Published November 13, 2023 · By CampaignsLive · Case Study
On the second of November 2023, Apple Corps released “Now and Then,” the first new Beatles song since 1995. The track was built around a vocal performance John Lennon had recorded onto a cassette tape in 1977 — a year before he was killed — that had been sitting in the family’s archive ever since. The cassette had been considered unusable for decades because Lennon’s vocal could not be separated from the piano and ambient noise on the same channel. In 2022, working on the Get Back documentary, Peter Jackson’s WingNut Films team adapted the audio-source-separation tools they had developed for that project and applied them to the cassette. The vocal came out clean.
The track that resulted was not, in any meaningful sense, an AI-generated Beatles song. It was a recording of John Lennon, made by John Lennon, that had been technically inaccessible until machine learning made it accessible. Paul McCartney and Ringo Starr added new instrumentation. George Harrison’s guitar parts, recorded in 1995 for an earlier attempt at the song, were already in the archive. The track was assembled, not generated.
The episode is worth studying because it is one of the cleanest demonstrations of what AI in creative production should look like, and one of the cleanest counterexamples to the slop conversation.
What the tool actually did
The Get Back team’s audio-source-separation tools — public-facing versions of similar techniques are now available in commercial tools like Audionamix, RX, and several Spleeter-derived offerings — work by training a model on a large set of audio recordings and learning to isolate the components of a mix: vocal, guitar, drums, ambient, and so on. Given a new mix, the model can produce estimates of each component on its own, even when the original recording did not separate them onto different channels.
For “Now and Then,” the relevant input was the 1977 cassette: Lennon at a piano, singing into a single microphone, with the piano bleeding into the vocal channel and the room’s reverberation in both. The desired output was Lennon’s voice on its own, in a usable state.
The model produced that output. It did not generate Lennon’s voice; it isolated it. The voice that came out was the voice that was already on the cassette, with the piano and the room stripped away.
This is a small but consequential distinction. The tool did not create new material. It made existing material accessible.
Why this version of AI works
The “Now and Then” project is a clean case because the model is doing something specific and well-defined: separation, not generation. The output is verifiable against the input. There is no creative judgment being delegated to the model — the model is doing a technical job that would otherwise have required either a multi-track recording that did not exist or a remix that was not possible.
This is the version of AI in creative production that almost no one objects to. The model is augmenting a human creative choice; it is not making the choice. The artist’s intent is the input. The model is in service to that intent.
Compare this to the cases that produce the strongest reactions:
- Marvel’s Secret Invasion title sequence. AI generated commercial visual material that would otherwise have been commissioned. The creative judgment moved from a title-design studio to a model. The displacement is the issue.
- AI-generated stock photography in retail. AI replaced a category of human commercial work that supported a real workforce. The labor is the issue.
- AI-generated influencer personas. AI replaced the human-presence value that the format depends on. The authenticity is the issue.
None of those frictions are present in “Now and Then.” The Beatles are the Beatles. Lennon’s voice is Lennon’s voice. The tool extracted what was already there.
What this teaches the brand creative conversation
The “Now and Then” model — AI as a separation, restoration, or augmentation tool, in service to existing human creative work — is a posture that brand creative has not adopted enough of.
Most of the AI-in-brand conversation in 2023 was about generation. Brands testing image generation. Brands testing copy generation. Brands testing fully AI-produced commercials. The generation framing is what attracted attention and what attracted criticism.
The augmentation framing is what most brands actually need. Brand creative production is full of tasks that look like separation, restoration, or extension:
- A campaign concept exists in one format; the team needs it in fifteen. The work is not generating new concepts; it is extending an approved one across placements.
- A photoshoot produced material the team wants to extend with environmental fill, sky replacement, or scene continuation. The creative judgment is in the original shoot; the AI work is in service to it.
- An older brand asset is being revived for a campaign. The asset is low-resolution, low-quality, but the original creative intent is intact. Restoration tools can make it usable without changing what it is.
- A brand archive is large enough that finding the right reference is itself a problem. Embedding and search tools — adjacent to but not identical to generative ones — make the archive accessible.
None of these uses produce slop. None of them attract the labor-displacement objections that the generation cases do. All of them, used well, augment what brand creative teams are already trying to do.
The broader takeaway
The Beatles project is not directly transferable to brand creative production, but it has been a useful reference point in the years since. The argument for AI in creative work is strongest when the technology is doing something the alternative could not do — recovering material that was inaccessible, scaling work that was uneconomic, holding consistency across formats that humans cannot hold reliably. The argument for AI in creative work is weakest when the technology is replacing a category of work that humans were doing perfectly well.
The brands and platforms that have done well with AI imagery since have tended, often without naming the principle, to operate on the augmentation framing rather than the generation one. The work is in service to an existing creative direction. The model is a tool, not an author. The intent is human.
“Now and Then” is the cleanest expression of that posture currently available in mass-market creative work. It is a useful template for the brands still working out where AI belongs in their production stack.