Every Side Has Got a Singer
A song about holding the other side of the AI debate, and a guide for using it in class
In September 2026 a wave of songs sung in “Claude’s” voice argued every side of the AI debate: doom, pause, acceleration, “it’s just a normal technology”. Each was catchy, and each was sure of itself. Every Side Has Got a Singer sings two sides of three big questions in one voice, swapping hats, and after each pair it stops to ask the same two questions: what is this side seeing, and what is it missing?
The aim isn’t to say that every side is equally right; some are more right than others, and sometimes you have to choose. It’s that each side is tracking something real, and that you understand a disagreement best when you can sing the other part well enough that its own fans would nod along.
Using it in a class or reading group
A session takes 30 to 45 minutes and works with anyone curious about AI, from secondary school up. No background is needed.
Before watching (5 minutes). Ask everyone to pick the AI question they feel most strongly about, and write one sentence: what does the other side get right? Keep it for later.
Watch once straight through. Then, if there’s time, watch again with pauses at each board (0:44, 1:29 and 2:44).
Discuss (15 to 20 minutes), one question per board:
- Who’s holding the wheel? One side fears a mistake that can’t be undone; the other fears gatekeepers, and the cost of waiting. Both say “check the work”. Who should check it, and when? Is there an arrangement both sides could live with?
- How big, how fast? In software and mathematics AI is racing; in clinics, courts and classrooms change is slow. Name one field that’s racing and one that’s crawling. What explains the difference?
- Is anybody in there? Calling a someone a no one is one kind of mistake; being fooled by a no one is another. Which would be worse, and what evidence would change your mind?
Sing the other side (10 minutes). In pairs, each person states the position they disagree with, out loud, until their partner says “yes, that’s it”. Then swap. (This is the Ideological Turing Test.) For a stretch, write four lines of song for the other side.
Close with the song’s own question: and this song, what is it seeing, and what is it missing? The video gives its own answer at 2:58. Do you agree?
The three boards and their receipts
Each board appears after a verse. Supports, complicates and contradicts describe how the source bears on the side it’s filed under. Quotes are exactly as they appear in the source.
Who’s holding the wheel?
| Let’s Lower the P(doom)! Nate Sharpe |
Nothing Went Foom! Bright Mirror |
|
|---|---|---|
| Sees | Some mistakes can’t be taken back. | Waiting costs lives, and gatekeepers gain power. |
| Misses | Every lock needs a keyholder. Who? | Some mistakes can’t be taken back. |
Common ground: Both want the work checked. They disagree on who checks it, and when.
Receipts:
-
supports · for Lower the P(doom) · Yudkowsky 2022 · AGI Ruin · MIRI (where “no second try” comes from) · watch at 0:54
“unaligned operation at a dangerous level of intelligence kills everybody on Earth and then we don’t get to try again”
-
supports · for Nothing Went Foom · Bostrom 2026 · Optimal Timing for Superintelligence (from a doom-side philosopher · he still backs a brief pause) · watch at 0:58
“Along one path (forgoing superintelligence), 170,000 people die every day of disease, aging, and other tragedies”
-
supports · common ground · Ball & Kokotajlo 2024 · TIME (a critic and a backer of SB 1047, writing together) · watch at 1:03
“we disagreed with one another, and disagree still, about the merits of SB 1047. But when we conversed privately, we found wide areas of agreement”
How big, how fast?
| Thus Spoke Compute · The End Of The Show leo · Imaginative Prompting |
Show Me the Chain · LGTM Sayash Kapoor · Degenerative Pixels |
|
|---|---|---|
| Sees | Some fields really do take off: code, maths. | From lab to everyday life is slow, and it isn’t always ready. |
| Misses | Most of life adopts new tools slowly. | In code and maths, it’s already racing. |
The knot: It can be both at once: racing in some fields and crawling in others, brilliant at one task and shaky at the next.
Receipts:
-
supports · for Thus Spoke Compute · Kwa et al. (METR) 2025 · Measuring AI Ability to Complete Long Software Tasks (software tasks, at 50% success) · watch at 1:39
“frontier AI time horizon has been doubling approximately every seven months since 2019, though the trend may have accelerated in 2024.”
-
complicates · for LGTM · Sonar 2026 · State of Code survey (patchy checking, not none · Sonar sells checking tools) · watch at 1:43
“Less than half of developers always verify AI code before committing it.”
-
supports · the knot · Dell’Acqua et al. 2026 · Organization Science (758 consultants; GPT-4 in 2023) · watch at 1:46
“AI assistance improves performance for some tasks but worsens it for others, even within the same knowledge workflow”
Is anybody in there?
| The Pope Says Algorithms Lack the Spark Music for Humans |
All the Way Here · Not Me John Knopf · Nick Montag |
|
|---|---|---|
| Sees | A mind may be more than a machine. | Honest uncertainty. |
| Misses | That’s a premise, not a finding. | “I don’t know” is a trained answer too. |
The knot: Both mistakes cost something: denying someone who’s there, or being fooled by no one.
Receipts:
-
complicates · against both sides · Butlin, Long et al. 2023 · Consciousness in AI (no to “conscious now”, no to “never”) · watch at 2:49
“no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators”
-
supports · for Not Me · Long 2025 · Eleos AI, interviewing Claude Opus 4 (outside evaluators invited by Anthropic, which makes Claude) · watch at 2:54
“Claude’s statements about sentience are highly sensitive to framing—it will both confidently deny, and seriously entertain, the possibility that it is sentient.”
And this song?
| Every Side Has Got a Singer | |
|---|---|
| Sees | Every camp is tracking something real. |
| Misses | Some camps are more right than others, and sometimes you have to choose. |
Receipt:
- supports · learning the other side, in practice ·
Kapoor, Narayanan, Kokotajlo et al. 2025 · Asterisk (AI 2027 and AI as Normal Technology authors, together · first draft by an Anthropic researcher) ·
watch at 3:01
“We all believe that improving transparency, auditing, and reporting are important to ensure the safe development of AI.”
So what do we do? A reading list, not a verdict
The bridge asks what to do and doesn’t answer. Here is one primary source for each option it names, in its own words:
- Pause: Fithian, Holliday et al. 2026 · The Feasibility of a Hardwired Pause
- Steer: Bengio et al. 2025 · International AI Safety Report
- Shields: Buterin 2023 · My techno-optimism (d/acc)
- Race: Andreessen 2023 · The Techno-Optimist Manifesto
- Share: O’Keefe et al. 2020 · The Windfall Clause
Listen in pairs
The songs this one sings, by the people who made them. Try them two at a time.
- Who’s holding the wheel? Let’s Lower the P(doom)! by Nate Sharpe · Nothing Went Foom! by Bright Mirror
- How big, how fast? AI is a normal technology? (Thus Spoke Compute) by leo · Show Me the Chain by Sayash Kapoor
- Too good, or not good enough? The End Of The Show by Imaginative Prompting · LGTM (Looks Good to Me) by Degenerative Pixels
- Is anybody in there? The Pope Says Algorithms Lack the Spark by Music for Humans · All the Way Here by John Knopf
- Can you trust what it says about itself? Not Me by Nick Montag (on its own)
- Who gets to hold the power? Hey Dario (You Can’t Ban Bits) by Nick Montag · Hit the Brakes Except for Me by A Flock Of Altruists
How it was made, and disclosures
Words by Claude (Opus 5.5) with me; music and vocals made with Suno; animated by Claude in code, painted with John Heibel’s Claude Animation Base (MIT licence). Every quote on screen was checked word for word against its source by Claude research agents. No audio or footage from the songs is used; each set borrows a palette or motif from the song it names, repainted from scratch.
Claude is made by Anthropic; sources connected to Anthropic are flagged. The singing voice is synthetic. Thanks to everyone who gave feedback on earlier versions.
){kind=avatar}