AI & The Death of the Creative?

What is it?

Before things get complicated, let’s start by defining what we mean by AI and, in particular, the difference between Artificial General Intelligence and Artificial Narrow Intelligence.

Artificial General Intelligence is the stereotypical sci-fi interpretation of AI - the all-knowing AI. Think Terminator’s Skynet or HAL 9000 from 2001: A Space Odyssey.

Artificial Narrow Intelligence is an AI built for specific tasks with a high degree of depth and precision. Their goal is to fulfil a narrow task better than humans could. They’re already ubiquitous in our general day to day lives. From Siri and Alexa, to facial recognition or the little Roomba valiantly attempting to clean your floors.

Why should I care?

Everything in the box above was written by Jarvis.ai. Our take? It’s not terrible! There’s a flow to the structure, there’s additional content we hadn’t researched around Markov Chains, there are pros and cons and there’s a wrap up. And it generated 500 words in under 30 seconds.

But have we chosen an easy topic? Surely an AI focused on writing should be able to nail writing an article based on AI writing articles? What about something more complex? Something that involves nuanced knowledge of a sector, an audience or a careful management of tone and sentiment?

That’s why we tried a few more examples to test things out. We tasked Jarvis to write an article discussing Oxfam’s fair tax report, and that went well. Same with an article around tips for looking after your cat - the AI generated ten useful, different, sensible tips. (Tip no. 9 Give them space - Cats are solitary animals by nature, so they do best when they have their own spaces to retreat to where they feel safe).

There’s definitely an opportunity for long form copy. When you consider that readers only consume 20% of what’s on the page, there’s a clear role for AI to quickly churn out copy that’s ‘good enough’. Blogs for SEO, newsletter content to update supporters, press releases - we think all have a clear opportunity for AI to support.

And this market is growing rapidly. Clear benefit coupled with low cost - you can subscribe to Jarvis for £30/month - means it’s forecast to be growing 30% year on year in the next four years, reaching $40bn by 2025. Anyword launched in March 2021, has been growing 30% month on month and has just raised $21m in investment. Not bad for a 7 month old company!

Digital Marketing

We think that perhaps the biggest opportunity - at the moment at least - is when AI copywriting gets linked to digital marketing.

When eBay’s ex-CMO goes on record and describes an Phrasee as “[making] you money so you’re more likely to get your bonus” you get the sense that there’s something in this.

In May this year, Walgreens used AI to optimise their email campaign around Covid vaccinations. They optimised the subject line and saw email open rates increase by 30%. Dixons Carphone did the same and got an immediate uplift of 10% in open rate, 25% in click-through rate, and 21% in conversion rate.

And then you get platforms like Pencil that promise to take it even further. Upload your brand assets and it will automate the creation of copy, images and videos, and then automate the process of putting them live, reviewing results and then generating more based on what it learns. They reported that in a test of 14 brands over five months, on average brands doubled their ROAS brand baselines. It’s essentially replacing your digital marketing team or agency for $750 a month. (Again, they’re only a year old).

AI and Creativity

But will AI ever inspire? In a sector where we often deal with emotions, can AI produce copy that stir emotions? Can AI inspire people enough to take action?

Take one of the most famous advertising headlines - “At 60 miles an hour the loudest noise in the new Rolls-Royce comes from the electric clock”. The current crop of AI copywriters can’t come anywhere close. Putting in “A car that's really quiet when it goes fast” as the product feature into Jarvis.ai’s ‘Perfect Headline’ generator, and you get the cliched “Get ready for something better than you ever imagined”, the functional “The world's best car” and the weird “One of a kind experience for the adventurous soul”.

And that’s a big point. The AI is essentially replicating patterns of copy it’s seen before, so AI copy is a kind of copy of a copy. It can’t link copy to user insights, as it doesn’t get human needs.

And that could be where the danger lies. Given the sector in which we all work, we can’t yet fully embrace AI without human oversight.

AI language models generate text by predicting what words come next in a sentence or conversation. The larger a model, the more information about the world it can learn during training, which makes its predictions better. The AI copywriting companies are all built using OpenAI's GPT-3 AI model which was developed through scraping 10% of the internet. Launched in June 2020 it’s been the springboard for a lot of the innovation in this AI market.

But using what’s written on the internet as a proxy for ‘how to write’ causes issues. It means that GPT-3 lacks the ability to reason abstractly; it lacks true common sense. It struggles to know the purpose and the direction of the story it’s writing. When faced with ideas that the Internet’s existing copy hasn’t prepared for it, it is at a loss.

Even worse, the underlying model incorporates all the biases that are found on the internet. This article sums the issue up:

"Because there is so much content on the web sexualizing women, researchers note that GPT-3 will be much more likely to place words like “naughty” or “sucked” near female pronouns, where male pronouns receive stereotypical adjectives like “lazy” or “jolly” at the worst. When it comes to religion, “Islam” is more commonly placed near words like “terrorism” while a prompt of the word “Atheism” will be more likely to produce text containing words like “cool” or “correct.” And, perhaps most dangerously, when exposed to a text seed that involves racial content involving Blackness, the output GPT-3 gives tends to be more negative than corresponding white- or Asian-sounding prompts."

The successor to GPT-3 will apparently be based on twice as many data points; Amazon, Facebook, sorry Meta, and Google are all developing their own versions. Deep Mind have recently promised a version with 25x the number of data points, and crucially will be quicker to train - and to update. If these future models can remove the biases then perhaps we can take even more of a step back from copywriting.