Case Study 1: "Now and Then" — The Machine That Un-Mixed a Cassette

This one is documented history — the recordings, dates, releases, and technology below are a matter of public record, told by the participants in interviews, official credits, and a making-of film released alongside the single. Where the story leans on session lore (and a few corners of it do), it's framed the honest way: as the story the people involved have told. It matters to this chapter because it is the cleanest answer anyone has yet produced to the tool-versus-replacement question — a case where the most advanced audio AI of its moment was pointed at a song, and every musical decision still belonged to humans.

A Voice and a Piano, Glued Together

Sometime in the late 1970s, John Lennon sat at a piano in his apartment in the Dakota in New York and sang a song called "Now and Then" into a cassette recorder.

Stop and hear that signal chain with your Chapter 3 ears, because everything that follows hangs on it. One built-in microphone. One mono cassette — consumer tape, consumer speed, the noise floor and wow of Chapter 2's worst-case scenarios. The piano and the voice arrive at that single mic already summed by the air of the room — not two tracks, not even two channels, but one waveform with two sources baked into it. And underneath everything, a low electrical hum from the apartment's wiring, the kind of grounding gremlin you learned to hunt in Chapter 3 — except nobody hunted it, because this wasn't a session. It was a man at home, getting an idea down before it left.

A demo, in other words. The exact artifact you've made a hundred of on your phone.

Lennon was murdered in 1980. In 1994, Yoko Ono gave the surviving Beatles cassettes of his home demos — among them "Free as a Bird," "Real Love," and "Now and Then" — for the Anthology project. Which set up one of the most instructive failures in recording history.

1995: The Attempt That Couldn't

For Anthology, Paul McCartney, George Harrison, and Ringo Starr — with Jeff Lynne producing — finished "Free as a Bird" and "Real Love" and released them in 1995 and 1996: the first "new" Beatles singles in a quarter century, built around Lennon's demo vocals with everything the era's tools could manage. Those two demos were workable, barely. The craft was heroic and the seams were audible; engineers fought tape hiss, timing drift, and the fundamental problem that the vocal could never be fully freed from the piano underneath it.

"Now and Then" went worse. The trio attempted it — by the participants' accounts, the work lasted a day or two — and abandoned it. The demo's vocal and piano were locked together harder than the others, the hum sat under everything, and there was no technology on Earth that could pull a voice cleanly out of a mono cassette. The story goes that George was the bluntest about the source's quality — his family has framed his objection as technical, about what the cassette could give them, not about the song — and Paul has retold versions of that disagreement for years. However you weight the lore, the engineering fact isn't lore: in 1995, the master-versus-elements problem had no solution. The session produced one durable asset — George recorded rhythm guitar parts before the attempt stopped — and then the song went back in the drawer.

Hold the shape of that failure: the music was there, the will was (mostly) there, the budget was effectively infinite. What was missing was a capability. Twenty-eight years of skill accumulation wouldn't have fixed it, because no amount of skill un-sums a waveform.

The Documentary Detour

The capability arrived from a sideways direction, which is how capabilities usually arrive.

When Peter Jackson's team began work on the 2021 documentary Get Back, they faced an audio archivist's nightmare: hours of 1969 mono recordings in which conversation, guitars, drums, and room noise share single tracks — including moments where the Beatles deliberately masked private conversations by playing over them. Jackson's audio team at his New Zealand production company developed a machine-learning separation system — trained, in plain Chapter 38 terms, to recognize what a voice is, what a guitar is, what a drum kit is, and to assign a mixture's energy to its owners. They called the approach MAL. It was the sidebar of this chapter's index, pointed at history: learned masks over time-frequency tiles, deciding this energy is John talking, that energy is the band.

It worked well enough to change the documentary. Then it changed the catalog: the 2022 Revolver special edition used the same demixing to take 1966 four-track tapes — where whole rhythm sections had been bounced together onto single tracks, exactly the constraint economy you read about in Chapter 5 — and pull the elements apart for a modern stereo remix. That release was the proof of concept the world mostly missed: demixing wasn't a party trick. It was restoration-grade.

In 2023, they pointed it at the Dakota cassette. The system separated Lennon's voice from the piano — clean, present, free of the hum — at a quality the 1995 team couldn't have imagined. Everyone who has described hearing the isolated vocal for the first time describes the same jolt: not "impressive technology," but John, in the room, clear as anything.

What the Machine Did — and Everything It Didn't

Now read the production credits of the finished single with this chapter's two axes in your hand, because the division of labor is the entire lesson.

What the machine did: one thing. It un-glued a mix — recovered a clean lead vocal from a mono cassette where voice, piano, and hum had been a single waveform for forty-five years. Tedium is too small a word; this was impossibility, converted into a capability. That is the purest "tool" reading the chapter's grid allows: it replaced nothing a producer does, because no producer could ever do it.

What humans did: literally everything else. Paul and Ringo built the track — bass, drums, piano, and a slide guitar solo Paul played deliberately in George's style, as tribute. George's actual 1995 rhythm guitar parts were woven in, so all four Beatles play on the record. A string arrangement — written by Giles Martin, Paul, and an orchestrator — was recorded with session players at Capitol Studios who reportedly weren't told what they were playing on (secrecy as session management; Chapter 19 would approve). Backing vocals were lifted from "Here, There and Everywhere," "Eleanor Rigby," and "Because" and woven into the stack — a sampling-as-séance move the Anthology singles had pioneered. The production was credited to Paul McCartney and Giles Martin; a veteran mix engineer mixed it; mastering, sequencing, release strategy, the lot — humans, all the way down the Chapter 31–35 pipeline.

Arrangement judgment: human. Performance: human. Taste — which takes, which tribute gestures, how much string, how loud the ghost of a 1979 living room should sit against a 2023 rhythm section: human, human, human. The AI's contribution was profound and bounded: it handed the producers a stem. Every decision about what that stem should mean was made the way this book taught you decisions get made.

The Disclosure Wobble

One more documented beat, because it previews the norms section of this chapter perfectly. In mid-2023, before release, McCartney mentioned in a radio interview that AI had been used to "extricate" John's voice for a final Beatles record. Headlines did what headlines do: within a day, the story circulating was that the Beatles had used AI to fabricate Lennon's voice. McCartney had to clarify publicly — nothing was generated, nothing synthetic, the AI separated a real performance from a real tape, full stop.

The lesson costs you nothing to learn from his bill: say precisely what the AI did, or the public will assume the worst version. "AI was used" is not a disclosure; it's a Rorschach test. "Machine-learning source separation recovered the original 1970s vocal; all other performances are human" is a disclosure. When the disclosure was made that precisely — and the release was accompanied by a short making-of film that showed the cassette, the process, and the people — the story flipped from suspicion to celebration.

The Verdict

"Now and Then" was released on November 2, 2023, billed as the last Beatles song, as a double A-side with "Love Me Do" — their first single, sixty-one years earlier, closing the loop with the tidy poetry history rarely provides. It went to number one in the UK — the band's first new chart-topper there in over half a century — and, as of this writing, it won the Grammy for Best Rock Performance, making the Beatles Grammy winners across an eight-decade span. A song technology had to abandon in 1995 became, with one capability added and every human judgment kept, a record people genuinely loved rather than merely tolerated as an artifact.

What This Case Teaches

1. Restoration is the proof that "tool" is a real category, not a cope. The loudest framing of music AI is generation — machines making the music. The most successful music-AI project in history to date was the opposite: a machine recovering human music that already existed, so that humans could finish it. When someone tells you the tool/replacement distinction is wishful thinking, this is your counterexample: the most scrutinized AI deployment in pop history, and the line held perfectly.

2. Capability limits are temporal — so capture well and keep everything. The 1995 team didn't lack taste or money; they lacked a capability that arrived twenty-eight years later. The producer-sized lesson lands directly on your Chapter 2 and Chapter 19 habits: your "failed" sessions, rough demos, and unfinishable ideas are not dead — they're waiting. Record at honest levels, label your files like an adult, archive multitracks even when the bounce seems fine, and keep the 3-2-1 backups alive, because the most valuable file in this story spent forty-five years as a consumer cassette in a drawer. Lennon's demo survived on two strokes of luck — somebody kept it, and somebody could find it — and luck is a terrible archival strategy. Yours shouldn't need any: the phone memo you almost deleted last month is, on the evidence of this case, potentially a record. You can't know which one. So you keep them all.

3. The stem is not the record. Demixing handed the producers a clean vocal, and there was still a full pipeline of judgment between that stem and a release: arrangement, performance, tribute, restraint, mix, master, sequencing, story. If AI ever hands you a miracle stem — and stem separation means it eventually will — your job begins there, exactly as theirs did.

4. Consent and disclosure are why this one felt good. Run the contrast that defines the voice-ethics section: "Heart on My Sleeve" took two living artists' voices without permission and collapsed in days. "Now and Then" used a dead man's actual performance, with his family's participation, his bandmates' hands on every fader, and (after one wobble) a precise public account of what the machine did and didn't do — and it became a beloved number one. Same year. Same underlying technological moment. Opposite architecture of consent, and opposite outcome. The technology didn't decide which story each became. The humans did — which, you may notice, is this chapter's entire argument wearing a Beatles record as evidence.