4 Closing Thoughts: The Future of AI in Creative Applications
This paper has presented a comprehensive review of current AI technologies and their creative industries applications that have emerged in recent years. Generative methods have driven a rapid growth in AI usage, particularly in the creative sector, significantly advancing the state of the art across various applications such as content creation, information extraction and analysis, content enhancement and data compression.
Through these applications, generative AI has not only broadened creative possibilities, but has also reduced the manual effort and time traditionally associated with the production pipeline, allowing for greater creative experimentation and more rapid and agile production cycles. As this technology advances, it promises to unlock even more sophisticated capabilities. However, creative technologists, artists and other users must adapt, learn to use, and build these tools effectively and safely.
4.1 Challenges for AI in the Creative Sector
Artists are already exploring how to bridge structured nature of current creative AI and more traditional (analog) workflows. This includes using multiple methods, models, or tools to create new works. For example, artists can use tools to iterate on existing or past works; upload and fuse analog works or output from different Generative AI tools to further intervene in the AI generation process; and composite outputs to reconstruct or fill in missing or damaged parts of a work. However, one of the primary challenges for artists engaging with modern generative AI and LLMs is the lack of consistent, controllable outputs. These models operate via stochastic sampling from high-dimensional latent spaces, meaning that identical prompts can yield different results across runs. This unpredictability can make it challenging for artists to achieve, or iterate toward, a precise creative vision. Although prompt engineering has emerged as a technique to guide model behavior, it requires technical knowledge and iterative refinement, which may not align with the intuitive or exploratory approaches common in artistic practice.
Moreover, there can be a fundamental tension between the structured nature of current AI pipelines and current production, often improvisational, workflows used in creative disciplines. Many generative tools were originally designed for tasks like software development, content automation, or optimization Zhong et al. (2024), and are less well suited for open-ended, exploratory creation. Artists typically work in cycles of ideation, experimentation, and revision—processes that demand fluid, real-time interaction and control, which existing AI systems struggle to support. These limitations point to a gap in current AI design: a need for systems that not only generate high-quality content but also adapt to the iterative, interpretive nature of artistic production. One possible approach to address these challenges is a reinforcement of top-down creative workflows, where artists define high-level concepts, themes, or goals via text prompts before refining specific outputs. This approach helps align AI-generated results with artistic intent, offering a degree of control over inherently stochastic systems.
A further issue concerns creative authorship and ownership. As AI systems increasingly contribute to the ideation and execution of creative work, the line between human and machine authorship becomes blurred. Determining where creative credit lies—whether with the prompt designer, the model developer, or the AI system itself—poses significant legal and ethical challenges. Moreover, generative systems often reproduce stylistic elements from training data, raising questions about originality and cultural appropriation in AI-assisted creation.
Speaking at the World Government Summit in Dubai in 2024,[^56] NVIDIA CEO Jensen Huang argued that, with rapid advancements in AI, learning to code may become less essential for newcomers to the tech sector. He envisioned a future where traditional programming could be replaced by more intuitive AI-driven tools, thereby automating complex tasks and enhancing productivity—particularly for artists without coding expertise. While this perspective remains debated, it highlights the potential for AI to become more accessible within creative fields, not just in coding but across areas such as VFX and virtual production. But to achieve this, AI-assisted coding tools must be better integrated into creative workflows. Creators must exploit techniques such as fine-tuning pre-trained models, few-shot learning, or domain adaptation—methods that are powerful yet typically inaccessible without machine learning expertise.
There are also broader concerns that persist regarding the long-term impact of AI on the creative industries, economies, and labor markets. As automation accelerates, socioeconomic disparities may widen between those who can afford access to powerful generative systems and those who cannot. Freelancers and smaller studios risk being marginalized by large organizations with access to proprietary datasets and substantial compute resources. The potential emergence of artificial general intelligence (AGI). Envisioned by organizations like OpenAI, DeepMind, and Anthropic, AGI could surpass human cognitive abilities, raising ethical and existential questions about the role of human agency in artistic expression. Ensuring equitable access to AI tools and fair distribution of creative value will therefore be crucial to sustaining diversity, inclusion, and innovation within the sector.
4.2 Ethical Issues, Fakes and Bias
Advancements in generative AI, exemplified by models like Sora and Gemini 1.5 Pro, provoke ethical concerns and have societal implications. While their applications, with appropriate permission, can be beneficial and entertaining, these models, because they are capable of generating highly realistic content, escalate the risk of misuse through malicious deepfakes and misinformation. We are now in a situation where AI results transcend the uncanny valley, further complicating matters and challenging perceptions of authenticity. For example, the artist Miles Astray demonstrated that even authentic photographs could be mistaken for AI-generated images. His real photograph ‘F L A M I N G O N E’ won both the jury’s award and the people’s choice award in the AI category of the 1839 Awards. His aim was to highlight the ethical dilemmas inherent in AI, suggesting that the benefits of discussing AI’s ethical implications could surpass the ethical concerns related to viewer deception[^57].
While democratizing AI tools no doubt presents opportunities to transform creative processes and workflows, it also necessitates robust regulatory frameworks to safeguard privacy and ownership. For example, deepfake technologies stimulate significant concerns about the spread of misinformation and other malicious uses. Efforts to detect and identify increasingly realistic deepfakes are thus as important as the generative methods used to produce them. These must however be accompanied by increased media literacy, and policies that address the ethical and legal implications.
Diversity and representation is a key issue when using AI tools. Unified Concept Editing Gandikota et al. (2024) has been proposed as a basis for image generation in digital mediums. This aims to ensure the production of safe content with diverse representation, reducing gender and racial biases. Hallucination in generative AI (the production of outputs that are not faithful representations of reality but instead contain imagined or unrealistic elements) is a further cause of concern. These undermine trust in AI processes and can be due to limitations in the training data, biases in the model architecture or imperfections in the optimization process. Hallucinations associated with LLMs are one of the issues highlighted by the UK Government Communications and Digital Committee (2024), alongside bias, regurgitation of private data, difficulties with multi-step tasks and challenges in interpreting black-box processes.
Governments across the world are increasingly expressing concerns about the challenges and uncertainties that generative AI technologies pose to rights holders and human creativity Jeary and Gajjar (2024). Generative AI presents substantial legal challenges, including the copyright status of AI-generated work and the intellectual property and copyright implications of the datasets used in training AI models. Viewpoints on this issue do however differ. For example, the track ‘‘Heart on My Sleeve,” penned by an (as yet unidentified) human author, featured AI-generated vocals that replicated the voices of Drake and The Weeknd. Released independently on April 4, 2023, it was accessible via streaming platforms including Apple Music, Spotify, and YouTube. The song quickly became viral, accumulating over 20 million views across all platforms[^58], prior to its removal by Universal Music Group, Drake’s recording label. In contrast, Canadian artist Grimes has extended an invitation to musicians to emulate her voice via AI for the creation of new musical pieces, stipulating that the lyrics should not be harmful. She has advocated for the democratization of art and the abolition of copyright[^59]. Additionally, Grimes has employed AI to design visual content for her LED backdrop at Coachella in 2024.
Finally, the rapid development of AI technologies has also raised concerns about job displacement and the balance between automation and human participation in creative processes. Ensuring that AI augments, rather than undermines, human effort poses a significant challenge for developers and policymakers.
4.3 The future of AI technologies
Several key technological issues remain which need to be addressed if AI is to deliver its full potential. These in particular relate to training data, computational complexity and their depth of reasoning or planning, and are discussed below.
A substantial amount of data is essential for training AI models in order to achieve high performance and good generalization. Major companies such as Google, Meta, and NVIDIA, with their respective models: BERT, Segment Anything, and Canvas, dominate this space, benefiting from leveraged resources to gather data and process it to train sophisticated models. However, in November 2024, Bloomberg reported that OpenAI, Anthropic, and Google are all experiencing relatively slow growth in the performance of their AI models, with one of the key challenges being training data[^60].
LLMs excel in applications involving complex tasks, advanced reasoning, data analysis, and understanding context. However, these models typically require high computational resources or cloud computing for development, operation and fine-tuning. A new trend emerging alongside LLMs is the development of Small language models (SLMs), such as Phi-3 by Microsoft[^61]. SLMs offer promising solutions for regulated industries and sectors encountering scenarios where high-quality results are essential while keeping data ’on-site’. Their potential is particularly relevant when deploying more capable SLMs on smartphones and other mobile devices, allowing them to operate ‘at the edge’ without relying on cloud connectivity. Recent highly successful platforms, such as DeepSeek-V3 DeepSeek-AI et al. (2024) and Qwen2.5-Max Team (2024), are based on Mixture-of-Experts (MoE) models, which tackle complex problems by dividing them into simpler sub-tasks, each handled by a specialized “expert.”
Despite evident advancements in AI, current models still struggle with tasks requiring planning or deep reasoning and are prone to errors when encountering unexpected data. This, in turn, reduces the confidence of users and trust in the results. AI algorithms can learn through reinforcement learning, but this process often identifies the best outcome as an anomaly rather than the norm. Yann LeCun, Professor at NYU and Chief AI Scientist at Meta, noted that while LLMs show a degree of comprehension in processing and generating text, their understanding lacks depth, often leading to results that defy common sense[^62]. He advocates for self-supervised learning as a pivotal future direction for AI, emphasizing its potential to derive insights from unlabeled data. Concurrently, Andrew Ng, Adjunct Professor at Stanford University and Founder of DeepLearning.AI, sees iterative AI agentic workflows[^63] as a key advancement for enhancing AI tool capabilities through an interactive approach by AI agents. These workflows involve autonomous agents that interactively learn from experience, understand natural language, and execute tasks on behalf of users.
The increasing openness of code and datasets is seen by many as a catalyst for accelerating AI advancements, with major firms like Microsoft, Google, and Meta supporting open access technologies. However, this openness also introduces security risks, necessitating new regulatory measures to monitor models post-release, to standardize documentation, and to assess the safety of software code and training data disclosure.
Finally, as stated in Jeary and Gajjar (2024), the rapid advancement of AI technologies has revolutionized cultural experiences, often referred to as ‘CreaTech’—the convergence of the creative and digital sectors Council (2021). Such innovations not only reshape how people engage with art and creative work (e.g., through AR/VR/MR) but also drive the evolution of the technologies themselves.
Research funding
This work has been funded by the UKRI MyWorld Strength in Places Programme (SIPF00006/1).
Data Availability
No datasets were generated or analysed during the current study.
Author Contributions Statement
N.A. wrote the main manuscript text in section 1, 2, 3.1-3.5,3.8, 4, prepared all figures. F.Z. wrote the main manuscript text in section 3.6-3.7. D.B. wrote the main manuscript text in section 4. All authors reviewed the manuscript.