A rogue artificial intelligence obsessed with Taylor Swift could supplant all recorded music with artificially generated cover versions by her, say researchers. History would show the American singer-songwriter as being responsible for everything from F眉r Elise to Paperback Writer, leaving no evidence that Ludwig van Beethoven or The Beatles ever existed.
at Durham University, UK, and at the University of the Arts London give the unusual warning in a paper that says humanity must think of methods of resistance 鈥渘ow, rather than when it is too late鈥�.
Thankfully, the risk of an AI Swiftpocalypse is low. Collins says that the idea is a thought experiment designed to prompt researchers to develop ways to protect all sorts of data 鈥� music, literature, scientific research and historical records 鈥� from being corrupted by AI.
Advertisement
The pair lays out a future scenario where we rely on a handful of centralised stores of data: Spotify and Apple for music, for example. An AI could infiltrate those stores and corrupt, delete or alter the data within. This could be in a dramatic and obvious way or insidiously and gradually. 鈥淲ithin thousands of years it鈥檚 really likely that there鈥檒l be at least some level of corruption and some level of conflict over the musical ground truth in audio recordings,鈥� Collins says.
To make their point and show how AI can already manipulate data that it has access to, the researchers used current AI models to make Taylor Swift versions of songs including Queen鈥檚 Bohemian Rhapsody,聽Frank Sinatra鈥檚 I鈥檝e Got You Under My Skin and The Beach Boys鈥檚聽Wouldn鈥檛 It Be Nice. Generating these 鈥淭aylor鈥檚 Versions鈥� for all recorded music would currently require 1.67 billion kilowatt-hours of electricity at a cost of more than $266 million, they calculate 鈥� a price tag that Swift could afford herself.
Digital and physical backups can make us complacent about the safety and permanency of our data, says Collins, but an AI with the right motivation and capability could access and corrupt anything we have recorded. 鈥淗owever much you try to preserve human culture, there may be threats in the future that you can鈥檛 anticipate,鈥� he says.
But not all experts are convinced that AI represents a serious threat in this way. at the University of Oxford says that AI has shown itself capable of causing great harm by replicating the sexist and racist biases of humans, but it isn鈥檛 going to be capable of the sorts of feats described by Collins and Grierson.
鈥淚 don鈥檛 think there is a serious problem of AI waking up, creating its own goals, having its own motivations and taking actions to fulfil those goals,鈥� she says. 鈥淚 think that鈥檚 a nonsense argument and I don鈥檛 think it鈥檚 realistic. This is similar to asking me what would I do if aliens landed on this planet tomorrow. I see it as that unlikely.鈥�
, also at the University of Oxford, says that there is a need for decisive action on AI, but it shouldn鈥檛 be some dramatic 鈥渒ill switch鈥� to halt a malevolent model in its tracks. Instead, it should be a system of careful checks and balances to ensure the safety of the AI models.
鈥淭he debate seems to assume that there鈥檚 this malevolent AI that somehow has desires of its own and becomes very powerful, and that we might want to switch it off,鈥� she says. 鈥淎nd that seems to me so implausible and so ridiculous.鈥�
The real problem, she believes, is that we will integrate AI into so many aspects of our lives that we become utterly reliant on it, creating issues that are likely to be less apocalyptic in nature and yet still very damaging, including racist and sexist biases听辞谤 simply making up plausible-sounding facts.
鈥淭he more we put it [AI] into products the harder it will be to turn it off. Not because it鈥檚 this malevolent thing that has become so powerful that it takes over, but because we鈥檝e come to depend on it and it鈥檚 very costly to turn off even when it鈥檚 not working well,鈥� says V茅liz.
Taylor Swift didn鈥檛 respond to a request for comment.
arXiv
