On generative AI and creativity

“Where were you when I laid the foundations of the earth? Tell me, if you understand” God rebuked Job’s lamenting in Job 38:4-7, scoffing a man’s futile attempts to understand anything beyond his petty existence.

A verse that set the stage for Terrence Malick’s 2011 masterpiece “The Tree of Life”, in which - at least through a theological lens - the two paths to go through life are painted: that of nature (brute, prideful, self-interested, temporal) or that of grace (soft, humble, accepting, eternal). Both paths interlinked with excruciating beauty through the notion of failing to understand the vastness of creation: as (imitation of a) reality, an act, a concept, a purpose maybe even.

Pretty much all civilisations that are part of written history have contributed creation - or creativity, the ability to produce out of nothing - to either nature, or the supernatural or divine; for some, this could also mean the same.?Human creation on the other hand, was merely the result of reproduction or combination of existing ideas, abiding to specific rules, and thus more craft than creativity, with one particular exception though: poetry. A conviction rooted in Greek philosophy, it was Horace that wrote that only poets possess the privilege to daringly conceptualise as they desire, void of preconception.

Player four has entered the chat

In that sense one could call it somewhat ironic that the fourth protagonist that has entered the creative stage - technology - has been primed to exploit our weakness for arguably the single most transformative human invention, language. AI hijacking our idiom has left us vulnerable to project all sorts of human qualities to computed output, galvanised by its makers - often with benign intent - to accomplish exactly that. And imagination (or better, ingenuity) is a somewhat unexpected human quality attributed to these generative AI models as well, in particular since the emergence of generative adversarial networks and transformer architectures.

Because weren't AIs supposed to be all logic, and no intent or imagination? Sure, HAL 9000 went rogue after he decided he had no choice but to kill the crew of Discovery One, but only to resolve a conflict in his directives. Today, we find ourselves discussing less lethal or existential threats (though prominent thinkers like Yudkowksy and Harari still do), yet watch the discourse gravitate around professional insecurities mostly, fuelled by the disconcern on the origin of (digital) artefacts: have they been produced by people or their machines or perhaps more visceral, conceived?

So, can AI be considered creative?

Theologic and philosophical contemplation on the nature of creativity has developed into more human-centric and scientific studies over the last century, attempting to converge to a conclusive definition of creativity; an ambition that proved to be equally susceptible to recursion and subjectivity as is defining what is art and what is not. Absent of absolute measure, inevitably relative scales and statistics came into play, with psychological tests developed to measure the ability of humans to come up with novel, surprising, meaningful and valuable ideas, despite well-known flaws in these tests. Flaws that weren't a big problem, until LLMs were able to pass all of these tests, with GPT-4 matching the top 1% of human thinkers. Various papers (see this, that or over here) have been published recently on the topic of measuring human and AI actors their respective or collaborative creative qualities, with some interesting conclusions:

• AI is already better than most people to generate creative ideas;
• Human top performers still outperform AI, and benefit less from human-AI collaboration than the majority of the population does;
• Current AI models display more similarity in their ideas than large groups of humans do.

Thus, in the most practical sense, even those who don't see themselves as creative, now have a vast array of tooling at their disposal that can produce ideas surpassing most humans. The result of this democratization however - at least on the surface - appears consistent with the ripple effects of most (digital) technology advancements of the last decades: the potential created for individual advocacy and mass enablement at increasingly lower cost of access has mostly resulted in a bottom-up explosion of noisy goo of sameness, driven by the amygdalan urges for attention, affection and affirmation.

What defines truly creative minds?

Infinitely more of the same doesn't necessarily feel like progress. So in order to better understand the trajectory we might be on in terms of computational creativity, perhaps it's worthwhile to focus on its edges and not its mass: what do top creative performers across various disciplines have in common?

In 2019, Arthur Miller published a meta-study on creative traits and its definitions, distilling seven hallmarks that shape excellence in (human) creativity:

• the need for introspection;
• the need to know your strengths;
• the need to focus, persevere and not be afraid to make mistakes;
• the need for collaboration and competition;
• the need to beg, borrow or steal great ideas;
• the need to thrive on ambiguity;
• the need for experience and suffering.

Furthermore, Miller identified two marks of high-caliber creativity, that are ill-taught:

• the ability to discover the key problem;
• the ability to spot connections.

Now, if we were to evaluate these characteristics on their fit with technology as a truly creative agent, a few observations emerge. First, all of these first seven traits appear closely tied to human behaviour and attitude, and not so much skill, or the ability to acquire them as such. What opportunity would these render to be modelled as high-dimensional vectors? Secondly, there appears to run a somewhat ethical thread throughout these traits as well. In particular the need for suffering stands out, and immediately sparks impressions of Van Gogh, of Picasso, Pollock, Kahlo, Klimt, Gaugain and so many others. Yet more importantly, it shapes the notion that suffering by definition requires consciousness, which is something completely different than intelligence; listen to Lex Fridman and Yoval Noah Harari speak on this at length in Lex's podcast.

Lastly, interestingly enough, it's the latter two traits of exceptional talent that appear to be prime candidates for algorithmic representation. Deduction, logic and extremely complex interconnectedness are some of the architectural foundations of current AIs. So why do we have this schism between the potential for excellence and the reality of mostly average or at best good (yet extremely fast) reproduction? Are those first traits preconditions? In short, is consciousness? And if so, are the computational behemoths that LLMs are, the 'right' direction?

Intent and execution

Questions well beyond my own intelligence I have to admit. Yet there is an angle to explore without stretching once's philosophical capabilities immediately. When dissecting the creative process - fully aware that it is anything but a linear flow - there's an important distinction to be made between the stages that shape the intent, and those that constitute the execution.

Perhaps it's best explained through one of my personal favourite post-modern art series, the Spot Paintings by Damien Hirst. His fascination - if not obsession - with the violent energy and petrified emotion of abstract expressionism met his 'contempt for the coldness of contemporary minimalism' of the Eighties in the realisation that a viewer can also be moved by the absence of emotion felt by the painter.

Adhering to a very strict set of rules, the paintings unfold as a grid of dots, seemingly mechanically placed at fixed intervals with mathematical precision. Every dot is coloured, yet no repetition is applied. So 'simple' that there's a Github repository with a source file little over 100 lines of code perfectly capable of reproducing a single work. No intelligence whatsoever - apart from the developer who wrote it - is involved.

Yet the quality of these works lies in a thorough understanding of human vision, perception, emotion and intent, something we all to various degrees carry. And the effect of looking at a work is only equalled by being confronted with a Rothko or a Jackson Pollock; an overwhelming feeling of unease, disempowerment and eventually content to have been part of that experience.

On the level of the series, there's another important quality to identify. For Hirst, the art lies in the conception, not the execution, a conviction and approach that has worked well throughout history: Raphael, Rembrandt, Rubens, Warhol, Koons and many others all have operated assembly lines to scale their production, without affecting its (perceived) value.

But what Hirst did better than, or before anyone else, was make the execution part of the intent - if not the intent itself - criticising and establishing a new form of art simultaneously; in his studio, interns were teaching other interns to paint the works according to those simple rules. And to further proof his point, generating considerable monetary value with limited marginal cost, even enjoying the act of professional buyers taking it one step further by cutting individual works up into singular dots and selling those at a premium, effectively extending and reinforcing the reach of the original intent.

Not until something has been done, it can appear so straightforward, like with so many things. The novelty, its value lies in the intent, the prompt as you will, and the consciousness to firmly place it in context and culture.

Dream on

And thus, back to generative AI. We've established that it's almost infinitely more cost-effective at execution than humans are. We've also established that on intent, humans still outperform AIs.

In that perspective, it's interesting to weigh Sam Altman's remarks a few weeks ago in a conversation with Marc Benioff during Dreamforce '23 on creativity, proclaiming that “one of the sort of non-obvious things is that a lot of value from these systems is heavily related to the fact that they do hallucinate”.

If we try to demystify the mechanics of these phenomena (just earlier that same day labeled by Benioff as "lies" in support of Salesforce's valid narrative on trust by the way), no hints of imagination surface; most causes seem to originate from a vector database's architecture or a model's training through either:

• Considerate gaps in training data that open up exposure to edge cases unfamiliar to the models deployed;
• Overfitting of a model on particular data sets that demote generalisation of training instances;
• Ambiguity in encoding/decoding, the proces of tokenising words and their semantics into vectors and vice-versa.

All three of these causes for adversarial hallucinations are classified by researchers as high-dimensional statistical phenomena, of which none seem unsolvable. If we were to extend this thinking, one could argue that these models will become 'better' if they 'dream' less. I would argue that these models can dream all they 'want', but their trajectory isn't one steering towards creative brilliance.

Therefore, to quote a dear friend, maybe artists are the people that should be least worried about their future, as they are better equipped than most of us to identify and articulate true novel intent. I would like that.