Early on in the breathless “Chatbots are intelligent” hype cycle, there were a bunch of examples of chatbots spitting out verbatim copies of existing work. OpenAI argued that it was rare, various organisations argued that it was a copyright violation, and Large Language Model (LLM) peddlers around the world worked feverishly to prevent it from happening, or at least to make it less likely people would notice.
The thing is, reproducing chunks of text verbatim is not surprising behaviour from a system that is designed to produce statistically likely sentences. If those sentences appear repeatedly in its training data, then its internal model will rank them as highly statistically likely, and boom, plagiarism is your uncle. So to speak.
Copyright concerns aside*, this raises questions about the much vaunted creativity of LLMs. Sam Altman says that ChatGPT 5 is a “PhD level” intelligence. A major criteria for awarding a PhD is that the research must be a novel contribution to its field. It has to be new. It has to be creative. So can an LLM truly be creative? What is creativity anyway?
Dictionary.com defines “creative” as “having the power to bring something new into being, as a creature, or to evolve something original from one’s own thought or imagination, as a work of art or invention.”
I think the key here is “something original”. When I write something, everything I have ever experienced contributes to my state of mind. Everything I have ever read was, in some sense, training data for this piece. It shaped my language skills, my writing ability, and my thinking. How is that different from ChatGPT taking its training data, mashing it through a statistical process, and “creating” a new set of text? The chances are that some of the phrases I use have been used before. Some of my sentences are, no doubt, statistically plausible. (Though some, I suspect, are wildly implausible!)
Can I say for sure that my work is creative and ChatGPT’s is not?
For a long time we have argued that humans are the only creative species on Earth, and yet, increasingly, research is showing that a diverse range of animals can actually be creative, and solve complex problems. From crows to dolphins, chimpanzees to octopuses, animals show an astonishing level of creativity, once you go looking for it.
So what about ChatGPT and all of the other chatbots flooding the market? Are we wrong about those, too?
Well, here’s the thing. If your system creates the kind of sentences it has seen before – an inevitable outcome of the “statistically plausible” model – then it can only create … well… the kind of sentences it has seen before. It won’t create truly original, startling sentences like Douglas Adams’ famous “It hung in the air in much the same way as bricks don’t.” or Terry Pratchett’s wonderful “Racism was not a problem on the Discworld, because—what with trolls and dwarfs and so on—speciesism was more interesting. Black and white lived in perfect harmony and ganged up on green.” It won’t come up with a wizard suffering from burnout like Premee Mohamed did in “By Salt, by Sea, by Light of Stars” (my goodness, it couldn’t even come up with that name!).
Creativity requires a new take on something. An original perspective. Something no one else has thought of. LLMs, by definition, produce the kinds of things they have already seen. The kinds of things that already exist. The kinds of things that reinforce the status quo, entrench bias, and emphasise the mundane.
When you use a chatbot to produce a piece of writing, or an image, the only work it can possibly produce is something like the things it has seen before.
It can’t have a revelation.
It can’t provide a new perspective.
It can’t take an old idea, turn it around by 15 degrees and force you to look at it with fresh eyes.
It can’t produce something that makes you go “Wow! I’ve never thought of it like that.” (Unless it has nicked it from something someone has already written.)
It can’t do investigative journalism, or connect old things in new ways.
It can’t understand a complex system and describe it in simpler terms.
Creativity, it turns out, is statistically implausible.
Chatbots can’t create something truly new. And goodness knows, we could use some new thinking around here.
*(which is not to dismiss them – copyright concerns around Large Language Model training are a significant conversation that we should be having at a society level)