CoChat
Sign Up
Cited by 113Ethics and Social Impacts of AICreativity in Education and NeuroscienceLiterary and Philosophical Studies

On the creativity of large language models

Open access
Franceschelli, Giorgio, Musolesi, Mirco·2025·AI & SOCIETY
DOI 10.1007/s00146-024-02127-3
Open this paper in CoChat — free account

Abstract

Large language models (LLMs) are revolutionizing several areas of Artificial Intelligence. One of the most remarkable applications is creative writing, e.g., poetry or storytelling: the generated outputs are often of astonishing quality. However, a natural question arises: can LLMs be really considered creative? In this article, we first analyze the development of LLMs under the lens of creativity theories, investigating the key open questions and challenges. In particular, we focus our discussion on the dimensions of value, novelty, and surprise as proposed by Margaret Boden in her work. Then, we consider different classic perspectives, namely product, process, press, and person. We discuss a set of “easy” and “hard” problems in machine creativity, presenting them in relation to LLMs. Finally, we examine the societal impact of these technologies with a particular focus on the creative industries, analyzing the opportunities offered, the challenges arising from them, and the potential associated risks, from both legal and ethical points of view.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper argues that current LLMs can produce valuable and weakly novel/surprising artifacts, but their autoregressive and RLHF/DPO-aligned nature limits them to combinatorial or exploratory creativity, not transformational creativity. It frames machine creativity through Boden’s three criteria and Rhodes’s product/process/press/person perspectives, then discusses legal and societal implications for creative industries.

Key findings

  • Under Boden’s three criteria, LLMs can satisfy value and a weak version of novelty and surprise, but P-/H-novelty and surprise are challenging because generation is probabilistic imitation of the training distribution; LLMs may at most express combinatorial or exploratory creativity, while transformational creativity is not achievable with current training solutions.
  • Beyond products, a creative process requires motivation, thinking, perception, and self-evaluation; LLMs lack intention, a self-feedback loop, and only output likely continuations. The “hard problem” is intentionality and self-awareness, and current AI systems are not conscious according to the best-supported neuroscientific theories.
  • The social/press dimension requires adaptation within a never-ending creative cycle; LLMs are frozen after training and cannot truly adapt or retain experience across environments, though in-context learning can simulate some adaptation. The person perspective requires consciousness and self-awareness, so creative-person qualities are out of discussion.
  • Practical implications include copyright protection requiring human original contribution, risks from memorization of protected training data, job displacement and deception risks, and opportunities for human-AI co-creativity. No quantitative results are reported.

Methods

  • Design: theoretical and philosophical/conceptual analysis, not an empirical study.
  • Data or sample: Not stated in the paper; no dataset or sample is used.
  • Main analysis: examines LLMs through Boden’s value, novelty, and surprise; Rhodes’s product, process, press, and person perspectives; easy and hard problems in machine creativity; and legal/ethical implications.

Limitations

  • The paper does not state limitations of its own analysis. Evident: it is conceptual and lacks empirical experiments or quantitative evaluation.
  • Evident: claims about “current” LLMs are tied to the state of the art at the time of writing and may not apply to future architectures or training methods.
  • Evident: the argument relies on selected philosophical frameworks and literature synthesis rather than a systematic review or primary data.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s00146-024-02127-3

Cites

93 of 93 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 53 cited sources checked · 0 supported · 0 flagged

  • Interpretationcitation
    § 3 Large Language Models and Boden's Three Criteria · p. 4

    The artifacts produced by LLMs are valuable.

    “Due to the large impact LLMs are already having (Bommasani et al., 2021) and the quality of outputs of the systems based on them (Stevenson et al., 2022b), it is possible to argue that the artifacts produced by them are indeed valuable.”
    • Also stated in: § 3 Large Language Models and Boden's Three Criteria · p. 5 · plain
      “In conclusion, while LLMs are capable of producing artifacts that are valuable, achieving P- or H-novelty and surprise appears to be more challenging.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Bommasani et al., 2021

      Not checked in the free run.

    • cites Stevenson et al., 2022b

      Not checked in the free run.

  • Interpretationreasoning
    § 3 Large Language Models and Boden's Three Criteria · p. 5

    The product of an LLM can appear to be creative-or be B-creative-even if it is not truly creative according to the theory of creativity.

    “In other words, the product of an LLM can appear to be creative - or be B-creative - even if it is not truly creative according to the theory of creativity.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: LLM outputs are often considered creative by the person interacting with them or exposed to their best productions.
  • Interpretationcitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 5

    The fact that LLMs will be able to generate creative outputs will not make them intrinsically creative.

    “However, the fact that they will be able to generate these outputs will not make them intrinsically creative.”
    Quote in the paperNot a substantial claim
  • Interpretationreasoning
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    The easy problems in machine creativity can be solved by considering the underlying training and inference processes.

    “these appear as easy problems to solve in order to achieve creativity, since solutions to them can be identified by taking into consideration the underlying training and inference processes.”
    Quote in the paperSubstantial claim
  • Interpretationreasoning
    § 4 Easy and Hard Problems in Machine Creativity · p. 7

    Placing LLMs in a different environment that requires a different prompt will make them start over, without the possibility of leveraging previously acquired experience.

    “Placing them in a different environment that requires a different prompt will make them start over, without the possibility of leveraging previously acquired experience.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 38
  • May combine two claims. 20

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationcitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    Individuals and their works cannot be isolated from the social and historical milieu in which their actions are carried out.

    “Individuals and their works cannot be isolated from the social and historical milieu in which their actions are carried out.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    LLMs can correct, modify, or rephrase outputs only by guessing what is the most likely re-casting of such responses or through the application of a set of given rules.

    “However, they would do it only by guessing what is the most likely re-casting of such responses or through the application of a set of given rules.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: No current LLM is able to self-evaluate its own responses.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § 4 Easy and Hard Problems in Machine Creativity · p. 7

    Fine-tuning actually updates network weights, but it requires a potentially large training dataset.

    “On the other hand, fine-tuning actually updates network weights, but it requires a potentially large training dataset.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning
    § 4 Easy and Hard Problems in Machine Creativity · p. 7

    Continual learning for LLMs represents a promising direction, yet unexplored for creative applications.

    “Continual learning (Kirkpatrick et al., 2017; Shin et al., 2017) for LLMs (Sun et al., 2020; Wu et al., 2022) represents a promising direction, yet unexplored for creative applications.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Kirkpatrick et al., 2017 source

      Not checked in the free run.

    • cites Shin et al., 2017

      Not checked in the free run.

    • cites Sun et al., 2020

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 7

    Increasing context length allows researchers to provide more information to LLMs without re-training them, although longer context might lead to performance degradation.

    “The constantly increasing context length (Hsieh et al., 2024) allows researchers to provide more and more information to LLMs without re-training them, although a longer context might lead to performance degradation (Li et al., 2024).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Hsieh et al., 2024

      Not checked in the free run.

    • cites Li et al., 2024

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation
    § 3 Large Language Models and Boden's Three Criteria · p. 4

    Strict alignment to very careful and pre-designed human responses leads to the generation of text that tends to be less diverse and that might be considered banal.

    “While in theory this could lead to potentially surprising generation, its strict alignment to very careful and pre-designed human responses leads to the generation of text that tends to be less diverse (Kirk et al., 2024) and that might be considered banal (Hoel, 2022).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Kirk et al., 2024

      Not checked in the free run.

    • cites Hoel, 2022

      Not checked in the free run.

    • Rests on an assumption: ChatGPT and all the other state-of-the-art LLMs introduced in Section 2 are fine-tuned with RLHF or DPO.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § 4 Easy and Hard Problems in Machine Creativity · p. 7

    To be part of the never-ending creative cycle, LLMs should constantly adapt.

    “In order to be able to be part of the never-ending creative cycle mentioned above, LLMs should constantly adapt.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 3 Large Language Models and Boden's Three Criteria · p. 4

    LLMs play a sort of imitation game, without a focus on computational novelty.

    “LLMs still play a sort of imitation game , without a focus on (computational) novelty (Fazi, 2019).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Fazi, 2019 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    LLMs cannot currently adapt through multiple iterations; they rely on one fixed version of the domain and generate works based on it.

    “However, LLMs cannot currently adapt through multiple iterations in the way described above; they just rely on one, fixed version of the domain and generate works based on it.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    The hard problem in machine creativity is about the intentionality and the self-awareness of the creative process in itself.

    “The hard problem in machine creativity is about the intentionality and the self-awareness of the creative process in itself.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given
    § 3 Large Language Models and Boden's Three Criteria · p. 5

    Our perception does not usually align with theoretical definitions of creativity.

    “we can explain this phenomenon by considering the fact that our perception does not usually align with theoretical definitions of creativity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § 3 Large Language Models and Boden's Three Criteria · p. 5

    LLMs may be deemed able to generate creative products if we assume the definition of combinatorial creativity.

    “It is possible to argue that LLMs may be deemed able to generate creative products if we assume the definition of combinatorial creativity.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    Four perspectives have to be considered to explain creativity: product, process, press, and person.

    “Rhodes (1961) theorizes that four perspectives have to be considered: product (see Section 3) and process (discussed above), but also the so-called press and person .”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Rhodes (1961)

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    A crucial aspect of the creative process is the perception and the ability of self-evaluating the generated outputs.

    “Indeed, a crucial aspect of the creative process is the perception and the ability of self-evaluating the generated outputs (Amabile, 1983).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Amabile, 1983 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationcitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 7

    Properties related to the creative person appear out of discussion for generative AI.

    “While several of the properties of press and process might be achieved - or at least simulated - by generative learning solutions, those related to the creative person appear out of discussion (Browning, 2023).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Browning, 2023 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given
    § 3 Large Language Models and Boden's Three Criteria · p. 5

    Current probabilistic solutions are intrinsically limiting in terms of expressivity.

    “current probabilistic solutions are intrinsically limiting in terms of expressivity.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 6

    Products have to be accepted as creative by the society, and producers are influenced by previously accepted works, i.e., the domain.

    “Products have to be accepted as creative by the society, and producers are influenced by the previously accepted works, i.e., the domain (Csikszentmihalyi, 1988).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Csikszentmihalyi, 1988

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 5

    LLMs stop at the first stage of creative learning, i.e., imitation, not implementing the remaining ones, i.e., exploration and intentional deviation from conventions.

    “In other words, they stop at the first stage of creative learning, i.e., imitation, not implementing the remaining ones, i.e., exploration and intentional deviation from conventions (Riedl, 2018).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Riedl, 2018

      Not checked in the free run.

    • Rests on an assumption: LLMs do not contain a self-feedback loop.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § 3 Large Language Models and Boden's Three Criteria · p. 5

    People do not typically judge the creativity of a product by considering its potential novelty and surprise in relation to its producer, but rather in relation to themselves.

    “Indeed, we do not typically judge the creativity of a product by considering its potential novelty and surprise in relation to its producer, but rather in relation to ourselves.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 5

    It is not what is achieved but how it is achieved that matters.

    “Indeed, as Floridi and Chiriatti (2020) puts it, it is not what is achieved but how it is achieved that matters.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Floridi and Chiriatti (2020) source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning
    § 3 Large Language Models and Boden's Three Criteria · p. 5

    Alternative learning architectures are probably necessary to achieve transformational creativity.

    “To achieve transformational creativity, alternative learning architectures are probably necessary; in fact, current probabilistic solutions are intrinsically limiting in terms of expressivity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation
    § 5 Practical Implications · p. 8

    The societal impact of LLM works will be tremendous.

    “Whether or not LLM works obtain protection, we believe their societal impact will be tremendous (see also Newton and Dhole (2023)).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Newton and Dhole (2023)

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning
    § 5 Practical Implications · p. 8

    More creative LLMs would diverge more consistently from existing works, reducing the risk of capitalizing on others' ideas.

    “Nonetheless, more creative LLMs would diverge more consistently from existing works, reducing the risk of capitalizing on others' ideas.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning
    § 5 Practical Implications · p. 9

    LLMs can adapt the same text to different styles, allowing an artifact to reach wider audiences.

    “LLMs can also adapt the same text to different styles (see combinatorial creativity in Section 3): by doing so, an artifact can be adapted to reach wider audiences.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 5 Practical Implications · p. 8

    Mitigation of such threats through dedicated policies or designed mechanisms of watermarks are already being developed.

    “Mitigation of such threats through dedicated policies 8 or designed mechanisms of watermarks (Kirchenbauer et al., 2023) are already being developed.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Kirchenbauer et al., 2023

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning
    § 5 Practical Implications · p. 9

    The role of writers and journalists may be transformed, but not replaced.

    “This means that the role of writers and journalists may be transformed, but not replaced.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisecitation
    § 5 Practical Implications · p. 9

    Generated responses from LLMs may augment writers' inherently multiversal imagination.

    “In the same way, the generated responses may augment writers' inherently multiversal imagination (Reynolds and McDonell, 2021).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Reynolds and McDonell, 2021

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 5 Practical Implications · p. 8

    LLMs are trained in a supervised fashion on real data, which also include protected works.

    “In particular, LLMs are trained in a supervised fashion on real data, which also include protected works (Bandy and Vincent, 2021).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Bandy and Vincent, 2021

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 4 Easy and Hard Problems in Machine Creativity · p. 7

    LLMs have no self to which to be true when generating text and are intrinsically unable to behave authentically as individuals.

    “As Ressler (2023) pointed out, LLMs have no self to which to be true when generating text and are intrinsically unable to behave authentically as individuals.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Ressler (2023)

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning
    § 5 Practical Implications · p. 9

    Humans are still required for prompting, curation, and pre-/post-production.

    “Given their characteristics, humans are still required, especially for prompting, curation, and pre-/post-production.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation
    § 5 Practical Implications · p. 8

    One last threat may be posed by human and artificial works being indistinguishable.

    “Finally, one last threat may be posed by human and artificial works being indistinguishable (Dehouche, 2021).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Dehouche, 2021 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 5 Practical Implications · p. 8

    Users obtaining LLM outputs might claim them as authors for deceiving readers, cheating during exams, or improving bibliometric indicators.

    “The users obtaining such outputs might therefore claim them as the authors, e.g., for deceiving readers (Grinbaum and Adomaitis, 2022), for cheating during exams (Fyfe, 2023), or for improving bibliometric indicators (Crothers et al., 2022).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Grinbaum and Adomaitis, 2022

      Not checked in the free run.

    • cites Fyfe, 2023 source

      Not checked in the free run.

    • cites Crothers et al., 2022

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § 5 Practical Implications · p. 8

    The lack of current copyright protection for generated works can foster replacements for tasks where a free-of-charge text would be preferable to a high-quality but costly one.

    “The lack of current copyright protection for generated works can also foster such replacements for tasks where a free-of-charge text would be preferable to a high-quality (but still costly) one.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation
    § 5 Practical Implications · p. 9

    LLMs represent a valuable tool in scientific research, especially for hypothesis generation.

    “In the same way, LLMs also represent a valuable tool in scientific research (Fecher et al., 2023), especially for hypothesis generation (Gero et al., 2022).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Fecher et al., 2023

      Not checked in the free run.

    • cites Gero et al., 2022

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning
    § 5 Practical Implications · p. 8

    Current LLMs might satisfy the originality condition for copyright, but they cannot be considered creative persons, therefore missing the personality requirement.

    “As discussed earlier, current LLMs might satisfy the first condition, but they cannot be considered creative persons, therefore missing the latter requirement.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given
    § 5 Practical Implications · p. 8

    Current LLMs are not as reliable as humans; they cannot verify their information and they can propagate biases from training data.

    “However, we must remind that current LLMs are not as reliable as humans, e.g., they cannot verify their information and they can propagate biases from training data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning
    § 5 Practical Implications · p. 9

    LLMs provide new opportunities for humans, who will be able to spend more time validating news or thinking up and testing ideas.

    “On the contrary, LLMs provide new opportunities for humans, who will be able to spend more time validating news or thinking up and testing ideas.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given
    § 6 Conclusion · p. 9

    LLMs are capable of value and a weak version of novelty and surprise.

    “While LLMs are capable of value and a weak version of novelty and surprise, their inner autoregressive nature seems to prevent them from reaching transformational creativity.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 6 Conclusion · p. 9

    LLMs provide opportunities for writers, especially in terms of human-AI cooperation.

    “However, LLMs also provide opportunities for writers, especially in terms of human-AI cooperation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 6 Conclusion · p. 9

    Specific fine-tuning techniques might help LLMs diversify productions and explore the conceptual space they learn from data.

    “Specific fine-tuning techniques might help LLMs diversify productions and explore the conceptual space they learn from data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 6 Conclusion · p. 9

    The current legal framework does not appear to be completely suited to the fast-moving field of generative AI.

    “In fact, the current legal framework does not appear to be completely suited to the fast-moving field of generative AI.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning
    § 6 Conclusion · p. 9

    LLMs' adaptive abilities are only at a very initial stage.

    “The social dimension of creativity (usually referred to as the press) would demand to be placed in and influenced by a society of creative agents, requiring LLMs adaptive abilities that are only at a very initial stage.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The social dimension of creativity requires being placed in and influenced by a society of creative agents.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 6 Conclusion · p. 9

    Continual learning can enable long-term deployments of LLMs in a variety of contexts.

    “Continual learning can enable long-term deployments of LLMs in a variety of contexts.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation
    § 5 Practical Implications · p. 9

    Through human-AI co-creation, LLMs may eventually allow the development of entire new arts.

    “Being intrinsically powerful tools, through human-AI co-creation, LLMs may eventually allow the development of entire new arts, as has been the case for any impactful technology in the past centuries (Eisenstein, 1979; Silva, 2022).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Eisenstein, 1979

      Not checked in the free run.

    • cites Silva, 2022

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given
    § 6 Conclusion · p. 9

    Current LLMs do not possess motivation, thinking, and perception.

    “A creative process would require motivation, thinking, and perception, properties that current LLMs do not possess.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: A creative process requires motivation, thinking, and perception.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given
    § 6 Conclusion · p. 9

    All these techniques would only simulate certain aspects of creativity.

    “While, of course, all these techniques would only simulate certain aspects of creativity, whether this would be sufficient to achieve artificial, i.e., non-human, creativity, is up to the humans themselves.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given
    § 6 Conclusion · p. 9

    The impact of LLMs on creative professions and the arts will definitely be considerable.

    “Moreover, the impact of these technologies on creative professions and the arts is difficult to forecast at this stage, but will definitely be considerable.”
    Quote in the paperNot a substantial claim
47 of 47

  • The paper relies on this without support.

    The question of whether LLMs can be considered creative has not been considered yet.

    “However, a critical question has not been considered yet: can LLMs be considered creative ?”

  • The paper relies on this without support.

    LLM outputs are often considered creative by the person interacting with them or exposed to their best productions.

    “Nonetheless, the outputs from such models are often considered creative by the person interacting with them or exposed to their best productions.”

  • The paper relies on this without support.

    LLMs are commonly trained on all available data.

    “since LLMs are commonly trained on all available data.”

  • The paper relies on this without support.

    LLMs lack the intention to write.

    “However, LLMs lack the intention to write.”

  • The paper relies on this without support.

    LLMs do not contain a self-feedback loop.

    “Again, LLMs do not contain such a self-feedback loop.”

  • The paper relies on this without support.

    No current LLM is able to self-evaluate its own responses.

    “However, no current LLM is able to selfevaluate its own responses.”

  • The paper relies on this without support.

    The quality of the output strictly depends on the prompt.

    “In addition, the quality of the output strictly depends on the prompt, which might in turn demand human skills and more time.”

  • The paper relies on this without support.

    LLMs can be used to produce artifacts that would be protected if made by humans.

    “Since LLMs can be used to produce artifacts that would be protected if made by humans, a first concern is the definition of legal frameworks in which they will be used.”

  • The paper relies on this without support.

    ChatGPT and all the other state-of-the-art LLMs introduced in Section 2 are fine-tuned with RLHF or DPO.

    “ChatGPT and all the other state-of-the-art LLMs introduced in Section 2 are finetuned with RLHF or DPO.”

  • The paper relies on this without support.

    LLMs are not trained to directly maximize value, novelty, or surprise.

    “At the same time, they are not trained to directly maximize value, novelty, or surprise.”

  • The paper relies on this without support.

    A creative process requires motivation, thinking, and perception.

    “A creative process would require motivation, thinking, and perception, properties that current LLMs do not possess.”

  • The paper relies on this without support.

    The social dimension of creativity requires being placed in and influenced by a society of creative agents.

    “The social dimension of creativity (usually referred to as the press) would demand to be placed in and influenced by a society of creative agents, requiring LLMs adaptive abilities that are only at a very initial stage.”
12 of 12

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Cites

Flagged first, then by citations

Research works printed but not linked: 50

  1. Amabile TM (1996) Creativity in context. Routledge, London
  2. Asay CD, Sloan A, Sobczak D (2020) Is transformative use eating the world? Boston Coll Law Rev 61(3):905–970
  3. Bahdanau D, Cho K, Bengio Y (2015) Neural machine translation by jointly learning to align and translate. In: Proc. of the 3rd international conference on learning representations (ICLR’15)
  4. Bandy J, Vincent N (2021) Addressing ”documentation debt” in machine learning: a retrospective datasheet for bookcorpus. In: Proc. of the 35th conference on neural information processing systems datasets and benchmarks track (round 1)
  5. Barron F (1955) The disposition toward originality. J Abnorm Psychol 51(3):478–485
  6. Boden MA (2009) Computer models of creativity. AI Mag 30(3):23–34
  7. Bonadio E, McDonagh L (2020) Artificial intelligence as producer and consumer of copyright works: evaluating the consequences of algorithmic creativity. Intellect Prop Q 2020(2):112–137
  8. Bond T, Blair S (2019) Artificial intelligence & copyright: section 9(3) or authorship without an author. J Intellect Prop Law Pract 14(6):423–423
  9. Bradley H, Dai A, Teufel H, Zhang J, Oostermeijer K, Bellagente M, Clune J, Stanley K, Schott G, Lehman J (2024) Quality-diversity through AI feedback. In: Proc. of the 12th international conference on learning representations (ICLR’24)
  10. Brown T, Mann B, Ryder N, Subbiah M, Kaplan JD, Dhariwal P, Neelakantan A, Shyam P, Sastry G, Askell A, Agarwal S, Herbert-Voss A, Krueger G, Henighan T, Child R, Ramesh A, Ziegler D, Wu J, Winter C et al (2020) Language models are few-shot learners. In: Advances in neural information processing systems (NIPS’20)
  11. Bunescu RC, Uduehi OO (2019) Learning to surprise: a composer-audience architecture. In: Proc. of the 10th international conference on computational creativity (ICCC’19)
  12. Calderwood A, Qiu V, Gero KI, Chilton LB (2020) How novelists use generative language models: an exploratory user study. In: Proc. of the IUI’20 workshop on human-AI co-creation with generative models
  13. Carlini N, Ippolito D, Jagielski M, Lee K, Tramer F, Zhang C (2023) Quantifying memorization across neural language models. In: Proc. of the 11th international conference on learning representations (ICLR’23)
  14. Chowdhery A, Narang S, Devlin J, Bosma M, Mishra G, Roberts A, Barham P, Chung HW, Sutton C, Gehrmann S, Schuh P, Shi K, Tsvyashchenko S, Maynez J, Rao A, Barnes P, Tay Y, Shazeer N, Prabhakaran V, Fiedel N (2023) PaLM: scaling language modeling with pathways. J Mach Learn Res 24(240):1–113
  15. Christiano PF, Leike J, Brown T, Martic M, Legg S, Amodei D (2017) Deep reinforcement learning from human preferences. In: Advances in neural information processing systems (NIPS’17)
  16. Dathathri S, Madotto A, Lan J, Hung J, Frank E, Molino P, Yosinski J, Liu R (2020) Plug and play language models: a simple approach to controlled text generation. In: Proc. of the 8th international conference on learning representations (ICLR’20)
  17. Deltorn JM (2017) Deep creations: intellectual property and the automata. Front Digit Humanit 4(3):1–13
  18. Devlin J, Chang MW, Lee K, Toutanova K (2019) BERT: pre-training of deep bidirectional transformers for language understanding. In: Proc. of the 2019 conference of the North American Chapter of the Association for Computational Linguistics: human language technologies, vol 1 (Long and Short Papers) (NAACL’19)
  19. Du N, Huang Y, Dai AM, Tong S, Lepikhin D, Xu Y, Krikun M, Zhou Y, Yu AW, Firat O, Zoph B, Fedus L, Bosma M, Zhou Z, Wang T, Wang YE, Webster K, Pellat M, Robinson K et al (2022) GLaM: efficient scaling of language models with mixture-of-experts. In: Proc. of the 39th international conference on machine learning (ICML’22)
  20. Franceschelli G, Musolesi M (2024a) Creative beam search: LLM-as-a-judge for improving response generation. In: Proc. of the 15th international conference on computational creativity (ICCC’24)
  21. Gervais DJ (2002) Feist goes global: a comparative analysis of the notion of originality in copyright law. J Copyr Soc USA 49:949–981
  22. Gervás P (2013) Computational modelling of poetry generation. In: Symposium on artificial intelligence and poetry (AISB’13)
  23. GPT-3, Thunström AO, Steingrimsson S (2022) Can GPT-3 write an academic paper on itself, with minimal human input? https://hal.archives-ouvertes.fr/hal-03701250v1. Accessed 30 Oct 2024
  24. Guadamuz A (2017) Do androids dream of electric copyright? Comparative analysis of originality in artificial intelligence generated works. Intellect Prop Q 2:1–24
  25. Hoffmann J, Borgeaud S, Mensch A, Buchatskaya E, Cai T, Rutherford E, de Las Casas D, Hendricks LA, Welbl J, Clark A, Hennigan T, Noland E, Millican K, van den Driessche G, Damoc B, Guy A, Osindero S, Simonyan K, Elsen E, Rae, JW, Vinyals O, Sifre L (2022) Training compute-optimal large language models. In: Advances in neural information processing systems (NIPS’22)
  26. Jurafsky D, Martin JH (2023) Speech and language processing (third (draft) ed.)
  27. Kirchenbauer J, Geiping J, Wen Y, Katz J, Miers I, Goldstein T (2023) A watermark for large language models. In: Proc. of the 40th international conference on machine learning (ICML’23)
  28. Kirk R, Mediratta I, Nalmpantis C, Luketina J, Hambro E, Grefenstette E, Raileanu R (2024) Understanding the effects of RLHF on LLM generalisation and diversity. In: Proc. of the 12th international conference on learning representations (ICLR’24)
  29. Lee HH, Shu K, Achananuparp P, Prasetyo PK, Liu Y, Lim EP, Varshney LR (2020) RecipeGPT: generative pre-training based cooking recipe generation and evaluation system. In: Companion proceedings of the web conference 2020 (WWW’20)
  30. Macedo L, Reisenzein R, Cardoso A (2004) Modeling forms of surprise in artificial agents: empirical and theoretical study of surprise functions. In: Proc. of the annual meeting of the Cognitive Science Society (CogSci’04)
  31. Maher ML (2010) Evaluating creativity in humans, computers, and collectively intelligent systems. In: Proc. of the 1st DESIRE Network Conference on Creativity and Innovation in Design (DESIRE’10)
  32. Meehan JR (1977) TALE-SPIN, an interactive program that writes stories. In: Proc. of the 5th international joint conference on artificial intelligence (IJCAI’77)
  33. Menabrea LF, Lovelace A (1843) Sketch of the analytical engine invented by Charles Babbage. In: Scientific memoirs, vol 3. Richard and John E. Taylor, pp 666–731
  34. Newton A, Dhole K (2023) Is AI art another industrial revolution in the making? In: Proc. of the AAAI-23 workshop on creative AI across modalities
  35. Pardinas R, Huang G, Vazquez D, Piché A (2023) Leveraging human preferences to master poetry. In: Proc. of the AAAI-23 workshop on creative AI across modalities
  36. Rafailov R, Sharma A, Mitchell E, Ermon S, Manning CD, Finn C (2023) Direct preference optimization: Your language model is secretly a reward model. In: Proc. of the 37th conference on neural information processing systems (NeurIPS’23)
  37. Raffel C, Shazeer N, Roberts A, Lee K, Narang S, Matena M, Zhou Y, Li W, Liu PJ (2020) Exploring the limits of transfer learning with a unified text-to-text transformer. J Mach Learn Res 21(140):1–67
  38. Rhodes M (1961) An analysis of creativity. Phi Delta Kappan 42(7):305–310
  39. Sawicki P, Grzés M, Jordanous A, Brown D, Peeperkorn M (2022) Training GPT-2 to represent two romantic-era authors: challenges, evaluations and pitfalls. In: Proc. of the 3th international conference on computational creativity (ICCC’22)
  40. Seth A (2021) Being you: a new science of consciousness. Penguin
  41. Shin H, Lee JK, Kim J, Kim J (2017) Continual learning with deep generative replay. In: Advances in neural information processing systems (NIPS’17)
  42. Stevenson C, Smal I, Baas M, Grasman R, van der Maas H (2022a) Putting GPT-3’s creativity to the (alternative uses) test. In: Proc. of the 13th international conference on computational creativity (ICCC’22)
  43. Stevenson C, Smal I, Baas M, Grasman R, van der Maas H (2022b) Putting GPT-3’s creativity to the (alternative uses) test. In: Proc. of the 13th international conference on computational creativity (ICCC’22)
  44. Stiennon N, Ouyang L, Wu J, Ziegler D, Lowe R, Voss C, Radford A, Amodei D, Christiano PF (2020) Learning to summarize with human feedback. In: Advances in neural information processing systems (NIPS’20)
  45. Sun FK, Ho CH, Lee HY (2020) LAMOL: LAnguage MOdeling for Lifelong Language Learning. In: Proc. of the 2020 international conference on learning representations (ICLR’20)
  46. Turner SR (1994) The creative process: a computer model of creativity and storytelling. Lawrence Erlbaum Associates, Inc., Hillsdale
  47. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L, Polosukhin I (2017) Attention is all you need. In: Advances in neural information processing systems (NIPS’17)
  48. Wertz L, Kuhn J (2022) Adapting transformer language models for application in computational creativity: generating German theater plays with varied topics. In: Proc. of the 13th international conference on computational creativity (ICCC’22)
  49. Wu T, Caccia M, Li Z, Li YF, Qi, G, Haffari G (2022) Pretrained language model in continual learning: a comparative study. In: Proc. of the 2022 international conference on learning representations (ICLR’22)
  50. Zhang Y, Gan Z, Fan K, Chen Z, Henao R, Shen D, Carin L (2017) Adversarial feature matching for text generation. In: Proc. of the 34th international conference on machine learning (ICML’17)

Cited by

Sorted by citations

Related

Authors
Franceschelli, G., & Musolesi, M.
Affiliations
  • University of Bologna (IT)
  • University College London (GB)
Corresponding author
  • Franceschelli, Giorgio
  • Musolesi, Mirco
ORCID
Venue
AI & SOCIETY
Publisher
Springer Science and Business Media LLC
Published
2024-11-28
Volume
40
Issue
5
Pages
3785-3795
ISSN
  • 0951-5666 (ISSN-L)
  • 1435-5655
DOI
10.1007/s00146-024-02127-3
OpenAlex
W4404826238
Semantic Scholar
1c7402843d8b586d945b3b030e3edd93f0ae3959
Cited by
113
References
93
Citations by year
2026: 47 · 2025: 39 · 2024: 20 · 2023: 7
FWCI
43.93
Type
journal article
Language
en
Open access
Licence
Locations
  • AI & Society (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • Archivio istituzionale della ricerca (Alma Mater Studiorum Università di Bologna) (submittedVersion, cc-by)
  • arXiv (Cornell University)
Topics
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Creativity in Education and Neuroscience
  • Social Sciences › Political Science and International Relations › Literary and Philosophical Studies
Keywords
large language models, machine creativity, creative writing, poetry generation, storytelling, Margaret Boden, creative industries, societal impact of AI

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review