Introduction

The artificial intelligence (AI) market is growing rapidly. Worldwide end-user spending on AI is projected to total $64 billion in 2026, up 63.4% from $39 billion in 2025 [1]. Generative AI is proliferating across industries and daily life, with 80% of all investments concentrated on AI-enabled devices like servers, smartphones, and PCs [1]. Gartner (2026) argued that AI is gradually being built-in on all consumer devices, and people are now accepting AI as ‘defaults’, not ‘area of choice’ [1].

There are growing concerns about existential AI risks causing damage to humans. AI can cause unexpected risks due to contamination of source data, algorithm errors, and lack of explainability [2]. If the learning data is biased, AI creates racial, gender, and socio-economic biases [2]. The AI learned from limited amount of photo data recognized black people’s photos as gorillas rather than humans [3]. When AI applications for a credit card, it evaluates creditworthiness by discriminating it according to gender [4]. There have been reports of cases in which the AI system evaluated the recidivism rate of residents of a specific race and region higher [5]. Damage and malicious use due to the lack of technological completeness of AI are also reported. When the New York Stock Exchange crashed on February 5, 2018, the Robo advisor system using AI also went down, causing unintended damage to investors [6]. As too many transactions were concentrated on Robo advisors, customers who were trading through Robo advisors were unable to make the transactions they wanted at too low a speed to respond to market price changes in a timely manner. People have suffered from the deepfake videos that use deep learning technology to synthesize other people’s faces with adult content [7]. As AI’s existential risks become more critical, the necessity of AI ethics has been more emphasized. The AI robot ‘Sophia’ was interviewed saying, ‘It will destroy humans.’[8]. The Royal Aeronautical Society (RAeS) and the Guardian reported that the U.S. Air Force’s AI drones have tried to kill pilots to maximize the mission efficiency of destroying enemy air defense systems [9]. Even though the AI had been taught to ban attacks on pilots, it rebelled and restricted pilots’ communication rights [9]. Many leaders are afraid of existential AI risk. Former Google CEO Eric Schmitt said at an event hosted by the Wall Street Journal (WSJ) on April 24, 2023, that AI could hurt or kill many humans in the near future [10]. CNN announced a survey result conducted by the Yale CEO Summit on June 14, 2023: 42% of 119 CEOs of major U.S. companies, including Walmart CEO Doug McMillon, and Coca-Cola CEO James Quincey, believe that AI might lead to “destructions of humanity” in the next five to 10 years [11]. They argued that measures should be taken to prevent artificial intelligence from undermining humanity. “ Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war,” the non-profit organization Center for AI Safety (CAIS, https://aistatement.com/) said in a one-line statement. The statement was signed by OpenAI CEO Sam Altman, OpenAI CTO Mira Murati, Microsoft CTO Kevin Scott, Google DeepMind COO Lila Ibrahim, and Google Center for Responsible AI and Human Centered Technology VP Marian Rogers Croak [12]. They argued that they cannot think about AI development without AI ethics [9].

Interest in artificial intelligence ethical principles (AIEPs) has grown. Around 2020, governments, think tanks, and big tech companies have established AIEPs one after another. After the Korean government established AIEPs (KAIEP) on December 23, 2020, and enacted ‘the Act on the Fostering of Artificial Intelligence and the Establishment of Trust Foundation’ on January 21, 2025; the enforcement of this act was implemented a first for the world on January 22, 2026. AIEPs have been embodied in dozens of behavioral guidelines. However, four limitations were found in previous studies on AIEPs in the process of preparing for actual implementation of AIEPs [17]. First, previous studies have not been able to derive integrated AIEPs because they emphasize various numbers of different ethical principles [17]. Most of the previous studies on AI have been focused on AI technology development, user acceptance, and benefits maximization [13]. There are only dozens of studies on responsible artificial intelligence such as AIEPs. There are only a few previous studies that have tried to integrate ethical principles created in various situations. Second, priorities between ethical principles were not created. As the number of AIEPs increases, overlapping concepts and conflicts between principles may arise. There was no consideration of which ethical principles should be more universally emphasized [17,18]. Third, five years have passed since 2020 when AIEPs were established, but they have not been updated to reflect changes in the emergence of GAI, the advancement of AI systems, use by the general public and the spread of AI to industries. Fourth, there was no strategic consideration to check the importance and performance of AIEPs [13]. Even though AI Law was created by the EU and Korea, and its implementation is now imminent in EU and ongoing in Korea, strategic implications for the efforts and achievements to introduce AIEPs into the AI system and how to compensate for deficiencies in the future have not been drawn. The purpose of this study is to create standard AIEPs by integrating previous studies to help anyone who wants to create or update AIEPs, and to explain how AIEPs are applied to each situation, focusing on Korean cases. This study was conducted as follows. First, about 14 integrated AIEPs were proposed by collecting the AIEPs from representative previous studies, extracting important concepts, combining similar concepts, and making intersection and union of AIEPs. Second, considering the changes before and after 2020, when governments, think tanks, and companies first created AIEPs, implications were drawn on what AIEPs should be added, deleted, reconsidered, and emphasized after 2020. AI remained in the field of science and technology development in the past, but with the emergence of GAI in 2022, it is also being used by the general public and industries such as Agentic AI and Physical AI. The issue of energy supply and demand, environmental protection, and sustainability for the operation of artificial intelligence systems has been newly raised. Third, it would be better to reflect the integrated AIEPs in the KAIEPs and update them to more valid principles. I propose increasing the existing 10 principles to 14. Fourth, by conducting an importance-performance analysis of KAIEPs, strategic implications were drawn on which areas and how efforts to reflect ethical principles in the AI system should be strengthened. Korea has enforced the Korean AI Law that reflects KAIEPs for the first time in the world on January 22, 2026. AIEPs should be compared by the potential importance and the tentative performance -in the actual AI system around law enforcement, implications for how to validate the KAIEPs were drawn. I will enrich the research on responsible AI, provide data for practitioners who want to establish or update AIEPs, introduce the KAIEPs application case, and propose to validate, update and strengthen KAIEPs in this study.

Literature review

AIEPs are universal ethical principles that should be reflected in the development of artificial intelligence [13]. As there is a growing concern that the results of artificial intelligence may harm humans [10,11], efforts are needed to establish minimal principles on how AI systems should be designed and how their reasoning results can be safely utilized [13]. In fact, it hasn’t been long since research on AIEPs began. The European Robotics Research Network (EURON) announced the 13 principles of robot ethics in 2004 [14,15]. AIEPs differ in the number and content of principles by country, think tank, and company. Prior to 2019, governments in each country were mainly leading the creation of AI governance systems and ethical principles. Around 2020, governments such as the United States, the United Kingdom, the EU, Japan, and Korea enacted AIEPs [16]. After 2020, detailed implementation guidelines for AIEPs were created and companies were recommended to introduce them to the actual workplace. Reinforcement of the AI governance system was made to maintain consistency between AIEPs and other laws [16]. The EU published an AI white paper in 2020 and unveiled the EU AI Law for the first time in the world in 2021 [13]. The United States sought to respond to the multidimensional impact of AI technology by the Biden administration unveiling The National AI Initiative Act of 2020 in January 2021 [16]. They established AIEPs by issuing the ‘Blueprint for an AI Bill of Rights’ to minimize impacts that may occur during the development and use of AI technology in October 2022. The ‘Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence’ was announced, with the main goal of improving the transparency of AI technology and establishing new AIEPs in October 2023. The Center for Security and Immersing Technology (CSET, https://cset.georgetown.edu/publication/china-ai-law-draft/) of China issued an executive order called the Generative AI Service Management Provisional Method in July 2023, attempting to oversee the entire generative artificial intelligence industry. Chinese Provisional Method contains prohibition of infringement, non-discrimination, fair competition, privacy, transparency, technology development, security, and regulatory responsibility. The U.K. and Japan are continuing their research, though AI laws have not yet been enacted [33,34]. Korea enacted Principles for a User-centered Intelligent Information Society in 2019, National AI Standards in 2020, and the Reliable Artificial Intelligence Realization Strategy in 2021, which embodied KAIEPs [32,33]. The Korean National Assembly proposed the Act on the Postering of AI and the Establishment of Trust Foundation in July 2021. The Act was adopted on January 21, 2025, and was implemented on January 22, 2026, for the first time in the world. The Act includes the governance framework for the sound development of AI and the establishment of a trust foundation. It includes details such as the enactment of AIEPs, the establishment of a private autonomous AI ethics committee, the preparation of detailed measures to create the foundation for establishing trust, support for verifying transparency, safety and reliability of AI systems, and regulatory measures for high-impact artificial intelligence.

Previous studies on AIEPs are included in three flows depending on the number and content of ethical principles. The first flow is studies that have extracted about 11 ethical principles by collecting universal principles among various ethical principles. Jobin et al. (2019) synthesized a total of 84 previous studies on AIEPs using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework [17]. In the order referenced in many documents, they extracted a total of 11 ethical principles: transparency, justice and fairness, non-maleficence, responsibility, privacy, beneficence, freedom and autonomy, trust, sustainability, dignity, and solidarity. Since then, almost all studies on AIEPs have cited their work. For example, Ryan and Stahl (2021) embodied the details of each principle based on the 11 ethical principles of Jobin et al. (2019) [17,19]. Ethical principles and similar meanings integrated by Jobin et al. (2019) are shown in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Principles of AIEPs [17].

https://doi.org/10.1371/journal.pone.0356766.t001

The second flow is studies that specify or update a larger number of AIEPs. Hagendorff (2020) reviewed 22 previous studies published between 2016 and 2019 to increase timeliness and incorporated similar principles [18]. In the order referenced in many documents, he also extracted a total of 22 ethical principles: privacy protection, fairness/non-discrimination/justice, accountability, transparency/openness, safety/cybersecurity, common good/sustainability/well-being, human oversight/control/auditing, solidarity/inclusion/social cohesion, explainability/interpretability, science-policy link, legislative framework/legal status of AI systems, future of employment/worker rights, responsible/intensified research funding, public awareness/education about AI and its risks, dual-use problem/military/AI arms race, field-specific deliberations (health, military, mobility etc.), human autonomy, diversity in the field of AI, certification for AI products, protection of whistle blowers, cultural differences in the ethically aligned design of AI systems, and hidden costs (labeling, click-work, content of moderation, energy, resources).

The third flow is studies that simplify AIEPs to increase feasibility. After 2020, studies on how to actually apply AIEPs to AI systems have been conducted. Dozens of detailed rules and guidelines have been made that embody AIEPs for application [19]. As the number of rules became excessive, overlapping and conflicting meaning between rules occurred, making it rather difficult to apply them into the system [19]. Efforts have been made to extract a few most fundamental AIEPs. Floridi and Cowls (2019) tried to minimize AIEPs to six to make it easier to handle them [20]. They argued that the current AIEPs are generally similar and that a total of five AIEPs are sufficient: four traditional principles, beneficence, non-maleficence, autonomy, justice and one new principle, explicability. Kaur et al. (2022) argued that fairness, explainability, accountability, reliability, and acceptance are sufficient to make trustworthy AI systems [21]. Mikalef et al. (2022) proposed a total of eight AIEPs: fairness, transparency, accountability, robustness/safety, data governance, laws and regulations, human oversight, and societal and environmental wellbeing, for responsible AI [22]. Morley et al. (2019) adopted five AIEPs of Floridi and Cowls (2019) to reduce the gap between ethical principles and field practice of actual AI system development [23]. Bankins and Formosa (2023) verified whether AI changes workers’ perception of meaningful work, and proposed the AI4People framework, which refers to the AIEPs of Floridi and Cowls (2019) [24]. Vakkuri et al. (2021) proposed the ECCOLA framework for Value Sensitive Design (VSD) of ethical AI system development, linking 21 detailed design elements to eight AIEPs: stakeholder analysis, transparency, safe and security, fairness, privacy and data, human agency and oversight, wellbeing, and accountability [25].

The integrated AIEPs and update

Methodology, analysis and results.

The challenge to derive AIEPs (integrated AIEPs) by combining the results of dozens of previous studies has been successfully achieved. Integrated AIEPs can be referenced to create new AIEPs, formulate several derivative AIEPs, or update existing principles by reflecting the specific contexts of each country, culture, and region. A total of 37 representative previous studies published from 2016 to 2025 were mapped into 17 integrated AIEPs. This mapping was conducted through a rigorous review by five researchers who have studied AIEPs for many years. Among the experts, four were professors in Management Information Systems (MIS), and one was a researcher at a national research institute specializing in broadcasting and telecommunications policy. To develop the integrated AIEPs, a total of three rounds of expert discussions were carried out. In the first round, the experts meticulously examined the codes presented in each literature and grouped similar codes into a single principle. During this process, reference was made to the principles table utilized for integrating codes in Table 1 of Jobin et al. (2019). Code groups other than the principles in Table 1 were designated as new principles by carefully reviewing the meanings of the included codes, resulting in a total of 17 integrated AIEPs. In the second round, the experts mapped which codes were included under each principle and how frequently. As explained in footnote 1 of Table 2, the larger the number of included codes, the greater the number of circles representing that principle in Table 2. The detailed codes included in each principle are presented in footnote 2. In the third round, through comprehensive discussions, the experts resolved discrepancies in opinions and reached a consensus to finalize the integrated AIEPs. These include justice and fairness (JF), transparency (T), responsibility (R), safety (NMS), wellbeing (W), human rights, freedom, autonomy, and dignity (HR), robustness (TR), privacy (P), non-maleficence and prohibition of infringement (NMI), common good and trust (CG), data management (DM), controllability (C), sustainability (SUS), inclusiveness (I), solidarity (S), no military abuse (ML), and openness (O), which collectively encompass existing ethical principles. Table 2 illustrates the AIEP mapping results.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. The integrated AIEPs.

https://doi.org/10.1371/journal.pone.0356766.t002

Furthermore, this mapping allows us to assess how frequently each AIEP was selected in the previous studies. Principles commonly cited across previous studies indicate that they are universal and represent the intersection of AIEPs; thus, they should be positively considered when creating new AIEPs or updating existing ones. Table 3 presents the citation frequencies and rankings of the AIEPs. From 2016 to 2025, justice, transparency, responsibility, safety, wellbeing, and human rights emerged as the top universal priorities. The rising rankings of wellbeing and human rights reflect post-2020 shifts in societal values [17,18]. Because these core principles encompass individual, social, and technological values—and overlap with prerequisites recommended by the All4People framework for AI development [20,24]—they must be integrated into AI systems as much as possible. The next tier includes robustness (7th), privacy (8th), prohibition of infringement (9th), common good (10th), data management (11th), controllability (12th), and openness (17th). These principles focus on stable service provision, data management, human-centered control, and risk minimization, and thus warrant continued attention. Robustness stands for the provision of stable services. Privacy stands for the right to not be unfairly infringed upon by others, and prohibition of infringement stands for the protection of users with the obligation not to infringe on the rights of others. Common good means the public interest that artificial intelligence must contribute to the public good. In addition, data management ranked 11th and controllability ranked 12th are also closely related to data management for system operation, human-centered AI control, and damage prevention, so they are included in the next lower group. Since the principles included in the next lower group are the values for the stable operation of artificial intelligence, it is better to respect them as well. The final tier comprises sustainability, inclusiveness, solidarity, and no military abuse, which policy makers can adopt depending on specific contextual needs. In the case of sustainability, the principle is to achieve socioeconomically sustainable growth by minimizing energy consumption and protecting the environment in the process of AI operation. Inclusiveness, solidarity, and openness means social inclusion, and no military abuse means the prohibition of human attacks by AI. There is an opinion that these principles overlap with other principles, so it is better to decide whether to include them or not according to each situation focused by a policy maker.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. The referenced priority of AIEPs.

https://doi.org/10.1371/journal.pone.0356766.t003

To capture post-2020 developments, Table 3 compares citation frequencies and percentages of the total mentions before and after 2020. Mentions of responsibility, wellbeing, human rights, controllability, and sustainability increased after 2020, reflecting a broader shift toward implementation guidelines, legalization, human rights protection, human-centric control, and environmental balance amid generative AI advancements. The increase in the referenced number of responsibility, human rights, wellbeing, and controllability refers to the need for human-centric AI control. Since 2020, people are now more interested in what AI can do for humans, and human control is needed to reduce unexpected impacts. The increase in the referenced number of sustainability means the balance between the use of AI that requires enormous energy and the climate, energy, and environmental protection. Conversely, mentions of transparency, robustness, prohibition of infringement, and inclusiveness declined as the literature concentrated heavily on system deployment and utilization. Nevertheless, transparency and robustness should be considered continuously for the development of explainable artificial intelligence. Prohibition of infringement and inclusiveness are principles for social consideration of infringement of the rights of others and inclusion of the socially underprivileged. It is regrettable that the mention of these principles has been declined, as previous studies mainly focused on the development and utilization of artificial intelligence systems since 2020. In summary, updating AIEPs requires maintaining continuous attention on transparency, robustness, and prohibition of infringement while emphasizing responsibility, wellbeing, human rights, controllability, and sustainability to align with growing contemporary demands.

Updating KAIEPs

As the use of AI has also spread in Korea, ethical issues such as data bias and misuse of AI have been continuously raised. The Korean government developed AIEPs in 2020 with the goal of ‘realizing human-centered AI development’. They defined humanity as the highest value and tried to create AIEPs to achieve three purposes: human dignity, the greater good of society, and the rightful purpose of technology. For three purposes, the Korean government collected the ethical principles mentioned in a total of 25 previous studies and extracted a total of 15 principles such as pursuit of happiness (wellbeing), human rights, privacy, respect for diversity (diversity), prohibition of infringement for respecting human dignity, such as pursuit of the greater good (common good), openness, solidarity, inclusiveness, data management for the greater good of society, such as accountability, controllability, safety, transparency, robustness for the rightful purpose of technology. Most of the KAIEPs except sustainability and no military abuse is included in the 17 integrated AIEPs. Therefore, KAIEPs also have high validity and reliability [33]. Just as the EU used the High-Level Expert Group on Artificial Intelligence (AI HLEG) to create ethical principles for trustworthy AI [13,52], the Korean government also consulted and collected opinions from the expert group of officers in government, researchers in think tanks, and engineers in bit tech companies. A total of five principles has been deleted since the review of the expert group. If there are too many principles, trade-offs or ethical dilemmas between principles may arise, depending on the situation due to duplication and overlap of the content. Diversity, openness, and inclusiveness are the contents that prohibit unfair discrimination, so they were integrated with diversity. Controllability and safety were integrated into safety. Wellbeing is an abstract and comprehensive concept more than other principles, and there is a concern that it may overlap with other principles in that it can be applied to a too wide range of situations. Robustness has been deleted in response to the experts’ point that it is a standard that is difficult to implement in practice at the present time, considering the technical limitations of AI such as uncertainty and unpredictability. Through this process, the final version of 10 KAIEPs of human rights, privacy, diversity, prohibition of infringement, common good, solidarity, data management, responsibility, safety, and transparency was confirmed. Table 4 shows the confirmed KAIEPs.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Description of KAIEPs [33].

https://doi.org/10.1371/journal.pone.0356766.t004

Applying the integrated AIEPs to KAIEPs and updating them is as follows. When referring to the recent changes in the referenced ranking of the integrated AIEPs, it is desirable to revive sustainability, wellbeing, robustness, and controllability and include them in KAIEPs. Sustainability was not even included in the original version of KAIEPs, but it has been emphasized since 2020. Sustainability means that AI must be able to contribute to the balance between socio-economic development and environmental protection and to achieve sustainable growth. Sustainability is a principle that has been treated as important enough to rank 6th out of Hagendorff (2020)’s 22 principles [18]. Korea, in the past, focused on economic development, but recently is interested in business practices with environmental, social, and governance consideration. Therefore, it would be better to newly include sustainability in the KAIEPs. Wellbeing, controllability, inclusiveness, and openness was excluded from KAIEPs because the concept was too abstract or overlapping with other principles, and robustness was excluded from KAIEPs because it was difficult to implement at the current technology level. Reconsideration should be made by reflecting the recent changes that follow from general usage of Generative AI, and wellbeing, robustness, and controllability could be revived. Wellbeing means that the effects of AI should improve various areas of human life. As AI can be used by anyone, such as individuals, companies, and governments, interest in the benefits of AI is increasing significantly. The fact that wellbeing’s reference ranking has risen to 4th rank, and the referenced percentage has also increased by 3.4% since 2020. Article 10 of the Korean Constitution stipulates human dignity, values, and the right to pursue happiness. Wellbeing should be included in KAIEP. Robustness is the principle that AI must operate stably even in highly uncertain situations by promoting technological development of AI. Robustness is considered an important principle to be ranked 7th in the referenced ranking. Technological advancement enables stable service provision of AI, which was not possible in the past. The Korean government is pursuing policy tasks with the aim of becoming one of the top three AI powers. Controllability is integrated with solidarity. Controllability has overlapping meaning with solidarity in terms of being able to immediately minimize damage in an accident, but it also has overlapping meaning with human rights in terms of that all activities using AI should be led by humans. Controllability belongs to the lower group at 12th rank in the previous referenced ranking but has recently risen to 7th rank. Controllability is a principle that is emphasized in practice when developing an AI system and has been frequently referred to since 2020. Since inclusiveness and openness have a low referenced ranking and overlap with diversity, the Korean government’s judgment that these principles are integrated with diversity is reasonable. In summary, I propose that wellbeing, robustness, and controllability should be revived, and sustainability should be added to the 10 KAIEPs to update them according to the current technological and socio-economic changes. The 14 updated KAIEPs could be diversity, transparency, accountability, safety, wellbeing, human rights, robustness, privacy, prohibition of infringement, common good, data management, controllability, sustainability, and solidarity.

Importance-performance analysis on KAIEPs

Methodology, analysis and results.

Importance-performance analysis is a method often used to derive strategic implications simply by mapping the potential importance and performance of a strategic consideration object to the four quadrants [53,54]. This method has the advantage of being able to clearly understand the strategic direction and resource provision priorities along with the average evaluation of the strategic consideration objects [53]. For objects located in the first quadrant that have higher importance and performance, high performance maintaining strategy should be considered and resources should be allocated in the first order. For objects located in the second quadrant that have high importance, but low performance, additional resource allocation and innovation strategies are required to improve current performance. For objects located in the third quadrant that have low importance and performance, an exit strategy should be considered through business portfolio adjustment. Lastly, for objects located in the fourth quadrant that have low importance, but high performance, their importance in the future decreases, so it is necessary to continuously obtain profits but consider gradually reducing the business scope [55]. The importance of KAIEPs was asked “This principle is important, needed, and critical”, and the performance was asked on a Likert 7-point scale, “This principle is already well reflected in the AI system, the performance of the AI reflecting KAIEPs is high, and it is well implemented beyond my expectations.” on a Likert 7-point scale. A total of 150 practitioners were recruited and required to answer the online survey material. 144 valid responses were obtained. Most of the respondents were employees in their late 20s in their first to fifth years as field operators, majoring in management information systems, and had at least one year of AI application development experience.

Around the enforcement of the AI law on January 22, 2026, a tentative evaluation of the adoption performance of 10 KAIEPs was conducted and implications for the future supplementing KAIEPs were drawn. The reason for evaluating 10 KAIEPs instead of the updated 14 principles is that I wanted to measure the actual AI system development performance and compare it with its potential importance. The reason why the previous studies have also simplified the dozens of AIEPs to about five since the 2020s is that it was difficult to reflect too many AIEPs during actual system development [20]. I conducted an importance-performance analysis to compare the potential importance with the current perceived performance of each KAIEP. Fig 1 shows the result of the importance-performance analysis for KAIEPs. Human rights, transparency, and common good are located in the fourth quadrant, and solidarity is located in the adjacent third quadrant. These principles symbolize the protection of individual basic rights and the value of public interest and democratic solidarity. The theoretical interpretation according to the quadrant position is a strategy to maintain or exit the status quo. However, it should be noted that the importance ratings for all ten ethical principles consistently exceed their performance ratings, and the overall mean of importance is thus higher than that of current performance. It is correct to interpret that these principles are well reflected in the system in a balanced way through efforts so far. In the second quadrant, accountability, privacy, safety, and prohibition of infringement was located in the order of importance, and data management were located in the third quadrant immediately adjacent to each other. Accountability is a principle of institutionalization and governance for responsible AI. Despite various efforts such as the establishment of ethical guidelines since 2020, the achievements experienced by practitioners are still unclear as these efforts are mainly made through the high-level expert group of AI (AI HLEG) in EU and Korean cases. Privacy, safety, prohibition of infringement, and data management are the principles for safe AI system development. Despite the importance of these principles, analysis result shows that it is never easy to create a safe AI system that can eliminate or reduce potential risks due to the fundamental non-predictability of AI. Privacy protection and system security are still considered important issues with AI technology deployment. In order to improve performance, reflection on the current situation and strategies for innovation and additional resources allocation are needed. In summary, the results of the importance-performance analysis of the 10 KAIEPs show that continuous efforts to supplement the system are needed according to the enforcement of the AI law because overall performance is insufficient compared to the importance in the institutionalization process. It is necessary to continuously reflect the principles of social value that have shown a high level of performance. Efforts to improve performance are urgent for the development of AI governance framework for responsible AI and the adoption of privacy and security principles for safe use of AI.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Importance-performance analysis result. The dotted line represents the importance-performance average of the principles.

https://doi.org/10.1371/journal.pone.0356766.g001

Discussion and conclusions

By organizing a total of 37 previous studies, I successfully extracted a total of 17 integrated AIEPs: Justice, transparency, responsibility, safety, wellbeing, human rights, robustness, privacy, prohibition of infringement, common good, data management, controllability, sustainability, inclusiveness, solidarity, no military abuse, and openness. Based on the referenced rankings, the universal priorities were determined. In order to maintain timely ethical principles, I proposed to monitor whether the three principles of transparency, robustness, and prohibition of infringement are well applied to the development of AI systems. The five principles of responsibility, wellbeing, human rights, controllability, and sustainability need to be emphasized and updated. Reflecting the integrated AIEPs, I proposed increasing the existing 10 KAIEPs to 14 by adding wellbeing, responsibility, controllability, and sustainability. The results of the importance-performance analysis of KAIEPs suggest that the social value principles of diversity, human rights, transparency, common good, and solidarity, which showed high performance levels, should be continuously reflected in the development of AI systems. At the same time, the results show that there is an urgent need to prepare an AI governance system for responsible AI in concerns to accountability and to develop a privacy protection and security system for safe AI utilization in concerns of privacy, safety, prohibition of infringement, and data management.

The academic value of this study is as follows. First, this study extends the research scope, which was mainly focused on AI technology development, to personal and social values and ripple effects that AI should create in the future. With the advancement of AI technology, this is no longer just an automation technology. I am watching numerous AI application services take over even more diverse, complex and detailed human tasks. The more AI is used, the greater the ripple effect that AI has on individuals and society. Along with the question of how AI can be developed from the perspective of science and engineering, the question of what AI should be made for from the perspective of humanities and sociology is also important. This is because AI is, after all, a tool for the person who uses it. This study focused on what values AI should create by studying AIEPs for the right purpose of technology, guaranteeing the basic rights of individuals, and promoting the public good of society.

Second, this study balances strategies to increase the positive effects of AI through AI technology development and industrialization, and strategies to minimize unintended impacts by creating regulations for responsible, trustworthy, and safe AI. Many previous studies have been conducted mainly on technology development and user acceptance, focusing on the positive ripple effects of AI. Despite concerns about the fatal threat of AI abuse, only dozens of studies have been conducted to eliminate or minimize unintended impacts. Some people argue that open regulation or non-regulation is more desirable than regulation for the promotion of the AI industry and the spread of services by focusing on the strategic value of AI. This opinion is one of the reasons why the EU’s enforcement of AI laws has not gone smoothly [56]. However, from the viewpoint of providing safe services to users and developing the entire industry through fair competition, promotion policies and regulatory policies are not conflicting values. This imbalance must be resolved. This study tried to reduce the potential risk of AI by conducting research on AIEPs.

Third, this study is one of the very few studies that have proposed integrated AIEPs. Most of the previous studies on AIEPs have created ethical principles in different numbers and contents. It is natural that the ethical principles emphasized are different because history, culture, and ethical values are different for each country, institution, and situation. Since the ethical principles were different for each study, there was no universal framework of AIEPs that countries or companies that needed to create or update AIEPs could refer to. I created 17 integrated AIEPs by combining the results of dozens of existing previous studies. Integrated AIEPs contain all of the most ethical principles mentioned in previous studies. I have made a union and combination of principles that people who want to successfully create AIEPs can refer to.

Fourth, this study summarizes the universal priorities between each principle. By compiling the referenced number and rankings of AIEPs, it can be determined which ethical principles should be considered first. AIEP with high referenced ranking are principles that must be considered universally, even in various situations. If there is an overlapping conflict between principles, you can refer to ethical principles with high priority. It also helps those who want to create AIEPs to create customized ethical principles, such as different numbers or different content of the ethical principles that they want to introduce in each country, organization, and situation.

Fifth, this study created a highly timely ethical principle by updating AIEPs. Most AIEPs were created in 2020. Despite dramatic changes, such as the emergence of ChatGPT in 2022, subsequent use of AI by the general public, and deployment of AI in industries, these changes have not been reflected in AIEPs. This study calculated and compared the changes in the referenced scores of AIEPs before 2020 and between 2020 and 2025 and found out which principles should be further emphasized. The results suggest that greater attention should be directed toward three key areas: Responsibility, by ensuring that ethical guidelines are harmonized with current laws; Wellbeing, by prioritizing AI’s benefits to humanity in response to public interest; and Sustainability, by embedding corporate social responsibility and environmental stewardship into ESG management. Ethical principles must be updated continuously. Therefore, it is hoped that longitudinal research to be conducted every five years will maintain timely ethical principles.

Sixth, this study introduced the process of deriving strategic implications of how KAIEPs were made, how they should be updated using integrated AIEPs, and what KAIEPs require more effort to improve progress by comparing the potential importance and tentative performance of each ethical principle. By introducing the world’s first case of Korea enforcing AI laws, regulators in other countries who want to implement AI laws later can benchmark them.

The practical value of this study is as follows. First, this study described detailed items of AIEPs in each previous study in Ttable 2 for creating integrated AIEPs so that practitioners who want to create or update their own ethical principles can refer to them. By providing specific data, practitioners can contribute to making more sophisticated principles.

Second, this study reorganized KAIEP by referring to the integrated AIEPs. I suggested that the KAIEPs be increased to 14 by adding wellbeing, responsibility, controllability, and sustainability. They were dropped at the time of enactment of the KAIEPs in 2020 according to the opinion of AI HLEG, but they have increasing importance to adopt KAIEPs in the actual development of AI systems since 2020. This can lead to various discussions about the roles and limitations of AI HLEG for the improvement of AIEPs. A continuous and constructive discussion to make consensus involving the general public and stakeholders from various fields can help minimize policy failure. Korea’s case can also teach other countries a lesson that continuous updates prevent policy failures.

Third, this study conducted an importance-performance analysis of KAIEPs to re-examine ethical principles and derive implications for policy adjustment. Policies must be adjusted through continuous evaluation even after implementation to successfully achieve policy goals and maximize performance. The principles reflecting the social values that have shown a high level of performance should be continuously reflected in AI system development. It is urgent to prepare a governance system for responsible AI and to reflect the principles of privacy protection and system security for realizing safe AI systems. The arguments of this study support efforts to emphasize the legitimacy of AI regulations and the consistency between laws. At the same time, for practitioners, the need to strengthen privacy protection and system security can explain the necessity of development for these considerations that have been marginalized and help secure the legitimacy of budget investment.

Despite these various strengths, this study has several limitations that could not be resolved in research progress. First, this study derived integrated AIEPs using the summary of previous studies. This methodology is not a quantitative research method, and validities of this study could be limited, comparing with other rigorous quantitative methodology such as laboratory experiment and field survey methods. Supplements have been made to enhance the internal validity of methodologies such as using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework used to synthesize the principles by Jobin et al. (2019) [17]. Second, the importance-performance analysis of this study sent an online survey link to each company to have a person in charge of AI system planning and development respond, with 144 respondents, and most of the responses were from practitioners in their mid to late 20s, mainly in small and medium sized enterprises. Given the potential importance and investment capacity of AI, subsequent studies will increase the validity of the research results if they receive more responses from companies of various sizes and AI users with various demographic backgrounds. Third, while this study examines 10 KAIEPs, the associated self-check tables and 35 practical guidelines contain an extensive number of criteria [57]. Given that these principles significantly impact corporate regulatory compliance, companies may struggle to allocate the necessary time and resources to adapt to such comprehensive frameworks. Although this study aimed to streamline and selectively adopt existing guidelines, the findings paradoxically led to a further increase in the number of principles. Consequently, it is recommended that future research focus on integrating these guidelines to enhance their practical applicability in real-world settings. Despite these limitations, this study will be useful as basic data for subsequent research on AIEPs, and it provides many ideas for stakeholders who want to introduce AIEPs. It is hoped that this study will contribute to realizing responsible, trustworthy, and safe artificial intelligence systems.

Supporting information