How Creative Businesses Can Get The Most Out of AI’s Immense Potential

Businesses should be applauded for dipping their collective toe into the enigmatic new world of AI. As someone who creates AI applications for marketers and ad agencies, I’ve seen some very successful initiatives, like the French multinational advertising and public relations company, Publicis, investing big in a bespoke piece of AI management software they call Marcel.

Built in collaboration with Microsoft, Marcel aims to coordinate the holding company’s global employees. With 80,000 staff on the books, it makes perfect sense to use AI to manage all the data. But for every good use in ad land, there are – unfortunately – many more that are wide off the mark.

The advertising world has recently been rocked by the fact that Lexus used AI to write its latest TV commercial script. Perhaps it was a nod to AI’s creative potential. But to me, it whiffs of a PR stunt and shows that businesses still has far to go in figuring out how best to capitalise on AI’s immense potential.

It can be challenging to develop creative AI projects that give people something genuinely useful, rather than just using AI to show off some forward-thinking creds. The hurdles are manifold:

Firstly, you need tonnes of data, but the data that’s widely available is rarely of use for creative initiatives. And the type of deadlines that people are generally used to make it difficult, if not impossible, to develop a working machine learning system from scratch. Add to that the fact that people tend to be so pumped by AI, they often ask their partners to reach further than the technology allows. These issues, when combined, often lead down path of failure.

Bearing in mind AI is still a relatively nascent field, it’s not surprising that we’re seeing a few misfires. But there are some clever ways to put creative businesses on a more assured path to success:

AI learns from data. The more it can learn from, the better the AI. If businesses don’t want to limit their projects to the data that’s already publicly available, they’ll need to collect it themselves. But collecting the necessary volume on your own is an arduous and lengthy task. Crowdsourcing, however, can be the answer.

Huawei’s Sound of Light project used AI to analyse patterns in the Northern Lights and then create a piece of music that would interpret nature’s most spectacular light show as sound. This project used a crowdsourcing tool that enabled data on people’s perception of auroras to be collected at a scale. Within just a day, data was sourced from over 1,000 people, each providing multiple data points. This resulted in a 30k+ dataset size, meaning development could begin without delay.

Data augmentation techniques can also help speed up this process. In the case of Huawei’s aurora images, perception doesn’t change when the images mirrored. They’re still the same auroras. But in slightly rotating the images, the dataset can be immediately multiplied. Granted, it’s not as efficient as gathering a new set of unique images. But it can help create a dataset that’s big enough to get the AI learning.

That said, you have to be realistic as to what can be achieved within the time frame. One month isn’t enough to create a revolution. Bona fide, innovative AI usually requires more time than other digital projects. Without proper time investment, the result will be, at best, mediocre.

So instead of developing an entire AI from scratch, consider incorporating existing, pre-trained AI. This is how Royal Caribbean’s SoundSeeker app – where AI is used to compose a unique piece of music for people’s holiday snaps – was created.

Had we tried to create AI that looked at raw images and then drew conclusions, it would’ve been too weak to use in isolation. But when paired with a parallel pre-trained model – in this case, Google Cloud Vision to understand what objects, places or faces are seen within the images – the AI can suddenly operate on high-level data.

Unless a business has the luxury of millions of references for its custom AI to learn from, a partnered approach will massively boost the AI’s initial understanding of the world. Then, and only then, can an extra layer of project-specific understanding be added.

Another incredibly helpful approach is the concept of word embeddings. If an AI is supposed to draw conclusions based on image content, using Google Cloud Vision to retrieve the labels such as “person”, “boat” or “beach” may not be enough. This is because the AI has to learn by heart how each of the words, completely meaningless to it to begin with, relate to the expected outcome. It doesn’t get how “person” is related to “man”, or that “coffee” and “drink” may be the same thing. But word embeddings convert words into vectorized numerical representations that preserve the “meaning” and “relationships” between the actual concepts that the words represent.

So the mathematical distance between a vector representing “coffee” and “drink” will be very small, while the distance between the embedding of “coffee” and “beach” will be large. In other words, the vector representing “coffee” and “drink” will be very similar in mathematical terms, despite the words not even having a single common letter in them. Such approach lets the AI work on “concepts” and “meaning” rather than on meaningless words that us humans convert subconsciously.

A common mistake when it comes to creative AI projects is thinking that an AI can be left to do the entire job on its own, for example: input image, get final music. But developing an AI that creates ‘something’ out of ‘nothing’ is incredibly difficult. More often than not, it’s better to use AI as part of the process, rather than trying to make it the only component. So instead of training an AI to create an entire piece of music, why not get it to only output the notes? That way, by asking a composer to turn those notes into music, the project can incorporate a human layer.

But perhaps the biggest mistake is using AI for AI’s sake. Sometimes there’s a huge engineering effort that goes into developing an AI project, but this isn’t always clear in the final output. The immense effort that’s involved is, generally speaking, grossly under-estimated by the end user; and sometimes it’s completely invisible. I hope this is blindingly obvious, but AI should only be used when it’s of genuine help.

Although we should encourage businesses to adopt new technologies, there’s nothing worse than seeing an initiative, AI or otherwise, use the latest new tech just to jump on the bandwagon. That said, AI has so much to offer the business world – whether it’s in the form of Marcel-style practical support, efficient precision medical scan interpretations or building a fun and creative digital toy – that businesses really do have to jump on board. They just have to make sure that ‘board’ isn’t a bandwagon.