Supporting Serious Game Development with Generative Artificial Intelligence: Mapping Solutions to Lifecycle Stages

Abstract

The emergence of effective Generative AI has sparked a revolution in video game development, enabling us to generate game assets and source code at a fraction of the cost and time needed compared to if human developers were involved. But the support available from GenAI goes far beyond the generation of game assets and code, especially in the case of serious games, which have to combine playability with non-entertainment purposes such as, but not only, education. In this paper, the potential forms of GenAI-based support for serious game development are explored and placed into the context of the respective phases of the serious game development lifecycle. As existing lifecycle models are either not specialized for the specifics of serious games or are otherwise too simple, a new serious game development lifecycle model has been proposed for this purpose.

1. Introduction

The oxymoronic term of serious games denotes “games that do not have entertainment, enjoyment or fun as their primary purpose” [1]. While some authors associate them solely with education [2], in reality, they may serve also other purposes [3], such as promoting a meaningful message to the player (and, if possible, promote change with that message; see, e.g., [4]) or creating direct real world outcomes (e.g., solving open research problems [5]).

Although serious games can and have been implemented as non-digital, e.g., board games, they are currently mostly implemented as video games [6]. Although we do not follow the notion of limiting serious games merely to the digital domain, as some authors do (see, e.g., [7]), nonetheless, in this paper, we will only deal with digital serious games. While such games share key development process elements and techniques with non-serious, i.e., entertainment, video games, their serving of purposes different to entertainment causes a number of differences. The primary difference lies in the need to combine a compelling gameplay experience for players (a requirement for every game) with the utilitarian function for which the serious game has been developed (a requirement typical for enterprise information systems rather than video games) [8]. No less important is the need for assessment or evaluation both in and of serious games: the former internally measuring the player’s performance for the sake of providing appropriate feedback to them, but also enabling adaptivity and personalization to meet individual needs with regard to, e.g., learning styles or information provision rates, and the latter verifying the effectiveness of a serious game in attaining its intended goals [9].

Generative Artificial Intelligence, or GenAI, represents “computational techniques that are capable of generating seemingly new, meaningful content such as text, images, or audio from training data” [10]. Prominent examples of GenAI-based solutions include OpenAI GPT (for text generation) [11], OpenAI DALL-E 2 (for image generation) [12], SUNO (for music generation) [13], and GitHub Copilot (for software code generation) [14].

The introduction of GenAI to game development is truly revolutionary, as it not only can save time and money, but is also capable of delivering high-quality assets [15]. The particularly important advantages are that creative power becomes available to people lacking artistic skills, that GenAI tools can create a large number of variations in an iterative way (without getting frustrated, as a human artist would, after receiving a request for changes multiple times), and that, once the model is trained, the results are obtainable almost instantaneously [15].

This potential has been noticed by the video game industry, and, by mid-2025, the usage of GenAI during production or gameplay has been disclosed for 7818 games published on Steam, which corresponds to 7% of the whole Steam catalogue [16]. The most frequently disclosed use case for game development (indicated in about 60% of disclosures) was visual asset generation (characters, backgrounds, 3D models, textures, etc.), followed by audio (e.g., background music, character voices, voice-overs), text and narrative (e.g., item descriptions or storyline), marketing and promotional materials, and game code/logic [16]. An increasing number of games (e.g., AI Roguelite, inZOI, Never Ending Dungeon, DREAMIO, AIdventure) use GenAI at runtime, e.g., to generate audio-visual assets, game levels, and non-playing characters’ dialogue responses and behaviours [16].

Although the adoption of GenAI by the game industry is hampered by several barriers, including intellectual property and copyright issues (indicated by 59% of respondents surveyed by Boston Consulting Group), the low reliability of GenAI outputs (50%), considering current GenAI technology as not ready or suitable for adoption (43%), cybersecurity and data privacy issues (33%), and the lack of employees’ capabilities and practical knowledge (28%) [17], the expected benefits make its further progress inevitable. For this reason, it is worthy to investigate the multitude of ways in which GenAI can support serious game development, considering not only the actual development phase, but the entire lifecycle of the game. And because, as explained earlier, the development of serious games differs to that of video games developed purely for entertainment, the scope of possible support also differs, which necessitates research that is explicitly focused on serious games rather than video games in general. This paper strives to address this gap by achieving the following goals:

• To scheme a model of the serious game development lifecycle that indicate its key phases of distinct character.
• To identify the literature sources, not necessarily devoted to the use of GenAI in the domain of serious games, but which nonetheless describe GenAI applications that could be used in the serious game development lifecycle.
• To assign identified GenAI applications to the respective phases of the proposed serious game development lifecycle model.

The first of these goals has been attained by performing a critical comparison of the classic system development lifecycle (SDLC) [18], video game development lifecycle (GDLC) [19], and a prior proposal of a lifecycle model developed explicitly for serious games [20], followed by the specification of the new model addressing their deficiencies with regard to the adequate coverage of serious games lifecycle phases (see Section 2).

The remaining two goals have been attained by performing a narrative review of the relevant literature in the context of the proposed lifecycle model. We chose to perform a narrative review rather than a systematic one because its purpose was not to map the existing literature, but rather to map the currently possible forms of support (backed by the existing literature) to serious game development lifecycle stages. By doing so, we both avoided a time-intensive analysis of all of the relevant literature (instead, we reported only a few sufficient examples for the respective types of GenAI applications) and were not bound by the search strategy defined in the systematic review protocol (instead, we devised additional narrowly focused queries to find papers describing GenAI applications that were known to exist, but were not identified by prior queries of wider scope). This choice came at the price of a higher risk of selection bias, characteristic of narrative reviews in comparison to systematic ones (see [21], p. 2, and the works cited therein), which is the key limitation of our study.

The identification of relevant papers was performed in August 2025 and consisted of three stages. First, using Scopus and Google Scholar, we identified the existing surveys of GenAI applications (not only to serious game development, but also to related areas) and then analysed them. The results of this stage are reported in Section 3. Next, we performed a number of bibliographic searches, focussing on general and specific forms of possible GenAI support (suggested by the results of stage one). We chose Google Scholar for this task, considering the novelty of the area, with a substantial portion of the relevant works only being published as preprints, which are not searchable in Scopus. Eventually, we performed backward and forward snowballing on the set of the most promising papers among those retrieved up to that point. For the sake of replicability and to propel future research in this area, Appendix A lists the search phrases used (

Table A1. List of query strings.

Query String	Targeted Lifecycle Phase
AI SDLC	All
AI Software Development Lifecycle	All
Generative Artificial Intelligence software	All
Generative AI software	All
GenAI gaming	All
GenAI Software Development	All
Generative AI creative ideas	Ideation
GenAI brainstorming software	Ideation
GenAI ideas	Ideation
GenAI software requirements	Design
Automatic 3d video games	Design, Development
Synthetic user experience GenAI	Design, Testing
GenAI narrative	Design, Using
Automatic code generation	Development
GenAI dialogues	Development, Using
GenAI game dialogs	Development, Using
GenAI Generating test cases	Testing
Serious game evaluation GenAI	Evaluation

) and the full list of identified papers, with the relevant phase of serious game development lifecycle indicated (

Table A2. List of works in the analysed dataset.

Ref.	Title	Phase
[37]	The Role of Generative AI in Software Development Productivity: A Pilot Case Study	Ideation
[39]	Supermind Ideator: Exploring generative AI to support creative problem-solving	Ideation
[40]	The Ink Splotch Effect: ACase Study on ChatGPT as a Co-Creative Game Designer	Ideation
[38]	LLMs in Design Thinking: Autoethnographic Insights and Design Implications	Ideation
[44]	Using ChatGPT in Software Requirements Engineering: A Comprehensive Review	Software Design
[45]	MAPE-K Loop-Based Goal Model Generation Using Generative AI	Software Design
[52]	GeneyMAP: Exploring the Potential of GenAI to Facilitate Mapping User Journeys for UX Design	Software Design
[54]	Generative AI for Front-End Development—Automating Design and Code with GPT-4 and Beyond	Software Design
[51]	Game Plot Design with an LLM-powered Assistant: An Empirical Study with Game Designers	Content Design
[42]	How ChatGPT can inspire and improve serious board game design	Content Design
[43]	ChatGPT and Other Large Language Models as Evolutionary Engines for Online Interactive Collaborative Game Design	Content Design
[53]	SuSketch: Surrogate Models of Gameplay as a Design Assistant	Content Design
[41]	A Preliminary Study on a Conceptual Game Feature Generation and Recommendation System	Content Design
[66]	LLMaker: A Game Level Design Interface Using (Only) Natural Language	Content Design
[50]	GPT2SP: A Transformer-Based Agile Story Point Estimation Approach	Design
[49]	Evaluation of Context-Aware Language Models and Experts for Effort Estimation of Software Maintenance Issues	Design
[46]	Generative Artificial Intelligence for Software Engineering—A Research Agenda	Design, Content Prototyping
[77]	The AI-Native Software Development Lifecycle: A Theoretical and Practical New Methodology	Design, Development, Testing
[69]	Doodle2App: Native App Code by Freehand UI Sketching	Software Prototyping
[70]	Automatic code generation from sketches of mobile applications in end-user development using Deep Learning	Software Prototyping
[67]	The Impact of AI on Developer Productivity: Evidence from GitHub Copilot	Software Prototyping
[68]	SKCODER: A Sketch-based Approach for Automatic Code Generation	Software Prototyping
[59]	MONstEr: A Deep Learning-Based System for the Automatic Generation of Gaming Assets	Content Prototyping
[58]	Generative AI for Architectural Design: A Literature Review	Content Prototyping
[105]	MarioGPT: Open-Ended Text2Level Generation through Large Language Models	Content Prototyping
[61]	SketchMetaFace: A Learning-based Sketching Interface for High-fidelity 3D Character Face Modeling	Content Prototyping
[62]	DreamFace: Progressive Generation of Animatable 3D Faces under Text Guidance	Content Prototyping
[60]	Text2AC: A Framework for Game-Ready 2D Agent Character(AC) Generation from Natural Language	Content Prototyping
[71]	Word2World: Generating Stories and Worlds through Large Language Models	Prototyping
[30]	Play my Thesis	Content Prototyping, Content Development
[63]	Adaptive Worlds: Generative AI in Game Design and Future of Gaming, and Interactive Media	Content Prototyping, Using
[74]	SimUser: Generating Usability Feedback by Simulating Various Users Interacting with Mobile Applications	User Feedback
[75]	SimTube: Simulating Audience Feedback on Videos using Generative AI and User Personas	User Feedback
[79]	Generative AI for Test Driven Development: Preliminary Results	Software Development
[80]	Large Language Models for Software Engineering: Survey and Open Problems	Software Development
[76]	Towards Immersive Computational Storytelling:Card-Framework for Enhanced Persona-Driven Dialogues	Content Development
[56]	Leveraging Large Language Models to Support Authoring Gamified Programming Exercises	Content Development
[55]	Mystery Game Script Compose Based on a Large Language Model	Content Development
[65]	Level Generation Through Large Language Models	Content Development
[106]	ChoreoMaster: choreography-oriented music-driven dance synthesis	Content Development
[81]	Automated Test Case Generation for Software Testing Using Generative AI	Testing
[82]	Retrieval-Augmented Simulacra: Generative Agents for Up-to-date and Knowledge-Adaptive Simulations	Testing
[84]	AutoTestGPT: A system for the automated generation of software test cases based on ChatGPT	Testing
[85]	No More Manual Tests? Evaluating and Improving ChatGPT for Unit Test Generation	Testing
[86]	Using Large Language Models to Generate JUnit Tests: An Empirical Study	Testing
[87]	Large Language Models are Few-shot Testers: Exploring LLM-based General Bug Reproduction	Testing
[83]	Applying Large Language Model to User Experience Testing	Testing
[88]	Enhancing Player Experience Through Generative Artificial Intelligence: Custom Interaction in Game Design	Using
[89]	Learning Through the Dialogue with NPCs Using Generative AI	Using
[90]	Developing an immersive game-based learning platform with generative artificial intelligence and virtual reality technologies—“LearningverseVR”	Using
[93]	Using a Large Language Model to turn Explorations of Virtual 3D-Worlds into Interactive Narrative Experiences	Using
[95]	Generative AI Meets Adventure: Elevating Text-Based Games for Engaging Language Learning Experiences	Using
[94]	LIGS: Developing an LLM-infused Game System for Emergent Narrative	Using
[96]	Generative Agents: Interactive Simulacra of Human Behavior	Using
[92]	Chatter Generation through Language Models	Using
[98]	Playing With Unicorns: AI Dungeon and Citizen NLP	Using
[99]	3D Building Generation in Minecraft via Large Language Models	Using
[97]	Assessing the proficiency of large language models in automatic feedback generation: An evaluation study	Using
[91]	Managing the Personality of NPCs with Your Interactions: A Game Design System Based on Large Language Models	Using
[29]	Machine Learning and Generative AI in Learning Analytics for Higher Education: A Systematic Review of Models, Trends, and Challenges	Using
[100]	Turning Real-Time Analytics into Adaptive Scaffolds for Self-Regulated Learning Using Generative Artificial Intelligence	Using
[101]	Multimodal large language models for inclusive collaboration learning tasks	Using
[100]	Turning Real-Time Analytics into Adaptive Scaffolds for Self-Regulated Learning Using Generative Artificial Intelligence	Evaluation
[104]	Evaluating and Optimizing Educational Content with Large Language Model Judgments	Evaluation
[103]	LLMs as Educational Analysts: Transforming Multimodal Data Traces into Actionable Reading Assessment Reports	Evaluation
[102]	Advanced large language models and visualization tools for data analytics learning	Evaluation

).

The mapping constituting the main result of our work is performed in Section 4, whose subsections correspond to the subsequent phases of the lifecycle. Section 5 strives to summarize the obtained results, whereas the final section concludes the paper.

2. Serious Game Development Lifecycle

As serious games implemented as digital games are a type of software, their lifecycle could be modelled after the classic system development lifecycle (SDLC), consisting of seven phases. The first of them is planning, when the project scope, goals, and requirements are identified and an initial project plan is created. Second comes the analysis phase, which involves gathering and reviewing data on project requirements to produce fully detailed requirement documentation. The design phase follows, comprising defining the project architecture and creating the Software Design Document (SDD). In the subsequent coding phase, developers write the initial code and build a functional software prototype. Then testing begins: the code is reviewed and bugs are eliminated to achieve a refined, optimized version of the software. Once testing is complete, in the deployment phase, the software is moved to the production environment and made available to end users. The final maintenance phase updates the code with continual fixes and improvements to ensure that it remains functional over time [18].

As mentioned earlier, game development differs from enterprise information systems development (for which SDLC was originally dedicated). Consequently, several definitions for game development lifecycle (GDLC) were proposed by various authors, later synthesized by Ramadan and Widyani into a model comprising six phases [19]:

• Initiation at which a rough concept of the game to be developed is created.
• Pre-production, which involves the creation and subsequent revisions of game design and the iterative development of game prototypes. This phase ends when the revised game design is approved.
• Production, consisting of the development of the source code, game assets, and the integration of both these elements.
• Testing, which is aimed at verifying the game usability and playability. It is an internal process involving game development team members.
• Beta testing, which involves external testers (either invited or self-recruited) in an attempt to discover bugs that have not been detected internally in the previous phase.
• Release, which denotes the time when the game has reached its final form and is released to the public. It includes planning for maintenance and future game extension.

Although the development of serious games can be considered in terms of both SDLC (see, e.g., [22]) and GDLC (see, e.g., [23]), a model not only selecting and combining the most relevant elements of these two but also adding elements they both miss (e.g., the evaluation whether a serious game achieves its intended utilitarian goals) would be more appropriate. And although there is a lifecycle model developed explicitly for serious games, it is very simplistic, as it identifies only three general phases: design (including learning design and game design), development (of the game and the analytics model), and evaluation (including metrics and interactions) [20]. Considering this, we have decided to propose an own serious game development lifecycle model, consisting of the following phases:

• Ideation, during which the initial concept is raised and basic design guidelines are set out. Usually a number of candidate concepts is considered from which one is selected, polished, and eventually approved for implementation.
• Design, consisting of content design and software design, in which documentation and materials are developed, but no viable prototype is obtained. Note that we do not formally distinguish game content and learning content, as even in a case where both are not the same, they can be developed using the same methods. Similarly, we do not distinguish game mechanics and game assets, as the former are either a kind of content like the latter (when a reusable game engine is used and game rules are specified as its input) or embedded in the software. The phase ends when a design decision is made to develop a prototype.
• Prototyping, consisting of content and software prototyping, in which the two components are combined into a playable form, more or less resembling, but not identical to, the future release version of the game. Once developed, a prototype is presented to human testers.
• User Feedback, in which human testers play the prototype and report their opinions. These are aggregated and result either in a request for changes and moving back to the design phase, or in user acceptance leading to the next phase.
• Development, consisting of content and software development, in which the game in version presumed to be final is developed.
• Testing, involving a closed group of human testers to which the alpha version of the game is presented. In contrast to the earlier prototypes, the alpha version should not contain obvious (that is, known to developers) bugs, scaffolding code, and missing contents. In contrast to the release version, the alpha testers may expect to reveal serious software bugs and gameplay glitches. The discovered bugs are reported, and the game moves back to the previous phase. If no important bugs persist, the game is deployed to the end users.
• Using, during which the game is played by the users. There still can be bugs discovered as late as this phase, and these can be reported to the development team for patches to be developed. Regardless, usage data and user assessments are collected.
• Evaluation, which is performed when a sufficient amount of data is collected. Its results not only verify the effectiveness of the game in attaining its serious purpose, but may also indicate its strong and weak points, and, as such, can spark the ideation of the next version of the game.

The proposed serious game development lifecycle model is visualized in Figure 1.

3. Related Work on GenAI Support

In order to establish the results of prior reviews on the potential use of GenAI in serious games, we have searched for literature reviews ([24,25,26,27,28,29]) and other types of publications containing relevant literature review sections ([30,31,32,33]). Note that we focused on GenAI and LLMs as its primary representative, avoiding older reviews on the use of AI (other than GenAI) in serious games, such as [34,35,36]. The results of this search are presented in

Table 1. Considered prior literature reviews.

Ref.	Focus	Application Area	Supported Stage	Type	Databases	Sources
[24]	GenAI	Game character creation	Ideation, Design, Prototyping, Development, Using	SLR	ACM DL, IEEE Xplore, SpringerLink, ScienceDirect, Scopus	72
[30]	GenAI	Serious games	Development	NR	(undisclosed)	5
[31]	GenAI	Video games	Development, Using	NR	(undisclosed)	6
[25]	GenAI	Video games	Ideation, Design, Prototyping, Development, Testing	ScR	ACM DL, IEEE Xplore, ProQuest, ScienceDirect, Scopus, Web of Science	61
[26]	GenAI	Video games	Ideation, Design, Prototyping, Development	QRS	ACM DL, Google Scholar, Scopus	10
[27]	LLMs	Any (Design)	Ideation, Design, Prototyping, User Feedback, Development	SLR	ACM DL, IEEE Xplore, Scopus, Web of Science	118
[28]	LLMs	Any (Ideation)	Ideation	SLR	ACM DL, Google Scholar, IEEE Xplore, Scopus, Web of Science	61
[32]	LLMs	Video games	Design, Using	NR	(undisclosed)	80
[29]	Combining ML and GenAI	Learning analytics systems for higher education	Using	SLR	ACM DL, Emerald, ERIC, IEEE Xplore, ProQuest, Sage, ScienceDirect, Scopus, Taylor & Francis, Web of Science, Wiley	101
[33]	Text-to-image generation	Visual game assets	Ideation, Design, Prototyping, Development	NR	Google	24

SLR = Systematic Literature Review; NR = Narrative Review; ScR = Scoping Review; QRS = Qualitative Research Synthesis.

.

As can be observed in Table 1, only one of the found ten works focused on serious games, and in a very narrow way, as it covered only five relevant primary sources [30]. There were four papers dealing with video games in general [25,26,31,32] and two more limited to specific elements of their development [24,33]. Two works covered specific phases of the creative process not restricted to (but including) video game development [27,28], whereas the last included paper dealt with educational software (which serious games could be an instance of) [29]. While six works covered any kind of GenAI [24,25,26,29,30,31], three considered only Large Language Models (LLMs) [27,28,32], and the scope of one was limited to image generation [33].

4. GenAI Support for Serious Game Development

4.1. Ideation

Generative Artificial Intelligence can support ideation processes. According to findings by Coutinho et al. [37], software professionals use AI in brainstorming sessions and design thinking. Experience-based insights into the impact of LLMs on idea generation are shared by Asadi [38]. An exemplary software that support users in thinking about their problems and generate potential solutions is “The Supermind Ideator” designed and developed by Rick et al. [39]. It applies techniques from “Supermind Design” methodology to generate ideas for consideration. The evaluation results show that, in general, participants responded positively to it.

The experiment performed by Anjum et al. has shown that GenAI does not yet provide a sufficient replacement for human game designers, but it is able to innovate and suggest interesting gameplay features; thus, it may have a place in the creative process of game ideation and design [40].

4.2. Design

GenAI can take an active part in serious game design. Charity et al. introduce a system recommending new game features using a model trained on descriptions of almost 60,000 games [41]. Apart from suggesting game themes and mechanisms aligned with curriculum and learning goals, LLMs can provide templates of game components and identify areas for improvement to enhance the game’s efficacy as an educational tool [42]. Lanzi and Loiacono propose a collaborative game design framework in which Large Language Models are used for the recombination and variation of ideas [43].

GenAI can be used to generate software requirements. According to the comprehensive review written by Marques, Silva, and Bernardino [44], ChatGPT has potential to transform software requirements engineering due to its ability to enhance brainstorming, reduce errors, cut costs, and improve efficiency, although challenges such as bias and hallucination remain significant concerns. Nakagawa and Honiden [45] propose a goal model generation process with the use of GenAI based on the MAPE-K loop mechanism. The results of two conducted case studies show that their process efficiently supports goal model constructions.

Design decisions can also be supported by the use of GenAI. Nguyen-Duc et al. [46], in their research agenda, show how other authors described multiple uses of GenAI to predict project cost. As demonstrated in prior research, traditional AI techniques are effective for cost and effort prediction in both agile [47] and non-agile projects [48]. Nonetheless, they are also outperformed by GenAI in this field, as indicated by the evaluation results of Alhamed et al. and Fu et al. involving BERT [49] and GPT-2 [50], respectively.

LLMs can support the design of game plot. A good example of that is “GamePlot”, a software powered by GPT-3.5-turbo assisting game designers in creating, testing, and refining narratives for turn-based games [51]. The tool includes features like real-time plot editing and designer control over NPC responses. Despite the problems with story simplicity and inconsistency, a study with 14 designers showed a high level of satisfaction with the generated plot.

LLMs may also be used to create user journey maps. One tool that integrates GenAI to improve the user journey mapping process is “GeneyMAP” by Mei et al. [52]. According to the feedback from 20 users involved in its evaluation, the tool managed to improve their efficiency, although participants also had concerns about losing control of design process and felt constant need to verify what GenAI had generated.

GenAI can be very useful in helping the designer with balancing the gameplay. For instance, SuSketch is a tool predicting balance between designer-specified character classes in a complete playthrough. It also proactively proposes alternatives to the level and class pairing to the designer that would improve the predicted balance of the game [53].

Another task that GenAI can effectively help with is the user interface layout. A comprehensive list of possible uses of GenAI in Front-End Development is provided by Alikhani [54]. The primary examples are the assistance in layout adjustment and design element suggestions. For example, popular design tools such as Figma and Adobe XD use AI features to assist in layout adjustments and design element suggestions.

LLMs can also help with recommending game design features. Charity et al. [41] introduced a system that suggests game design features from text. The designer-user inputs the basic idea for the game (e.g., multiplayer FPS with a list of game modes) and the system generates a list of features that would fit a games of such kind.

4.3. Prototyping

As LLMs are primarily text models, their most obvious use in content prototyping is for generating game script. According to Li et al., this not only allows for ordinary users to convert their ideas into a complete work with one click, but also helps professional scriptwriters to improve the efficiency of their creative work [55]. Importantly, LLMs can generate textual puzzle definitions along with the metadata needed by the game engine in a format it requires; the work by [56] describes this in the context of programming exercises embedded in a gamified learning platform, but the same approach could be applied to educational serious games featuring different types of puzzles.

GenAI can also support prototyping by the generation of gaming assets of various kinds. Colado et al. describe using several GenAI tools to generate game scenarios, characters, their poses, and their facial expressions [57].

GenAI is especially useful in generating visual game assets. Li et al. reviewed examples of imagery generated by different text-to-image and image-to-image models (the latter can generate high-quality images from sketches of limited quality) (see [58], Figure 9). Humble [30] reports using DALL-E to create 2D graphics for a visual novel game. He reports some limitations like AI working best only with certain artistic styles or having problems with the number of limbs (fingers, hands, legs, etc.). Brocchini et al. [59] present “MONstEr”, a system that automatically generates 3D assets for video games. Trained on a custom image dataset, it reduces the need for manual modeling with only minimal adjustments needed to integrate the object into the game scene.

GenAI capabilities extend from static to animated graphics. For instance, Sun et al. describe a framework for GenAI-based generation of animated 2D characters usable across various game engines from natural language text [60]. Moving the focus to the development of 3D avatars, high-fidelity 3D faces can be generated by amateur users in minutes with the help of GenAI as implemented in the sketching tool SketchMetaFace [61]. Realistic animated 3D faces can also be generated from video footage, even of fashion icons or cartoon characters, as demonstrated by the DreamFace tool [62].

GenAI can be exploited to generate 3D worlds prototypes from just sketches and text. As an illustration, GenAI-powered software “Cybever” can generate complex 3D worlds from simple sketches or text, thus helping to rapidly prototype the game worlds [63].

GenAI can also be used to create video game levels from text. A good example is provided by MarioGPT, a fine-tuned GPT-2 model that can be used to generate Super Mario Bros levels and also predict players’ interactions with the generated levels [64]. According to the results published by its authors, while this generation can also be guided with prompts, it occasionally fails to follow the provided instructions. In a similar vein, Todd et al. [65] used LLMs, more precisely GPT-2 and GPT-3, to generate new levels for the video game ’Sokoban’, whereas Gallotta et al. introduced LLMaker, a level generator for dungeon-crawler-type games only taking natural language instructions as input, allowing for the designer to express their intent in an intuitive way, but ensuring that the generated levels adhere to specified constraints [66].

Regarding software prototyping, the most obvious use of GenAI is code generation, so that a working prototype can be quickly obtained without much effort. While data exactly on supporting game code development were not found, there is a number of reports regarding other not-so-distant use cases. For instance, the development of a HTTP server in JavaScript (a potential component of web-based serious games) with the support of GitHub Copilot has been completed 55.8% faster than without using it [67]. While most of code generation research considers natural language text as the input, this is not the only possibility. For instance, Li et al. introduce SKCoder, a GenAI tool capable of generating code from sketches [68].

LMM can also be applied to generate a graphical user interface (GUI). A good example is Doodle2App, developed by Mohian and Csallner and based on a model pre-trained with Google Quick, Draw! dataset of 50M sketch strokes and retrained with sketch strokes collected from Amazon Mechanical Turk [69]. The app can convert sketches into single-page Android applications. On the 712 test samples, Doodle2App achieved 93.9% accuracy. A similar solution is Sketch2aia by Baulé et al., which generates the App Inventor code of the wireframe from hand-drawn sketches [70]. It achieves 87.72% accuracy for an average user interface component classification.

Word2World is a system that goes much further than those described above, as it combines LLMs’ strengths in storytelling and content placement to generate coherent and playable game levels from stories without fine-tuning, achieving a 90% success rate [71].

4.4. User Feedback

As shown by Abdelqader, LLMs can be useful in quantifying and analyzing game reviews written by customers of the Steam and Meta Quest stores, transforming unstructured text into structured feedback on key game design elements, monetization models, and platform-specific trends [72].

Tyni et al. make one more step and use LLMs for providing constructive feedback on game design (including using visual prototypes as input), thus reducing the need for human consultants. According to the results of their study, ChatGPT is able to provide feedback covering various aspects of serious game design, including, e.g., learning and educational content or game modes and mechanics similar in extent and quality to human consultants [73].

Another related idea is to use LLMs to generate usability feedback by simulating the interaction between users and applications. The users are modeled considering by their characteristics and contextual factors. Early results of one such system (SimUser) show that, depending on the user group and usability category, the feedback was from 35.7% to 100% similar to that obtained from human testers [74].

Similarly, SimTube is a GenAI-based tool that creates simulated audience comments ahead of a video’s release [75]. By combining video elements such as visuals, sound, and metadata with a range of virtual audience personas, it produces believable and diverse feedback. Its evaluation shows that the SimTube-generated comments are often more detailed and useful than actual viewer comments. Although this solution is designed for videos, it could also be applied in serious games to obtain feedback for any kind of in-game videos, such as, e.g., gameplay recordings, cutscenes, tutorials, or other video narrative elements shown during the game.

4.5. Development

Regarding development of game content, particularly visual game assets, the same GenAI-based tools can be used as in the prototyping phase. Nonetheless, there are types of game content which, in this phase, are needed in a much larger quantity (and also quality) than for prototyping.

GenAI can help with the development of dialogues. This potential has been explored by Humble, using his own game as a testing ground [30]. ChatGPT and Gemini were used to generate dialogues in the game. The author observes that GenAI-generated dialogues are most often not perfect, and suggests generating “good enough” dialogues to be later manually adjusted rather than to repeat the generation process until a fully satisfactory result is obtained. A problem with achieving dialogue accuracy in the context of many layers of provided information is also reported.

An even more positive example comes from Liao and Vargas [76], who introduced a card-based framework powered by Baichuan-7B to generate persona-driven dialogues for “Call of Cthulhu”, an RPG. They used character cards to include traits like how a character talks and their job, while scene cards were used to guide the conversation with setting details, important objects, character goals, and obstacles. Their findings show that the LLM was able to create context-aware dialogues, even for previously unseen characters and scenarios.

Another way that GenAI can help with the development of content is by generating game music. In the game by Humble mentioned earlier, GenAI (in particular, the BeatBot tool) was also used to generate music with a style that fitted the graphical design of the game. The author reports experiencing difficulties in getting the desired outcome and a problem consisting of the lack of style diversity [30].

Moving to the development of software, in a fully AI-native SDLC, as envisioned by Hymel, GenAI-based tools can be applied to every phase of software development, allowing us to shift the role of humans from primary implementers to primarily validators and verifiers [77].

More typically, however, GenAI is used for supporting rather than replacing programmers in their core job. This includes automatic code completion, code search support, automatic code translation between programming languages, and bug fixing (see [78] and the works cited therein).

GenAI can support specific software development methodologies. For example, Mock, Melegati, and Russo introduce Test Driven Development (TDD) automated with GPT 3.5 [79]. They conclude that GenAI can be effectively used in TDD, but the code generated by the AI must be supervised.

A number of studies (see [80] and the works cited therein) indicate that LLMs can also be used for source code optimization to improve performance and decrease memory consumption.

4.6. Testing

GenAI can be employed for automating test case generation in software testing, allowing for human staff to focus on higher-order problems. The findings of Bandi et al. reveal that ChatGPT is especially effective in this task, taking into consideration accuracy, adaptability, and handling of complex constructs, such as recursive algorithms and nested structures [81]. More examples of using GenAI for automated test case generation, along with other related uses of GenAI such as test adequacy evaluation and test output prediction, are presented in [80] and the works cited therein.

Shimadzu et al. show that LLMs can be used to simulate user interactions with the software. Although their case study regards social networking services, the same approach can be applied for game testing [82]. A similar solution developed by Hsueh et al. for testing web user interfaces was found, depending on test aspect, to be equal or to outperform human testers [83].

LLMs can be used to automatically generate software test cases. AutoTestGPT is a system powered by ChatGPT whose efficiency in test suite generation surpasses 70%, compared to the manually constructed testing suites [84]. In similar fashion, ChatTester utilizes ChatGPT to generate unit test cases [85]. Compared to directly using ChatGPT for this purpose, ChatTester produces 34.3% more compilable tests and 18.7% more accurate assertions. A valuable comparison is provided by Sidiqq et al. [86] who evaluated Codex, GPT-3.5-Turbo, and StarCoder on Java unit test generation using HumanEval and EvoSuite SF110. Their analysis covered compilation, correctness, coverage, and test smells. Codex achieved over 80% coverage on HumanEval, but all evaluated models performed poorly on EvoSuite SF110 with under 2% coverage. Another test-supporting functionality is featured in LIBRO by Kang, Yoon, and Yoo, a framework that uses LLMs to automatically generate bug-reproducing tests from bug reports [87]. On the Defects4J benchmark, LIBRO reproduced failures for 33% of 750 bugs, demonstrating its potential to support software testers and developers.

4.7. Using

GenAI can enhance player engagement by helping them to co-create within the game worlds with the help of procedural content generation [88].

GenAI can adjust the game based on player behaviour. This is demonstrated by Ratican and Hutson [63], who report that GameNGen can dynamically recreate the video game DOOM and adjust it to player’s actions like dodging bullets, interacting with enemies, and generating the next frame accordingly. Also, the previously cited work [88] analyses how GenAI-driven systems can adjust to personal preferences and emotional reactions to deliver a customized gaming experience.

GenAI enables to implement in-game interaction with non-player characters (NPCs) without scripted dialogues. As an example, Rodrigues and Silva [89] created a 3D character based on the Portuguese navigator Tristao Vaz Teixeira to let students learn history through textual communication or by directly talking with him. The evaluation showed that it met the expectations of the students, with the majority of them agreeing that AI can make history classes more fun and interactive. Another example is LearningverseVR, a game-based learning platform which uses GenAI to create NPCs with different personalities and backgrounds with whom learners can communicate without scripted dialogues [90].

Dai et al. went one step further and developed a system capable of creating NPCs not only with their own personalities, but which also produce a variety of reasonable feedback adaptively based on the player’s interactions. The system comprises an adaptive personality model manager and a drama mechanism manager based on retrieval-augmented generation [91].

Müller-Brockhausen, Barbero, and Preuss [92] examined the use of LLMs to generate chatter that stays in the context of a character’s description. According to their results, in 500 generated samples, 79% stayed within their respective context. The same authors also claim that around 70% of consumers have a gaming hardware with enough memory to store a small 7B 4-bit quantized language model on top of a demanding AAA game.

Yet another use for GenAI during gameplay is narrating the game based on the user’s actions. This is exemplified by “Inner Voice”, a prototype system developed by Rist [93]. It prompts the LLM to generate narration when the user interacts with invisible triggers. Inner Voice can also be adjusted to narrate in a neutral or sarcastic way.

The task of GenAI in the game can be more ambitious: for instance, it can take care of emergent narratives based on actual gameplay experiences and progression, dynamically generating content tailored to evolving gameplay contexts [94]. This is directly applicable to serious games, as illustrated by the work by DaCosta, who applied ChatGPT to enhance a text-based game for language learning with contextually appropriate responses, making the game more interactive and engaging [95].

GenAI may also simulate human behavior of NPCs. This is demonstrated by Park et al. [96], who introduced generative agents in environment modelled after The Sims game. Agents simulated human behavior (e.g., cooking, working, initiate conversation). Interestingly, they also produced believable social behaviours such as organizing and attending a Valentine’s Day party, beginning with a user idea that one agent hoped to host a Valentine’s Day party. The agents spread invitations, built new connections, and arrived at the party together.

GenAI can help with providing real-time learning feedback. According to the results of the study by Dai et al., both GPT-3.5 and GPT-4 were able to generate more readable feedback with greater consistency than human instructors [97]. The previously mentioned “LearningverseVR” platform uses GenAI to generate badges for correct answers [90]. Correct answers also earn points which are added to a generated leaderboard. These elements provide real-time feedback and serve as a reference for teacher assessment.

GenAI can also act as a dungeon master. For example, “AI dungeon 2” is a text adventure game that uses GPT-2 to generate and narrate adventures [98]. One functionality worth noting is the possibility to use the “alter” command and edit the textual output to guide the narrative in the chosen direction.

GenAI can refine users’ prompts and generate buildings in world-building game environments. For instance, Hu et al. [99] used LLMs to generate 3D buildings in Minecraft from text. The model was able to successfully generate complete and satisfactory buildings. A highly interesting part of that work is the use of LLMs to refine user prompts to give more specific information and enhance the quality of the outcomes.

Moving to the instructor’s side, hybrid learning analytics systems that combine ML and GenAI can be used to enable personalized feedback and early dropout risk detection (see [29] and the works cited therein). Large Language Models (LLMs) also facilitate real-time reflective scaffolding [100] and the detection of inclusive interactions [101].

4.8. Evaluation

LLMs can make the processing and analysing of gathered data easier by automating tasks such as data cleaning or analytics code generation [102]. They can be especially helpful for analysing multimodal educational data (including diverse sources, such as eye-tracking, learning outcomes, assessment content) into actionable reading assessment reports in a form easily comprehensible to educators [103].

The results of the experiment by Li et al. show that the integration of analytics of self-regulated learning (SRL) with GenAI can turn insights from real-time analytics about students’ SRL processes and conditions into adaptive scaffolds, and promotes more metacognitive learning patterns [100].

Instead of supporting the evaluation process involving humans, GenAI can substitute it with an automatized one. LLMs have been demonstrated to be capable of evaluating the overall effect of instructional materials on the learning outcomes of different student groups, replicating well-established educational findings such as the Expertise Reversal Effect and the Variability Effect [104].

5. Discussion

As the number of potential points of support is astonishing, it is worthy to summarize the key findings here.

Table 2. Activities of serious game development lifecycle supported by GenAI.

Support Activity	Phase(s)	Sources
Support during the brainstorming sessions	Ideation	[37]
Suggesting new ideas	Ideation	[39]
Co-creating game design	Ideation	[40]
Supporting the design thinking process	Ideation	[38]
Generating software requirements	Software Design	[44,45]
Generating user journey maps	Software Design	[52]
Predicting potential bottlenecks and resource constraints	Software Design	[77]
Suggesting design improvements to enhance user experience	Software Design	[77]
Assisting in layout adjustment and design element suggestions	Software Design	[54]
Generating game narrative	Content Design	[51]
Designing boards for serious games	Content Design	[42]
Selecting, combining, and mutating the most promising designs	Content Design	[43]
Assisting in the level design process	Content Design	[53]
Recommending game features	Content Design	[41]
Level designing through Natural Language	Content Design	[66]
Predicting project cost	Design	[46]
Agile story-point estimation	Design	[50]
Effort estimation	Design	[49]
Automating entire software development lifecycle	Design, Development, Testing	[77]
Generating UI from freehand sketches	Software Prototyping	[69,70]
Code generation from prompt	Software Prototyping	[67]
Automatic code generation based on similar snippets	Software Prototyping	[68]
Generating 3D world models from sketches and text	Content Prototyping	[63]
Automatic generation of gaming assets	Content Prototyping	[59]
Generating 2D graphics from text	Content Prototyping	[30]
Generating images from sketches	Content Prototyping	[58]
Generating video game levels from text	Content Prototyping	[105]
Generating 3D models from sketches	Content Prototyping	[61]
Generating 3D models of faces	Content Prototyping	[62]
Generating 2D characters from text	Content Prototyping	[60]
Generate game levels from stories	Prototyping	[71]
Simulating audience feedback on videos	User Feedback	[75]
Generating usability feedback	User Feedback	[74]
Automating Test Driven Development (TDD)	Software Development	[79]
Software optimization for performance improvement	Software Development	[80]
AI-assisted code completion	Software Development	[80]
Generating dialogs for NPCs	Content Development	[30]
Generating in-game music	Content Development	[30]
Generating dialogues	Content Development	[76]
Generating puzzles	Content Development	[56]
Generating game script	Content Development	[55]
Generating game levels	Content Development	[65]
Generating dancing models	Content Development	[106]
Generating test cases	Testing	[81,84]
Simulating user interactions	Testing	[82]
Generating unit tests	Testing	[85,86]
Bug reproduction	Testing	[87]
UX Testing	Testing	[83]
Procedural content generation	Using	[88]
Adjusting the game difficulty in real-time	Using	[63]
Real-time GenAI NPCs without scripted dialogs	Using	[89,90]
Adaptive gameplay mechanics	Using	[88]
Dynamically generating badges	Using	[90]
Generating leaderboards	Using	[90]
Generating video game narrator	Using	[93]
Providing contextual hints to reduce frustration	Using	[95]
Generating emergent game narratives	Using	[94]
Simulating human behavior in a Sims-like environment	Using	[96]
Generating chatter	Using	[92]
GenAI playing the role of dungeon master	Using	[98]
Refining prompt for another LLM	Using	[99]
Generating in-game 3D buildings	Using	[99]
Generating feedback in real time	Using	[97]
Managing the personality of NPCs based on player interactions	Using	[91]
Supporting real-time learning analytics	Using	[29]
Supporting real-time reflective scaffolding	Using	[100]
Supporting the detection of inclusive interactions	Using	[101]
Supporting self-regulated learning	Evaluation	[100]
Evaluation of educational content	Evaluation	[104]
Evaluation based on multimodal data	Evaluation	[103]
Cleaning data for evaluation	Evaluation	[102]

lists the diverse forms of support from the use of GenAI mentioned in Section 4, indicating the respective phases of serious games development lifecycle as well as the relevant references.

Considering the presented potential for the use of GenAI in serious game development, the following general statements can be made:

• GenAI can effectively support all phases of serious game development lifecycle, including non-obvious ones, such as user feedback and evaluation.
• GenAI can effectively support the development of both serious game content (including educational content) and source code.
• While most of the forms of support are same for serious and non-serious games (e.g., the generation of code and visual assets), others are shared by serious games with educational software (e.g., the support of learning analytics or automatic learners’ assessment and feedback).
• Although automatic content generation is most beneficial at the prototyping phase, allowing us to quickly see how the proposed concepts could be implemented and to compare implementations of multiple alternative design choices, it also makes the development phase much more cost-effective by automating tedious and mundane tasks that have to be performed in large quantities (e.g., mass-generation of artefacts each having their own distinct characteristics, converting 2D graphics to 3D models, generating animation frames for static graphic objects).
• While some types of content are not needed before the development phase (e.g., background music is not necessary in an early game prototype-unless the game is of the rhythm genre [107], for which the music is essential), some (e.g., graphics) can be useful also in other phases, e.g., ideation (for effective sharing of visual concepts) or using (for illustrating marketing or informational materials).
• Due to deficiencies in generating large collections of code by the currently available models [108], automatic code generation is primarily useful during prototyping when quickly obtaining a playable game prototype (even if full of bugs) is worth it. For the development stage, where quality matters most, the ways of supporting human programmers (such as code search or automatic code completion) are more adequate. In testing, the currently available LLMs already rival human testers [81].
• Although the GenAI potential of generating various types of game assets on the fly during the using phase is impressive, in reality, it is hampered by the relatively high cost of accessing GenAI API providers [58] or the purchase and upkeep of hardware infrastructure [109], especially considering that serious games are often provided by educational institutions to their students free of charge.
• Another aspect worth consideration is the environmental cost of using GenAI. For instance, according to the results of the study by Cheung et al., LLM-assisted code generation effects, on average, a 32.72-times higher carbon footprint than manual code [110]. Also, the code generated by GenAI is not energy-efficient: considering the results obtained from analysing LLM-generated solutions of 878 coding problems, on average, the most energy-efficient LLM (DeepSeek-v3) produced code which consumed 16.7% more energy than human-written canonical solutions, whereas the code generated by GPT-4 Turbo consumed more than twice the energy used by canonical solutions [111].
• While GenAI decreases the pressure of having skilled graphic artists, level designers, or puzzle creators in the development team, mastering new skills (such as prompt engineering) and acquiring knowledge of new GenAI tools (best suited for specific types of tasks) will become a necessity.

The enthusiasm found in the literature regarding the introduction of new forms of application of GenAI is somewhat mitigated by various concerns. We briefly mentioned this topic in the Introduction, quoting the results of the Boston Consulting Group’s survey [17]; however, a more comprehensive list of concerns has been identified by Alharthi (see [112] and the works cited therein):

• The GenAI output’s impact on the integrity of human-authored content;
• The potential erosion of originality and the risk of over-reliance on AI-generated assets;
• The doubtful provenance of AI-generated content, particularly when models are trained on copyrighted or unattributed material;
• The ambiguity surrounding the ownership of AI-generated content (irrelevant for in-house phases such as ideation and prototyping, but crucial for the development of the final product);
• The risk of replacing human labour, potentially hurting job security, fair compensation, and the long-term sustainability of creative professions.

The above list indicates the initial directions for future works: developing policies and industry-wide good practices that would effectively and ethically address all of the presented concerns, at least in the context of serious game development (or game development in general).

A second, highly intriguing direction for future works is investigating the use of LLMs in the comprehensive evaluation of the attainment of serious game goals. For instance, an educational serious game could be verified by testing the final level of knowledge of a model (possibly connected to a database using Retrieval Augmented Generation) which initially possessed no domain-specific knowledge and learned it only via human-user-simulating interactions with the game. Another similar vein of research lies in employing LLMs to assess the balance of the playability and serious aspects (e.g., educational impact) of a serious game.

6. Conclusions

This paper contributes to the pool of research on the application of GenAI to the development of serious games. The presented wide scope of potential GenAI applications proves that GenAI is not confined to isolated stages of the serious game development lifecycle (such as content prototyping), but rather can be integrated across the whole lifecycle.

This paper’s novelty is threefold. First, it identifies potential points of GenAI-based support without limiting its scope to known examples of applications to serious game development, but also considers applications to (non-serious) game development and other types of software (mainly educational software).

Secondly, in order to put the identified forms of support in a practically useful context, they are matched to phases of the serious game development lifecycle. Thus, the presented findings are directly applicable in practice by serious game designers and developers, as well as educators defining the requirements for serious games and evaluating them.

Thirdly, as existing models are either not specialized for the specifics of serious games or are otherwise too simple, a new serious game development lifecycle model has been proposed for this purpose. As the proposed model is conceptual, its correspondence with the actual development lifecycles of existing serious games poses an interesting question to be empirically validated in future research.