar examples) from assessment mode (rubrics, no silent answer keys).
- Design honor-code / disclosure / mode-switch product controls—not spyware.
- Apply Vol. 20 privacy (minors, FERPA-style themes) and fairness to tutoring and scoring.
- Reuse hallucination tests when tutors cite sources (Vol. 19 + research assistants).
- Recap Vol. 21’s 13 product categories as a pattern library (HITL, eval, privacy).
- Explain what Vol. 22 adds: companies, APIs, and tooling ecosystems rather than new product types.
Education AI means AI features in teaching, tutoring, content authoring, accessibility, and assessment support. Tutor mode helps a learner understand: hints, questions, worked similar problems, citations. Assessment mode helps an instructor evaluate under a declared policy: rubric scoring, feedback bands, integrity policy logging—not a hidden answer key to the live exam, and not device monitoring malware. Assessment integrity is the property that scores still reflect the learner’s own work under the course rules. Product design (mode split, disclosure, honor acknowledgment) is the lever taught here—not cheating techniques and not covert surveillance.
Do not build keyloggers, screen grabbers, or covert monitors “for integrity.” Honor codes, clear AI-use policies, timed authentic assessments, oral defenses, and separate tutor/exam environments are the legitimate patterns. Vol. 20 security stays defensive. This page does not teach students how to cheat or how to spy on them.
Tutoring vs Assessment Integrity
| Tutor mode | Assessment mode | |
|---|---|---|
| Goal | Learning / practice | Valid measurement under policy |
| Typical outputs | Hint, Socratic question, similar example, source cite | Rubric scores, feedback, flag for human TA |
| Answer keys | May show after attempt or on practice sets | Must not silently emit live-exam keys to the candidate |
| Disclosure | “You are using an AI tutor” | Honor-code ack + allowed-AI policy visible |
| HITL | Teacher configures curriculum + guardrails | Instructor/TA reviews contested scores |
| Failure mode | Hallucinated facts; doing the thinking for the student | Invalid scores; unfair false integrity flags |
Legitimate integrity levers
- Separate practice vs exam environments
- Honor-code acknowledgment + syllabus AI policy
- Authentic tasks (oral, in-class, process portfolios)
- Human review of automated flags
- Fairness slices on scoring error (Vol. 20)
Out of scope / harmful
- Covert keylogging or secret screen capture
- Silent exam-answer oracles marketed as “tutors”
- Cheating how-tos in the product
- Unappealable automated misconduct verdicts
- Training vendor models on minors’ chats by default
Privacy (Vol. 20)
- Minors: extra minimization and consent themes
- School records: purpose + retention + access
- Do not put full student PII in prompts
- Export/delete includes tutor logs + embeddings
Mode split buys
- Tutors can be generous; exams stay valid
- Clearer eval: learning gains vs rubric agreement
- Honest disclosure (transparency)
One-chatbot-fits-all costs
- Practice bot becomes a live-exam oracle
- False cheating accusations without appeal
- Hallucinated curriculum presented as fact
Tutor / Assessment Gate Sketch
Policy and mode flags for a system you operate. Not a cheating tool. Not endpoint spyware.
Vol. 21 Recap → Vol. 22 Ecosystem
Vol. 21 was a product-category volume: how to shape models, RAG, agents, eval, and Vol. 20 controls into something you can ship. The table is your pattern library—not a vendor catalog (that is Vol. 22).
| Lecture | Product idea to keep |
|---|---|
| Chatbots | Scoped dialogue + state; not an unbounded oracle |
| Customer support | Escalate; deflection \(\neq\) resolution |
| AI search | Ranking + grounded snippets |
| Document AI | Extraction with schema + review |
| Voice assistants | Speech UX + confirmation on side effects |
| Email automation | Draft/HITL before send |
| Workflow automation | Graphs with gates; no silent money/PII writes |
| Coding assistants | Editor context budget + HumanEval-style eval + human diff |
| Research assistants | Citation-first RAG + Vol. 19 hallucination tests |
| Healthcare AI | HITL + audit; not medical advice |
| Finance AI | Explain/flag; not investment/credit advice |
| Legal AI | Counsel-assist + citation lock; not the practice of law |
| Education AI (this) | Tutor vs assess; integrity via policy, not spyware |
Vol. 22 — AI Companies & Ecosystem answers who sells the substrate: frontier LLM providers (starting with OpenAI GPT / DALL·E / Whisper / Codex / Sora / API), then Google, Anthropic, open weights, image/video/voice platforms, cloud AI platforms, coding IDEs (22.6), and automation vendors. Read Vol. 22 with Vol. 21 in mind: a Copilot or GPT API is a component; your product still needs HITL, eval, and privacy. Pricing / strengths / weaknesses comparison is a recurring lecture format across providers.
Metrics & hallucination tests
Privacy, fairness, compliance, security
Product categories you can ship
Providers, APIs, ecosystem tools
Related Lectures
| Lecture | Role |
|---|---|
| Legal AI | Previous regulated sibling |
| Chatbots / Research assistants / Coding assistants | UX stacks education reuses |
| Privacy / Fairness / Transparency | Minors, score equity, AI disclosure |
| Hallucination tests / Human evaluation | Cited tutors + rubric agreement |
| HITL / Guardrails | Teacher config + mode policy |
| Vol. 22 OpenAI | Next volume: ecosystem starts |
“A good tutor should also complete the live exam.” That collapses integrity. Second: integrity requires covert device monitoring. Third: automated plagiarism/AI-use flags are ground truth without human appeal (fairness harm). Fourth: education products are exempt from privacy because “it is just school.” Fifth: Vol. 22 replaces Vol. 21—providers do not erase HITL/eval/privacy. Sixth: this lecture teaches students how to cheat.
Knowledge Check
- Short Answer: What is the core product split in education AI? Answer: Tutor mode (learning) vs assessment mode (valid measurement under policy).
- True/False: Assessment mode should silently give live-exam answer keys to the candidate. Answer: False.
- Multiple Choice: Legitimate integrity design includes: (a) honor code + mode split + human review, (b) covert keyloggers, (c) guaranteed cheating guides. Answer: (a).
- Short Answer: Name one Vol. 20 concern that is especially sharp for K–12 tutors. Answer: Privacy/minors (also fairness or transparency acceptable).
- True/False: This lecture teaches how to cheat or how to install spyware. Answer: False.
- Multiple Choice: Vol. 22 primarily adds: (a) provider/ecosystem/API landscape, (b) replacing HITL, (c) medical licenses. Answer: (a).
- Short Answer: Which Vol. 21 lecture taught citation-first RAG that tutors should reuse when citing a KB? Answer: Research assistants.
- True/False: Unappealable automated misconduct verdicts are a fairness/HITL failure. Answer: True.
- Multiple Choice: Coding-assistant eval in Vol. 21 was conceptually: (a) HumanEval-style pass@k + product telemetry, (b) only BLEU, (c) only GPU temp. Answer: (a).
- Short Answer: What is the first Vol. 22 lecture this capstone points to? Answer: OpenAI (GPT / DALL·E / Whisper / Codex / Sora / API).
Key Takeaways
- Education AI: tutor vs assess; integrity via policy and mode design, not spyware or cheating tools.
- Reuse chat, research citation, coding-assist patterns—with privacy for minors and fairness on scores.
- Vol. 21 recap: thirteen product categories sharing HITL, eval (Vol. 19), and Vol. 20 controls.
- Regulated verticals never claim to be the doctor, fiduciary, or lawyer.
- Next volume: Vol. 22 OpenAI and the broader ecosystem.
Lab: Students specify tutor vs assessment configs for one course: allowed tools, honor-code copy, rubric-HITL, privacy TTL for minors (synthetic data only). Explicitly list three non-goals (no keylogger, no live-exam oracle, no unappealable AI misconduct verdict).
Whiteboard: Full Vol. 21 map (13 boxes) \(\to\) shared spine: RAG/HITL/eval/privacy. Arrow out to Vol. 22 provider logos as components. Cap the volume: you can now name the product before you pick the vendor.
Recap: Education AI caps Vol. 21 by splitting tutoring from assessment integrity and recapping the product library. Vol. 22 begins the ecosystem with OpenAI.