Catch up on the latest AI articles

Is There Creativity In LLM?

Is There Creativity In LLM?


3 main points
✔️ Consider whether LLM has creativity according to Margaret Boden's three criteria
✔️ Organize current issues in the creative process of machines
✔️ Organize issues in the application of LLM to creative tasks

On the Creativity of Large Language Models
written by Giorgio FranceschelliMirco Musolesi
(Submitted on 27 Mar 2023 (v1), last revised 9 Jul 2023 (this version, v3))
Comments: Published on arxiv.

Subjects:  Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY)


The images used in this article are from the paper, the introductory slides, or were created based on them.

First of all

Recent developments in large-scale language modeling (LLM) have been remarkable. Text-based reasoning capabilities, for example, are among the most prominent. On the other hand, how much creative thinking power does today's LLM have?

The paper presented here captures the creativity of LLMs based on creativity theory and organizes future issues.


Over the past decade, the field of natural language processing has seen a significant development, and along with the proposal of the Transformer model structure, large data sets and increased computational power have also contributed to this development. As a result, the recent mainstream of large-scale language models (LLMs) has attracted a large number of users, who have found a variety of creative ways to use them. For example, they write recipes, compose poems and stories, and so on.

On the other hand, it is not clear whether LLM itself is truly creative.

The paper presented here addresses this question based on three criteria proposed by a researcher named Margaret Boden.

From Ada Lovelace to foundation models

In 1843, Ada Lovelace criticized Charles Babbage's Analytical Engine, saying that computers were not interested in inventing anything new, but only in doing what we tell them to do. Alan Turing called this the Lovelace objection, and Turing himself argued that machines can never surprise us.

From this point on, a philosophical debate would ensue as to what constitutes human creativity and what constitutes machine creativity.

Computer scientists have attempted to create machines that allow creative self-expression through writing. With the advent of the personal computer and the development of techniques based on planning and evolutionary strategies, a variety of elemental technologies have been incorporated to propose systems for text generation.

And with the recent emergence of neural network systems, we can now see a major qualitative change. The elemental technologies range from recurrent neural networks with recursive structures to the recently mainstream Transformer structures.

The latter Transformer-based model requires a large model size and training data, but has broad applicability, and recent research has focused on how to utilize the learned large-scale model (the underlying model).

Margaret Boden's definition of creativity

Margaret Boden defined creativity as "the ability to come up with ideas or artifacts that are new, surprising and valuable. Margaret Boden defined creativity as "the ability to come up with ideas or artifacts that are new, surprising and valuable. These three aspects (novelty, surprise, and value) are the criteria for judging creativity.

Let's look at these three items in detail.


First, value refers to utility, performance, attractiveness, quality of output, and acceptance by society.

Because of its high quality output, LLM has a high impact on society, and from this perspective, what LLM produces is valuable.


Novelty refers to the dissimilarity between what is generated and other examples, the property of not having existed before.

(As for the scope of "before that," novelty in the individual is called P-creativity (psychological creativity), and novelty in the history of humanity as a whole is called H-creativity (historical creativity); there are two kinds of novelty.)

In principle, LLM is less dissimilar because its behavioral design is to return a probabilistically likely output.

If an output with a high novelty value were to occur, it would either be accidental or the result of an ingenious prompt.


Surprises refer to how much the input stimuli defy expectations.

There are three possible types of surprise, each with a different kind of creativity.

One is Combinational creativity, which combines familiar ideas in unfamiliar ways.

The other is Exploratory creativity, which discovers new, unexplored situations within the scope of current ways of thinking.

The last one is transformational creativity, which changes the current way of thinking itself.

Current LLMs, which make autoregressive predictions within a given distribution, may be able to achieve at least Combinational creativity, but Transformational creativity will be difficult, because it is fine-tuned by fine-tuning and RLHF to fit human values, for better or worse. Transformational creativity is difficult to achieve, because it is fine-tuned or fine-tuned by RLHF to fit human values, for better or worse. While this may reduce output that is offensive to humans, it may also lead to output of mediocre answers.

In summary, while the current LLM can produce valuable output, it still has many difficulties in producing output that offers novelty and surprise.

Problems in machine creativity

The ability of an LLM to produce creative output is not equivalent to the LLM itself being creative; as Floridi and Chiriatti (2020) argue in this paper, it is not what is accomplished, but how it is accomplished that matters.

According to Gaut (2003), creativity is the ability to produce something original and valuable through talent/sense. To exercise this talent/sense is to exercise relevant purpose, understanding, judgment, and evaluation skills, which are deeply related to processes such as motivation, perception, learning, thinking, and communication (Rhodes, 1961).

The current LLM is still in the mimetic stage, returning plausible outputs in response to input stimuli, and is not structured to allow for exploration or intentional modification, processes that are typical of human creativity.

Rhodes (1961) and Csikszentmihalyi (1988) also say that when evaluating creativity, it is necessary to take into account not only the product or process, but also who made it and how it is evaluated in relation to society and history. Creative acts occur in a cycle of interaction between the individual and society, but the current LLM is not designed to do this because the weights are fixed after a single learning session is completed. In the future, LLMs will be required to learn continuously.

Regarding practical use

The first problem in practical use is copyright.

Copyright applies to works that reflect a minimum of originality and the humanity of the author. The current LLM's creations may satisfy the first part, originality, but not the latter part, humanity, since the LLM does not have a personality of his own. Therefore, in the current situation, copyright applies to the user who wrote the prompt.

Another potential problem is that if the training data contains data from copyrighted works, the output may be a partial reproduction of those works.

There is also the risk that writing jobs will disappear as the ability to write papers and novels increases, and the problem of distinguishing human work from that produced by generative models.

Despite these issues, the authors of this paper anticipate that the impact of LLM technology will be positive, as it will allow more time to devise and test ideas and expand opportunities for creative activity.


The paper introduced here discussed the creativity of current LLMs and other generative models.

In line with Boden's three criteria of creativity (value, novelty, and surprise), it does not yet appear that the current LLM has all three.

He also introduced some of the issues that current LLMs face in the creative process and the problems they face in practical use.

Expectations are high for the future development of research on LLM creativity.

If you have any suggestions for improvement of the content of the article,
please contact the AI-SCHOLAR editorial team through the contact form.

Contact Us