Is Generative AI like ChatGPT and Dall-E Creative?

Intelligence without Creativity is not possible. But, ask 5 people what Creativity is, and you will get 6 different answers. One of the reasons we are so inspired and hyped about AI lately, the image generation of Dall-E, MidJourney and ImaGen - or the textual generation from ChatGPT, is that it's seen as creative. But is it?

To answer that we have to understand what Creativity is. There are many ways to describe it, but I tend to lean toward frameworks that decompose creativity into distinct subtypes. A suitable model to use when discussing Generative AI is to decompose creativity into 5 types of thinking: Convergent, Divergent, Associative, Synthetic and Contradictory.

• Convergent Thinking is finding a solution to a problem after considering "all" possible options. This can also be said to be the analytical approach common with engineers.
• Divergent Thinking is coming up with multiple solutions to a problem. This is the brainstorming way of approaching problem solving, although I would argue that applying design thinking methods for divergent thinking is vastly superior to post-its in a sweaty meeting room.
• Associative Thinking is coming up with connections between ideas that would not normally be connected. This is often seen as spontaneous inspiration, and often described as happening when observing or doing something else. The story of Newton figuring out Newtonian Gravity from a falling apple is a good example.
• Synthetic Thinking is combining existing things in new ways, creating unpredictable emergent properties. Good examples include musical creativity, where notes and sounds are combined in ways that evoke something much grander than the parts. Or creation of a new dish of food, where ingredients and preparation is combined in new ways.
• Contradictory Thinking is to change an assumption, constraint or viewpoint - and thus changing what the obvious solution would be. A great example is how taxi's assumed that dedicated professional drivers and dedicated cars were necessary for transporting people - unable to see that if you removed those assumptions you would get rideshare. Which is obvious in hindsight, but contradicted the existing worldview at the time.

With that established, how does this relate to Generative AI?

Firstly, as humans we employ all 5 strategies at different times - and quite often we combine them. Some humans are really good at one, or maybe more than one - often seen as creative and intelligent by their peers. Imagine that these five ways of thinking is an equalizer with the sliders placed differently for every human.

Machine Creativity can do some of the above strategies - but more importantly, it can make us better at some of these strategies. So let's go through the list again and see how Machine Creativity is either employing these strategies or helping us do them:

• Convergent Machine Thinking. A lot of problems can be solved better by machines because they can excel at computation. When you have a 100 different solutions, finding the best one can be about creating a way to test all the approaches. Chess computers do this, there is many moves, what is the best one? Genetic algorithms, reinforcement learning, and generative-adversarial networks (GANs) all can be argued to be partly convergent in their approach. AI can help us test more solutions faster, and can greatly accelerate our ability to do convergent thinking. The human contribution is how we formulate the definition of success.
• Divergent Machine Thinking. I don't believe this is very common. Having a machine brainstorm is possible, but from a divergent standpoint it's hard to see how machines will deliver anything of both novelty and utility that is not already present in some form in the training data. Although, employing AI to generate hundreds, if not millions of variations from a prompt can have tremendous utility. The key here though is that the human contribution is the prompt. Without a human supplying a good prompt, again, a human formulates what they are looking for... the machine supercharge our ability to iterate.
• Associative Machine Thinking. Now we are getting to things that are really hard for machines. Theoretically it shouldn't be, but associative thinking also needs a feedback loop to determine when an association has the potential to be novel and have utility. So when machines do associative thinking, the associations are often generated in a divergent way... the shotgun approach. Famous creative environments like Bell Labs and Pixar was consciously created to maximize serendipity, where people in different fields bumped into each other - this creates new associations. Suddenly an idea is created in Architecture because the Architect had a conversation with an Evolutionary Biologist and applied a principle from that field to their field. Associative thinking is hard for machines, and how machines will supercharge humans to be better at associative thinking is not necessarily an AI problem, but will be up to how we build AI into products to help us make novel associations.
• Synthetic Machine Thinking. This is where machines excel. There are already neural networks that create music and food dishes at a high level. This is because combining things in new ways can be brute-forced. As long as you can define what good is, so the millions of combinations can be tested. AlphaFold for proteins is another. The properties of the molecules are known, the machine can combine it in millions of ways and simulate how it will behave, and then isolate for novelty and utility. As long as humans can define what we are looking for, the hard work of trying different things can be automated.
• Contradictory Machine Thinking. This is perhaps what machines are worst at. Trained on huge amounts of data that is generated by humans, they are bound to have the biases and assumptions of humans trained as well. The contrarian creatives will still be humans in my opinion, because they see what is not there, instead of seeing what is there. The ability to stop and say, what if... is still mostly in the realm of humans. Now, one can imagine that having conversations with machines could trigger contrarian ideas - if the machine was good at simulating outcomes. I could then supercharge my contrarian ideas by being able to validate them quickly. But the idea itself, would be the realm of humans.

As you can see. With all these different ways of thinking, there are still elements which is in the realm of human thinking. There is two commonalities across all different ways of thinking:

• Humans as the seed of creativity. For machines to help us with creating content, solutions or really anything - the seed is still human. A text without a purpose is meaningless... the machine does not create the purpose, human does.
• Humans as the evaluator of creativity. Machines can help us to create content and solutions. They can even select the best solutions. But the definition of best, the definition of quality is human... the machine does not define what a good artwork is.

So when we play around with Generative AI like ChatGPT, Dall-E and MidJourney - and the now thousands of services popping up to harness the power of these models - keep in mind that the machine is not making you more creative. All it's doing is to accelerate your ability to act on your creativity. The ability to seed, and the ability to evaluate (or define a way to evalute)... is distinctly human. So nourish your ability to do creative thinking, and use generative AI to supercharge your ability to execute ideas.

#ChatGPT, #generativeai, #conversationalai, #ai, #creativity