Opinion

Artificial Intelligence appears to be here to stay, despite its ‘hallucinations’. When it comes to music, however, Barry Fox is yet to be convinced it knows how to write a catchy tune

Advertisements will soon be promising, ‘We do not use AI’. Like ‘no MSG’ or ‘no added sugar’. People are starting to hate AI, and the moneymen are fearing it’s another dotcom-style boom-to-bust. I have been trying to renew broadband subscriptions online and by phone, using infuriating consumer helplines powered by Artificial Stupidity. One automated system could not recognise the spoken word ‘Yes’, offered to connect me to a human ‘agent’, then regretted no agent was available and told me to ‘call back later’. Another wrote me an offer and then wrote that it wasn’t available.

Are you sure?

It’s now more widely recognised that AI ‘hallucinates’. It makes up answers. Tutors are seeing this in students’ essays. In autumn 2025, financial services giant Deloitte got caught when it turned in a report to the Australian government that included AI-generated references. The idea that AI may be used to design or subjectively review hi-fi is downright scary. Don’t think it couldn’t happen. LG recently pitched its ‘innovations powered by Affectionate Intelligence’ that will provide ‘Personalised Experiences that Transform and Enrich Everyday Life, Making You Feel Understood’.

Before the late Paul Messenger started writing about hi-fi he worked for British loudspeaker company Spendor. He once told me how founder Spencer Hughes handled the unwelcome arrival of a new batch of components that did not match the previous batches. He would tweak other components by ear to retain the neutral sound of his BC1 standmount. Hughes apparently needed to spend hours picking himself up from his self-imposed gruelling listening tests.

Group listening tests need double-blind testing to reveal whose ears can reliably distinguish subtleties and whose are coloured by mind-wander or another listener’s audible sighs. I know because I’ve failed such testing. Japanese hi-fi manufacturers used to put visiting audio journalists through sessions aimed at quantifying English, German and American loudspeaker ‘sounds’. We had to press buttons to judge snatches of sound but the methodology was so pressured the results might have set Japanese loudspeaker design back by decades.

AI works by learning from what has been previously learned. It is good at recommending new tracks to listen to, on the basis of previous listens. TiVo did it years ago with TV. Spotify does it with music.

Rhyme, not reason

AI can also write new songs based on what’s come before. But there is no formula to learn for creating an earworm. The music industry uses ‘the old grey whistle test’, noting whether the janitor goes home whistling a new tune.

Irving Berlin [pictured above], one of the very few composers who was as good at writing lyrics as melodies, reckoned that writing strings of long rhyming words is easy. ‘It’s the short ones that give the trouble’, he said. He was also musically untrained, could only play by ear in only one key (the musically awkward F sharp), and needed mechanical levers in his piano to transpose his tunes to another key.

Berlin wrote 'Top Hat White Tie And Tails' for Fred Astaire. It sounds much like any other popular song, made up from eight-bar chunks, each with four regular beats. But it’s actually very different. You don’t need to be a musician to try counting Berlin’s bars and beats.

The intro, ‘I just got an invitation through the mails’, is so syncopated it’s extremely difficult for us mere mortals to count. The next two verses, ‘I’m putting on my top hat’, are like most popular songs with eight-bar chunks. The middle section, ‘I’m stepping out my dear’, sounds completely normal, but is actually wholly abnormal. Two five-bar chunks make up a ten-bar section.

This totally defeats amateur singers and players who think they can busk it. Never say never, but will AI ever write anything comparably creative?