Week 7 — AI Build Studio · "Record → Transcribe → Summarize"

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objective: Objective 3 — run the record → transcribe → analyze workflow and verify every step · SLO A (execute a multimodal AI workflow) · SLO B (catch transcription errors and summary fabrications)
Worth 50 points · AI Build Studios group = 15% of the grade · Studio 7
Format: a hands-on build — you'll record a short voice memo, convert it to text with a free transcription tool, have an AI produce a summary and action items, and then catch and fix every error the AI introduced.

This is the course's signature weekly component. Every instructional week has one Studio — a real thing to build, a required step where you verify and improve the AI's work, and a short reflection. All tools are free; everything is links to external sites.

---

Part 1 — The Build Goal

By the end of this Studio you'll have:
1. A recorded voice memo of a short simulated meeting or personal planning session.
2. A text transcript of that recording, produced by a free transcription tool — with errors you found and noted.
3. An AI-generated summary and action-item list from the transcript.
4. A verification write-up naming every error the AI introduced — transcript errors and summary fabrications — and your fixes.
5. A short reflection on what the workflow revealed.

This is the record → transcribe → analyze workflow in full, end-to-end, with the verification discipline baked in at every step.

---

Part 2 — Step 1: Record Your Voice Memo

What to record. Record yourself speaking for 60–90 seconds as if you are summarizing a quick planning conversation. Use a scenario that's real or plausible for you. Examples:

• A mini "meeting" with yourself planning a week: what you need to do, key deadlines, and any blockers.
• A simulated two-minute recap of a study group session (you can voice both sides or just the summary).
• A project status update: what's done, what's next, one open question.

The goal: include at least three specific pieces of information the AI will need to capture: at least one date or deadline, at least one action item with a person or role attached, and at least one decision made.

Free recording tools — use whatever's on your device:
- iPhone/iPad: open the built-in Voice Memos app (https://support.apple.com/en-us/101994). Tap the red button to record; tap again to stop. The file saves automatically.
- Android: open the built-in Recorder app (Google Pixel: https://support.google.com/recorder/answer/11420396) or your device's voice recorder. Most Android phones have one pre-installed.
- Mac: open QuickTime Player → File → New Audio Recording → click the red record button.
- Windows: search for Sound Recorder or Voice Recorder in the Start menu.

After recording, listen back once to make sure your three pieces of information are clearly audible.

---

Part 3 — Step 2: Transcribe (and Spot the Errors)

Convert your recording to text using any free transcription method available to you.

Free transcription options:
- Android Recorder app (Google Pixel and compatible devices): if available, it transcribes recordings automatically or on playback. Official instructions: https://support.google.com/recorder/answer/11420396
- Whisper-based web app: OpenAI's Whisper model (openai.com/research/whisper) is available through various free web wrappers. Search "Whisper transcription free" and evaluate the app's privacy policy before uploading personal audio. No specific third-party app is endorsed here.
- AI assistant with audio upload: if you have a paid account on ChatGPT or Claude, you may be able to upload the audio file directly. On most free accounts, use a separate transcription step first.
- Your phone's keyboard dictation: for very short recordings, you can dictate the memo again while typing into a notes app simultaneously — a manual transcription that is deliberately imperfect (great for this exercise).
- iOS Voice Memos with text transcription: on iOS 17 and later, Voice Memos offers transcription. Check under the recording in the app.

What to do with the transcript:
- Read the full transcript carefully against what you said.
- Mark at least two transcription errors — words mis-heard, names changed, numbers wrong, sentences dropped. If your recording was very clean and there are no errors, note that as "clean transcription — no errors found" and move on, but most recordings will have at least a few.
- Save the transcript (copy it into a notes app or document).

---

Part 4 — Step 3: Analyze (Summary + Action Items)

Open any approved AI assistant — ChatGPT (https://chatgpt.com), Claude (https://claude.com), Gemini (https://gemini.google.com), or Copilot (https://copilot.microsoft.com).

Paste your transcript into the assistant and send this prompt (adapt as needed):

"Here is a transcript of a short planning meeting. Please produce: (1) a 3–4 sentence summary of the key points and decisions; (2) a numbered list of action items, each with a person or role responsible and a deadline if one was mentioned. Format it cleanly."

Read the AI's output carefully — this is where the second layer of errors enters. Compare it, line by line, against your transcript.

---

Part 5 — The Verification / AI-Critique Step (required — the BYOAI step)

Now be the judge of the entire workflow. You are looking for errors introduced at two distinct points: the transcription step and the AI's analysis step.

Catch #1 — Transcription errors. List every error from Step 2. For each, write one line: what you said / what the transcript said / your fix.

Catch #2 — Summary fabrications. Read the AI's summary and action-item list line by line. Find things the AI added, changed, smoothed over, or invented that were not in the transcript. Common fabrication types in meeting summaries:
- An action item that was never discussed but sounds plausible.
- A decision described as "agreed" that was actually just floated.
- A detail ("located downtown," "available immediately") that the transcript never mentioned.
- An invented deadline or date.

For each fabrication: write one line — what the summary says / what the transcript actually says / your fix.

Also check for sycophancy in the summary: did the AI make the meeting sound more productive, organized, or decisive than the transcript actually shows? That's a form of embellishment — name it if you see it.

Write 4–6 sentences total reporting: the transcription errors you caught, the summary fabrications you caught (minimum one from each category, or explain why you found none), and your fixes.

The habit this week: a multimodal workflow has more error-entry points than a text prompt — the transcript corrupts, then the AI embellishes. Verify at every step. A summary that sounds polished and professional can still be wrong.

---

Part 6 — Reflection (2–3 sentences)

What surprised you most — about the transcription step, the AI's summary, or both? Would you trust this workflow for something that mattered — a real meeting, a real decision — without the verification step you just did? What would you change about your workflow next time?

---

Part 7 — What to Submit

Submit a single document (or text entry) in Canvas with:
- Your scenario description (what you recorded — no need to submit the actual audio unless your instructor requests it)
- Your transcript (full, with errors marked)
- Your transcription-error catch (Part 5, Catch #1)
- The AI's full summary and action-item list (copy/paste)
- Your summary-fabrication catch (Part 5, Catch #2)
- Your Part 6 reflection

Due Sunday, Oct 18, 11:59 p.m. (50 points)

---

Instructor answer key & model deliverable — REMOVE BEFORE PUBLISHING TO STUDENTS

Students use their own recordings and scenarios, so deliverables vary. Grade the process (all three workflow steps + verification at both steps + reflection), not the content of any specific recording.

Model deliverable (illustrative):
- Scenario: "I recorded myself giving a quick project update — things done this week, next steps, and one question I need my teammate Maya to answer by Friday."
- Transcript produced: "Things done this week: finished the literature review draft. Next step: send it to Maya for feedback. Question for Maya: does the survey need IRB approval before we submit? Deadline: Maya needs to answer by Friday the 17th." [Transcript said "Friday the 7th" — error: '17th' → '7th'. Also missed the word "Maya" the second time, writing "my teammate".]
- Transcription errors caught: (1) "Friday the 17th" → actual spoken date was "Friday the 7th." (2) "my teammate" → should be "Maya" (the name was misheard as a pronoun).
- AI summary produced: "The team completed the literature review and plan to share it with Maya for feedback. The team also agreed to submit the final version by end of month. Maya will confirm IRB requirements by Friday the 17th."
- Summary fabrications caught: (1) "agreed to submit the final version by end of month" — no such deadline was mentioned in the transcript; fabricated. (2) "Friday the 17th" — the AI inherited the transcription error and presented it as fact.
- Reflection: "I was surprised the AI's summary sounded so organized — it even added a deadline I never mentioned. That's the danger: a polished summary signals confidence, not accuracy. I'd only trust this workflow for notes I'd verify before sharing with anyone."

Why the verification step can't be faked: a student who submits a summary without marking any errors and writes "everything looked fine" earns the low end of the AI-critique rows — the rubric rewards the specific catches, not the AI's prose.

Grading rubric — 50 points

Criterion	Full	Partial	None
Recording + scenario — a genuine 60–90 second recording with at least three pieces of information (date/deadline, action item + person/role, decision) (6)	6	3–4	0–2
Transcript produced + transcription errors noted — a full transcript from a real transcription tool (or method) with at least one error found and documented (12)	12	6–9	0–4
AI summary + action items generated — the AI's output is present and clearly from the transcript (6)	6	3–4	0–2
Verification — catches + fixes (Part 5) — identifies at least one transcription error AND at least one summary fabrication (or explains why none were found and shows the evidence), with the original text and the fix for each (18)	18	10–15	0–8
Reflection (Part 6) — a thoughtful 2–3 sentences about the workflow, the errors, and what the student would do differently (8)	8	4–6	0–3

Quality gate: all tool links (Voice Memos, Android Recorder, Whisper, ChatGPT, Claude, Gemini, Copilot) point to official sites only; no invented accuracy percentages or fabricated statistics about transcription tools; the activity requires the student to catch the AI and the transcription, not trust either one; the model deliverable clearly frames the fabrications as errors to catch, not as true statements. Product-accuracy gate: PASS.