Skip to main content

From MrBeast to MrBot?

Generative AI seems to have entered the mainstream hype cycle with the recent release of ChatGPT becoming the fastest growing website to reach more than 100 million users, generating a lot of speculation  about AI disrupting and redefining how we navigate and access knowledge and information.

Having worked with but not on machine learning the last few years, I am myself amused, amazed and somewhat terrified by the impressive progress in what publicly available text and image generators are able to do now.

At its current state, ChatGPT is essentially an entertainment platform in the form of an infuriatingly polite, well-read, hallucinating bullshitter. One should not ask it any question for which we don't already know the answer, but like a mechanical automation in a victorian era parlour, we can marvel at how well it performs some amazing tricks some times or laugh at how it spectacularly fails some other times.

Unless there is an unexpected ceiling, another decades long "AI winter", we can expect generative AI to be some of the most significant disruption of our lives since the introduction of computers, the Internet or mobile phones.

Most of the large AI research labs are working frantically at trying to teach the latest generation of large language models to be less fast and loose with the facts and become more like a responsible librarian or researcher than a fast talking con artist. If or when they succeed, this will have tremendous impact on the life of any knowledge worker. Anybody who reads and writes for a living, who analyses sources and compiles reports based on them. In short, almost anybody working in an office today.

But there are also fields where truth and facts are far less important than sheer popularity: for example in entertainment and politics. In the realms of commercial pop-culture, success is very easy to measure, in terms of clicks for web publishing, best-seller lists for books, billboard 100 charts for music or Nielsen ratings on TV or views, likes, fans or subscriber counts on streaming video or social media distribution platforms. And one thing that machine learning is good at, is to optimise for a well defined success metric.

Over the last few decades, the digital revolution and the internet has reduced the cost of content creation and distribution to almost zero. Making a record or a TV show no longer requires a studio with equipment worth millions of dollars but can be done today with a cellphone or a laptop. Ironically this democratisation of access to the means of content creation has fuelled a winner-takes-it all race, where fewer and fewer of the most famous and successful creators manage to stand out of the noise and get the lion's share of attention, fame and fortune. And for the last decade, machine learning driven recommendation systems of content distribution platforms have indirectly played a role as king-maker and gatekeeper.

Would it be so hard to imagine how only slight evolutions of the current generative AI models could flood the current content distribution channels and optimise for what we want to see, hear and read in a similar but more systematic way the successful commercial content creators have been doing for decades?

AI creators might not (yet) sweep the Nobel prize for literature, the Pulitzer prize for journalism or the top outstanding achievement awards for music film or TV, but that's not really where they money is. Would it really be so hard to imagine an AI generated paperback novels ghost-written celebrity memoirs, clickbait articles, pop music tracks or social media influencer post?