Art at the speed of thought

You may have seen the story that someone in an American art competition submitted an image generated by artificial intelligence (AI). Tools like DALL-E 2, Midjourney and Stable Diffusion have made it possible for rank amateurs to produce images that would be well beyond their skills using conventional artistic tools.

Given that I definitely reside in the rank amateur class, this sounds incredibly tempting! I’ve recently dived into a trial of one of these tools – all images in this article are pretty unrefined outputs from that process. What follows are some impressions and thoughts.

First up, it is important to say how magical the experience is – the pictures in this article were generated quickly, and already given me a lot of joy as pictures are generated with distinctive moods and intricate detail.

There is also genuine artistry in the process too – experienced users know how to edit and tweak in a way to produce stunning images. Used in the right way, it is a fantastic tool that can engage more people in art and design.

Figure 1 – Generated images can be near-photo quality in some cases. Midjourney output of ‘city skyline at sunset, pink and purpose, Sydney Harbour, detailed’. Under a CC BY 4.0 license.

But as is ever the case, new technology creates new questions and potential problems to solve:

1. The first issue is ownership – if I use someone else’s AI machine to produce an image, is it owned by me, the AI machine, or the AI creator? This is probably the easiest issue to navigate. Clear terms of service agreements should resolve ownership disputes.

2. The second relates to AI tools as a mirror of society. Any machine that scours the internet and uses that as a basis of learning will take on biases and prejudices that already exist. Such issues are difficult to unpick from the overall learning process, creating a massive challenges for designers who want tools to be ethical and inoffensive. By reflecting societal bias, we potentially entrench rather than remove these issues.

A classic example of this are gender biases – producing images of nurses or teachers will generate pictures of females, while pictures of lawyers or surgeons might skew male. Other versions of the issue can range from the comical through to offensive – apparently Dall-E has learnt the innuendo attached to eggplant emojis and can throw up unappreciated explicit content as a result.

Figure 2 – Midjourney outputs for ‘nurse’ (left), ‘kindergarten teacher’ (middle) and ‘actuary’ (right). While actuary is delightfully abstract (perhaps the internet does not know what one is?), gender norms are well and truly in place for nurses and kindergarten teachers. Under a CC BY 4.0 license.

3. The third relates to challenges of imitation and derivation. The AIs are trained on existing images and inputs. This means that, in a real but indirect sense, its capability is very much derived from the skill and experience of others. This is uncomfortable territory, but not altogether new – for example, there has been discomfort over the years of services like google news benefiting from snippets provided by other news services.

This imitation of style can border on the specific – the technology can be requested to produce art ‘in the style of’ different people, which makes the imitation explicit. Conversely, it can be argued that all art has elements of derivation from prior work, so it is challenging to determine what is fair.

4. And fourth, when it comes to issues of what to do about AI art in competitions, the solution appears straightforward – introducing separate categories for pictures that really substantially on computer generation. It’s not too dissimilar to previous revolutions (photography competing with paintings, or digital art versus analogue pictures) that have required new categories.

5. The technology (as with any) has the potential for evil too, such as false news and deepfakes too – but perhaps this is a topic for a separate article!

Figure 3 – ‘Midjourney creation, with input ‘A steampunk robot hand holding a paintbrush and painting a picture’. under a CC BY 4.0 license.

While there are lots of issues to think through, it’s hard to deny the significant achievement such models represent. Creativity is enhanced by better tools, and this AI-led movement will be no exception.

As a final reflection, while I’m personally awed by the power of these tools, I cannot help but feel that there remains unrealised potential in how it can be used collaboratively with human input.

Being able to iterate, providing specific feedback (e.g. point to sections of a picture that require changes) and build images interactively may well be the next frontier, with efforts already in train.

CPD: Actuaries Institute Members can claim two CPD points for every hour of reading articles on Actuaries Digital.