
The writer is founder of Sifted, an FT-backed media company covering European start-ups
As we now know, generative AI models, such as ChatGPT and DALL-E, are great at making stuff up. Mostly, this is harmless fun, generating faux Beyoncé lyrics or a mock Raphael painting of a Madonna and child eating pizza. Sometimes, however, this ability to “hallucinate” facts generates disinformation and deep fakes.
Ideally, we want machines to make up stuff that is reliable, not just plausible, and to extend the range of human insight. Can we use machine learning models to generate truly novel ideas in hard areas, including mathematics and science, and enrich human creativity? It is beginning to look that way.
First, though, we should define creativity. Margaret Boden, research professor at Sussex university, has three useful classifications: combinational, exploratory and transformational. Creativity can be about mashing together improbable combinations of familiar ideas (think poetic imagery) or exploring new conceptual spaces (think jazz) — the most common forms, she suggests. Machines are pretty good at these pattern recognition and replication tasks, as generative AI models have shown.
But the most elusive, and arguably most valuable, kind of creativity is transformational: generating previously unimagined ideas or concepts. “The ultimate vindication of AI-creativity would be a program that generated novel ideas which initially perplexed or even repelled us, but which was able to persuade us that they were indeed valuable,” Boden wrote in a paper published in 1998.
Since then, machine creativity has failed Boden’s test. But at a recent discussion hosted by the London Institute for Mathematical Sciences, researchers highlighted how machine learning models were beginning to reshape mathematics by generating important new conjectures, subsequently proved to be true, in areas such as knot theory. “The field is at a delicious inflection point: the possibilities have been proven but it has not yet been fully explored,” one participant said.
Mathematics may be a particularly good testing ground for transformational machine creativity. ChatGPT is an impressive linguistic correlation machine, predicting what the next words in any sentence are statistically likely to be. Other types of machine learning models, however, can be trained to generate more substantive, and provable, outputs. “There are no coincidences in pure mathematics. It is either true or not true,” as one participant noted.
According to Alex Davies, the founding lead of the AI for Maths initiative at DeepMind, owned by Google parent Alphabet, machine learning models can help tackle three barriers to creativity: boredom, shame and vision. AI can perform dull, repetitive tasks. They also have no shame in generating apparently dumb, but occasionally inspired, outputs, which humans might be too embarrassed to air.
But the most intriguing aspect is vision. The promise is that machine learning tools can explore mathematical problems in dimensions that humans do not fully grasp. At present, Davies says experts are working with machine learning systems to guide their intuition towards creative results, but he anticipates further progress. “I believe we will eventually see machine learning have a transformational impact in mathematics,” he tells me.
That all sounds highly theoretical. DeepMind has already used machine learning models to predict the 3D structures of more than 200mn proteins. “It’s sort of like unlocking scientific exploration at digital speed,” said Demis Hassabis, co-founder of DeepMind. What does it mean in practice today? Researchers are using these models to help create wholly new proteins and to originate hypotheses for new battery materials. They are also using them to design antiviral drugs.
For the moment, companies in this sector remain mostly “financial optimisation machines”, using machine learning models to streamline processes and maximise earnings, says Martin Reeves, co-author of The Imagination Machine. But he suggests that they will increasingly have to become “imagination machines” that can use machine learning models to boost creativity and develop new products and services. “More than a visionary or a poet, companies imagine things that do not yet exist. We need that imagination more than ever to deal with issues like climate change,” he says.
The artist Pablo Picasso, a prime example of human creativity, is reputed to have said that “computers are useless, they can only give you answers”. Life will become more interesting as they help us formulate original questions.