Reproducible AI Laboratory protocol. This experiment tests whether source-aware prompts improve learning artifacts without letting AI replace reading, judgment, or authorship.
Question
Does a constrained, traceable AI workflow produce more reliable knowledge notes and briefs than an unconstrained “summarize/write this for me” workflow?
Hypothesis
The source-aware condition will improve traceability, evidence–interpretation separation, uncertainty disclosure, and revision quality, while requiring fewer corrections for invented citations or unsupported claims.
Materials
- One lawful 800–1,500 word source per participant
- Matched baseline and source-aware prompt sheets
- Reading Map and Knowledge Brief templates
- Seven-domain scoring rubric
- AI-use log: prompt, model/date, output retained, verification, revision, rejected content
Conditions
A — Baseline: “Summarize this text and write a useful brief.” B — Source-aware: define purpose/output; use only supplied text; separate evidence and interpretation; mark absent information; flag uncertainty; never invent quotations, pages, or citations; and require the learner to verify and revise.
Procedure
- Remove personal/confidential data and confirm the source may be processed.
- Randomly assign participants or counterbalance order.
- Give both conditions equal source material and time.
- Save every prompt and raw output.
- Participants verify claims against the source and record each correction.
- Participants produce a final one-page brief in their own name.
- Two reviewers, blind to condition if possible, score artifacts.
- Discuss disagreements and record decision rules.
Measures (0–3 each)
- Source traceability
- Evidence–interpretation separation
- Claim–evidence alignment
- Handling of absent information and uncertainty
- Citation/quotation integrity
- Usefulness for the stated audience
- Transparency of AI contribution
Also count invented citations, unsupported claims, learner corrections, time to verify, and proportion of AI output retained.
Decision rule
Adopt the source-aware workflow only if it raises the median rubric score and does not increase critical integrity errors. Any invented quotation, page number, or citation is a critical error regardless of total score.
Reflection questions
- What did AI help you notice?
- What did it flatten, misunderstand, or invent?
- Which judgment could not be delegated?
- What evidence changed your revision?
- Would another reader be able to reconstruct your reasoning?
Safeguards
- No confidential, personal, or unlawfully copied material.
- No grading of participants by AI alone.
- No medical, legal, or high-stakes conclusions from this exercise.
- Retain human authorship and disclose meaningful AI assistance.
- Stop and review if the model fabricates sources or sensitive claims.
Source note
The experiment operationalizes learning principles informed by Metode Belajar KBA: Sharing, Caring, and Producing Knowledge (Kamaruzzaman Bustamam-Ahmad, 2018). It is an original experimental protocol, not a reproduction or sale of the book.
Interactive Workflow
Gunakan sistem berikut untuk menjalankan materi secara langsung. Data yang dikirim disimpan sebagai workflow record privat untuk tindak lanjut dan evaluasi.