The AI-resistance method

The problem with the take-home essay

A generation of English teachers has built their assessment practice around the take-home essay: read this, write 500 words on it, hand it in Friday. The reasoning was sound. Sustained writing forces sustained thinking. The form is also durable — students have been writing essays since the 16th century, and the questions teachers asked in 1995 mostly still worked in 2015.

By 2023 they had stopped working. A student with a phone can produce a competent five-paragraph essay on Robert Frost in under two minutes. The essay no longer measures what it used to measure. Teachers know this and are tired of being told to “just check for AI” with tools that don’t work, or to lean harder on in-class writing that doesn’t fit the period schedule.

qed’bop is one answer to a smaller question: what shape of assignment is hard for a chatbot to do well, without being hard for a human to do at all?

Music as the irreducible thing

A poem on a page is text. A chatbot has read every poem ever published and every essay about every poem ever published. Asking a student to write 300 words about Frost’s “Stopping by Woods” is asking them to do something the machine has done a million times.

A poem set to music is not text. It’s a recording — an unrepeatable interpretive act by a specific musician, made of arrangement choices, tempo choices, breath, instrumentation, silence. Most musical settings of public-domain poems are obscure: small ensembles, art-song traditions, folk recordings on labels nobody has scraped. Even if the audio is online, what the machine has indexed is the metadata and maybe a transcript — not the interpretive work the performance is doing on the poem.

So the lesson stops being “write about this poem.” The lesson becomes “listen to this performance of this poem and tell me what the music argues that the words alone don’t.” That’s a question a student with headphones can answer in twenty minutes. It’s a question a chatbot can’t answer without hearing the recording, which it can’t do.

The four rules

Every assignment qed’bop generates obeys four rules, baked into the system prompts that produce the discussion questions. They’re simple and load-bearing.

1. No timestamps, no invented specifics. A qed’bop question never says “at 1:20 the bass drops out” or “in the second verse the singer whispers.” Questions about music are always general: “identify a moment where the instrumentation changes unexpectedly and explain its interpretive effect.” The student supplies the specificity from their own listening. That specificity is the proof of engagement.
2. At least one “find and describe.” Every assignment includes one question whose structure is: locate a specific moment, describe what you heard, explain its interpretive effect. This is the structural moat. A response that doesn’t cite a moment can’t answer it; a chatbot that hasn’t heard the recording can’t fake a moment without inventing one, and an invented moment is immediately falsifiable.
3. Engage the interaction, not the poem alone. Questions are about the relationship between the words and the arrangement — what the music emphasizes, strips, contradicts, extends. Pure poem-analysis questions get filtered out at generation time. The whole point is the pairing.
4. Audience-calibrated. A middle-school question and a graduate-seminar question are different shapes. qed’bop tunes vocabulary, abstraction level, and expected response length to the audience you pick. Middle-school questions are concrete and observable; graduate-level questions invite theoretical framing.

How a lesson actually runs

The intended shape of a qed’bop lesson is one class period, roughly fifty minutes. The teacher opens the share URL on the projector; students open it on their phones or laptops. There is no login, no class roster sync, no “assignment” in the LMS. Just a URL.

• Minutes 0–5. Read the poem aloud, once, before the music. Don’t discuss yet.
• Minutes 5–20. Play the recording once through, uninterrupted. Students take loose notes — moments that surprised them, words the music slowed down on, instruments they didn’t expect. Then play it again with the lyric sheet, and ask them to mark which parts of the text the arrangement emphasized and which it raced past.
• Minutes 20–40. Open the questions on the shared URL. Students work in pairs or individually. The “find and describe” question is the one that needs specific listening — give it real time. Other questions can be done aloud as a full-class discussion.
• Minutes 40–50. Debrief. The richest conversations come from comparing what different students heard in the same moment. That comparison is what no chatbot can participate in, because the comparison is between human ears.

What if a student uses a chatbot anyway?

They will. That’s the realistic baseline. The honest claim qed’bop makes is not “chatbots cannot do this” — it is “the part of this assignment a chatbot can do is worthless without the part it can’t.”

A student who pastes the questions into a chatbot will get back confident, fluent prose about the poem and about music in general. What they will not get is a specific moment from a specific recording, accurately cited, with an interpretive claim grounded in it. The “find and describe” question filters this cleanly: an answer either points to a moment a teacher can verify by replaying the recording, or it doesn’t.

When a chatbot does invent a moment, the invention is checkable in seconds. The teacher plays the recording at the cited timestamp and the claim either holds or it doesn’t. The asymmetry between “a student writes a real answer in twenty minutes” and “a teacher catches a fabricated answer in thirty seconds” is what makes this work pedagogically.

A note on the teacher edition

Every share URL has a twin at /t/... that only the teacher uses. It contains the same poem and the same questions, plus an AI-generated lesson agenda, poet biography, historical context, and per-question teaching commentary — material that can quietly include specific musical moments and timestamps, because it’s for the teacher’s eyes only. Student-facing content never sees those details. The two-track design is the structural commitment to rule one.

Where this goes next

qed’bop is built on the bet that the assignments that survive the next decade of language models will be assignments that require interaction with something specific, unrepeatable, and embodied. Music is one such thing. Field recordings, primary documents, live performance, lab observations, oral histories — there are others. For poetry specifically, music turned out to be the natural pairing, because the canon is already full of settings: Schubert lieder, Britten cycles, folk arrangements, hip-hop versions. Teachers have been making this pairing for centuries; we’re just trying to make it cheap and shareable.