News · ai music — By WATMM Desk — · Updated
The Beatles' "Now and Then" won a Grammy with AI, but the AI didn't make Lennon's voice
"Now and Then" won Best Rock Performance at the 2025 Grammys. The machine learning behind it separated John Lennon's real vocal from a late-1970s home demo; it did not generate a voice. That distinction is what kept the record inside the Recording Academy's human-authorship rules, and it is the line the industry should be arguing about instead of "AI or not".
Key facts
- What happened: on 2 Feb 2025 the Beatles' "Now and Then" won Best Rock Performance, beating the Black Keys, Green Day, Idles, Pearl Jam and St. Vincent; it was also nominated for Record of the Year, which went to Kendrick Lamar's "Not Like Us".
- How: Peter Jackson's team used WingNut Films' MAL audio technology, built for the Get Back documentary, to separate Lennon's vocal from his piano on a home demo the band had shelved in 1995.
- Why it matters: the Recording Academy bars work with no human authorship and requires the human contribution to sit in the category being judged; a separated human vocal clears that bar, a generated one would not.
The Beatles won a Grammy on Sunday night for a song finished with machine learning, and that sentence alone will be misread. "Now and Then" took Best Rock Performance, ahead of the Black Keys, Green Day, Idles, Pearl Jam and St. Vincent. It was also up for Record of the Year, which went to Kendrick Lamar's "Not Like Us". TechCrunch's headline put it as a win "thanks to AI", then spent its first paragraphs explaining that nobody built a Lennon voice bot.
Both halves are true, and the gap between them is the useful part. The tool behind the record separated a sound that already existed. It did not make a new one.
What did the machine learning actually do?
According to the band's announcement, the song started as a vocal-and-piano demo John Lennon recorded at home in New York in the late 1970s. Yoko Ono gave it to Paul McCartney, George Harrison and Ringo Starr in 1994 with the demos that became "Free As A Bird" and "Real Love". The band worked on it in 1995 but shelved it, because Lennon's voice and piano couldn't be pulled apart.
The fix came from film. For the Get Back documentary, Peter Jackson's company WingNut Films built an audio technology called MAL to de-mix old recordings into separate voices and instruments. Jackson's sound team, led by Emile de la Rey, ran the same process on Lennon's demo and separated the vocal from the piano. The Official Charts Company describes it as MAL being used to extract and clean up Lennon's original vocal.
Giles Martin, who produced the finished track with McCartney, explained the mechanics to NPR's World Cafe: the original was Lennon at the piano with a TV on in the background, and what the machine learning does is recognise a voice. He also said the process only works if you have the raw signal. His shortest version, in a MusicRadar video interview: "AI is not creating John's voice. John's voice existed on that cassette and we made the song around him."
Everything else on the record was played by people. The announcement lists Harrison's electric and acoustic guitar from 1995, a new drum part from Starr, and bass, guitar, piano and a slide solo from McCartney, plus strings arranged by Martin, McCartney and Ben Foster. Martin told Variety, as reported by MusicTech, that the backing vocals came from the band's own old multitrack tapes, including "Eleanor Rigby", and not from AI.
Why did "AI" stick to it anyway?
Partly because the people involved used the word loosely before they used it carefully. In June 2023, after speculation about an "AI Beatles" record, McCartney posted that "nothing has been artificially or synthetically created" and that the band had cleaned up existing recordings, as MusicTech reported. By then the story was already out.
The bigger reason is that "AI" now covers tools that do opposite things. Separation takes a recording and splits it into the parts a person played. Generation takes a prompt or a dataset and produces sound nobody played. One is restoration with better maths; the other is a new performer. Martin had already used Jackson's separation tech on the Beatles' catalogue: he told MusicRadar it was what made a stereo remix of Revolver possible.
Where do the Grammy rules draw the line?
The Recording Academy's AI protocols, introduced in June 2023, say that "a work that contains no human authorship is not eligible in any category", as reported by NBC, and that the human authorship component must be meaningful. TheWrap reported a second condition: the human contribution has to be in the category the work is being considered for. In songwriting categories, for example, the human authorship of the music or lyrics has to be meaningful and more than de minimis. The Academy's Harvey Mason Jr. gave the mirror-image example to Okayplayer: a song sung by a human but written by AI would not be eligible in a songwriting category.
Apply that to "Now and Then". Best Rock Performance judges a performance. Every performance on the record, Lennon's included, is a person playing or singing, so the human contribution sits squarely in the category. Our reading is that if MAL had synthesised Lennon's voice instead of separating it, the lead vocal would have been a machine's performance and the entry would have been on much weaker ground. The Academy has not ruled on that hypothetical, so treat it as our analysis.
What we think
The Grammy is a fair result, and the separation-versus-generation line is the right one to hold. But it should be stated as a rule about whose performance you hear, not as a rule about which software was in the chain. "Did it use AI?" is now a useless question for artists, voters and listeners. "Did a person perform this, and did they agree to it?" is the useful one.
Three consequences follow.
First, separation changes what an archive is worth. A demo that was unusable in 1995 became a number-one single (it topped the UK chart in November 2023) and a Grammy winner because the tools caught up. Every estate, label and artist with shelved tapes now holds raw material that may be finishable. That makes control of demos and session tapes a rights question, not a storage one. Artists should check their contracts for who owns unreleased recordings and who may finish them after they are gone.
Second, the Beatles had what most artists don't: the surviving band, a film studio's engineers and a demo handed over by Lennon's widow. That is the consent model working. Separation tools are spreading well beyond film studios, and a cleanly isolated vocal is exactly the material a voice model is trained on. The rules should treat separated vocals of someone who can't approve the use as needing the estate's sign-off, and say so.
Third, assistive separation belongs in ordinary workflows, where it already sits. Browser DAWs such as Veena list AI stem separation for isolating vocals and instruments, free plan included, and for most producers that is the realistic use: pulling a part out of an old bounce, a voice memo or a lost session.
The industry is better served by labels that say "separated" or "restored" than by a blanket "made with AI" that lumps Lennon's real voice in with a clone of it.
Sources
- Recording Academy, Best Rock Performance, 2025
- Recording Academy, Record of the Year, 2025
- Abbey Road Studios, The Beatles announce last-ever song Now And Then (26 Oct 2023)
- MusicRadar, 'AI is not creating John's voice': Giles Martin on Now and Then
- TechCrunch, The Beatles won a Grammy last night, thanks to AI (3 Feb 2025)
- NPR World Cafe via WAMC, How producers used AI to finish The Beatles' 'last' song (2 Nov 2023)
- MusicTech, Giles Martin on Now and Then and AI (3 Nov 2023)
- MusicTech, Paul McCartney responds to claims that new Beatles song will be "artificially created" (23 Jun 2023)
- Official Charts Company, The Beatles, Now And Then
- MusicRadar, Giles Martin on the Revolver stereo remix
- TheWrap, Grammy Awards restrict eligibility for music containing AI content (19 Jun 2023)
- NBC 6 South Florida, Only 'human creators' are eligible to win Grammys (16 Jun 2023)
- Okayplayer, Music generated with AI will be eligible for Grammys, AI songwriting will not (5 Jul 2023)
- Veena Studio, Veena
Sources
- Best Rock Performance, 2025, Recording Academy (GRAMMY.com) (primary source). Accessed 23 Oct 2026.
- Record of the Year, 2025, Recording Academy (GRAMMY.com) (primary source). Accessed 23 Oct 2026.
- The Beatles announce last-ever song Now And Then along with Red and Blue album reissues, Abbey Road Studios (primary source). Accessed 23 Oct 2026.
- The Beatles, Now And Then, Official Charts Company (primary source). Accessed 23 Oct 2026.
- Veena, Veena Studio (primary source). Accessed 23 Oct 2026.
- 'AI is not creating John's voice': Giles Martin explains why you'd be wrong to think 'AI' created Lennon's parts for The Beatles' Now and Then, MusicRadar. Accessed 23 Oct 2026.
- The Beatles won a Grammy last night, thanks to AI, TechCrunch. Accessed 23 Oct 2026.
- How producers used AI to finish The Beatles' 'last' song, 'Now And Then', NPR World Cafe (via WAMC). Accessed 23 Oct 2026.
- Giles Martin on Now and Then production and AI, MusicTech. Accessed 23 Oct 2026.
- Paul McCartney responds to claims that new Beatles song will be "artificially created", MusicTech. Accessed 23 Oct 2026.
- Giles Martin on Peter Jackson's separation tech and the Revolver stereo remix, MusicRadar. Accessed 23 Oct 2026.
- Grammy Awards Restrict Eligibility for Music Containing AI Content: 'Only Human Creators', TheWrap. Accessed 23 Oct 2026.
- Only 'human creators' — not AI — are eligible to win Grammys, Recording Academy says, NBC 6 South Florida. Accessed 23 Oct 2026.
- Music Generated With AI Will be Eligible for Grammys, AI Songwriting Will Not, Okayplayer. Accessed 23 Oct 2026.
Researched with AI assistance, written and checked by WATMM Desk.
Change log
- : Backfilled in October 2026; later sources were added for context.
Published by the company behind Veena. Disclosures