Introduction: In July 2023, the Democratizing AI Futures dialogue was launched using the Ideathon platform, and the Democratizing AI Futures Deliberative Workshop sessions were held in Taipei and Tainan Cities in August and September of the same year, respectively. The event brought together AI experts, scholars and the people to discuss the societal impact of generative AI and explore solutions through continued dialogues.

Center for Innovative Democracy and Sustainability (CIDS), National Chengchi University (NCCU) was commissioned by the moda's Department of Digital Strategy to hold two Democratizing AI Futures Deliberative Workshops. The Center comprises team members from the public sector, practice fields and academic circles dedicated to improving democratic participation procedures and social dialogue methods. The center's vision for social dialogue is to start with the issues and progress through various stages of communication and collaboration with social partners. Specifically, social partners with policy interests, such as citizen activists, organizational groups, experts and scholars, and government departments, engage in cross-field or cross-level dialogue via discussions, consultations and joint actions, to confront today's complex issues, build consensus and promote democratic participation.

CIDS members' creative thinking increases public participation channels in public policy. The center's recent research has focused on the governance of technological subjects, such as NIMBY technology facilities and human rights, social impact assessment on technological innovation, nuclear waste site selection and citizen participation. The goal is to bring public issues to the table to find the best solution for social development after integrating opinions from various forums.

Generative AI combines Big Data computing, real-time content production, and machine deep-learning technology to develop the ability to imitate and create. The application of AIGC has the potential to revolutionize modern society's labor structure, industrial layout and even daily habits, bringing significant impacts and opportunities to social operations. According to GitHub, the world's largest program code management platform, 92 percent of American program developers indicated they use generative AI to write programs at work, and 70 percent of interviewees believed that using AI can improve programming quality and work efficiency. In the past, however, most people in the art and creative industries regarded human intellectual property worthy of legal protection. Nowadays, combining generative AI with personal creativity can vastly increase efficiency and precision, but may also trigger debates on intellectual property ownership. Moreover, regarding the negative impact on society, if people's social biases and prejudices can be generated using generative AI and quickly spread through the communication system, discrimination will be accelerated to some extent, and ideological dissemination and social antagonism may escalate. These factors must be carefully considered as we develop and utilize generative AI.

Generative AI is rapidly evolving and becoming a competitive field among major technology industries. The development of generative AI has marked the beginning of a new era in society, and the international community has also begun to consider and propose related control plans. For example, the Organization for Economic Cooperation and Development (OECD) defines AI as "machine-based systems that can predict and provide suggestions or decisions to affect the physical or virtual environment to achieve human-set goals. This system has a certain degree of autonomy and is a machine capable of performing human-like cognitive functions." It also proposed the "OECD AI Principles," which include recommendations for national policies and international collaboration. These recommendations are as follows: (1) Invest in AI R&D; (2) Build an AI-friendly digital ecosystem; (3) Create an AI-friendly policy environment; (4) Cultivate human resources and prepare for labor market transformation; and (5) Implement international cooperation on trustworthy AI (OECD, 2023 17). The OECD research report also said that generative AI can filter large amounts of data and assist governments in making policy decisions. It can predict and help prevent election interference events and improve public services. The OECD agrees that generative AI should be used in government governance, but cautions that it may undermine democratic values by perpetuating and amplifying social inequality. Generative AI may further undermine trust and social contracts if not adequately regulated.

In the context of Taiwan, the impact of generative AI on society is evident as per the international trend, and AI's interactions with democracy, society and politics are also apparent. However, most current discussions in Taiwan continue to center on industrial development, such as evaluating the use of AI by the private sector to cut costs or create benefits for enterprises. Beyond the development and economic aspects, the impact of generative AI on democracy, society and politics should be critical questions for governance and long-term social development. Regarding academic research development, many studies have recently raised concerns about the various ethical threats that generative AI may pose to the public sector. Since different technological waves will result in other aspects of government transformation, they will also generate other public services and values. On the other hand, research directions that are policy-responsive, social demand-oriented, and focus on social entities or different industrial categories have resulted in more policy-responsive and socially demand-oriented research directions.

Accordingly, NCCU was commissioned by the moda's Department of Digital Strategy to collect data from netizens who are interested in discussing AI issues using the ministry's polis.tw public opinion survey report. The goal is to evaluate the more divergent ideas about specific generative AI issues among the people. Taiwan-specific issues can be addressed through such discussions, enabling policymakers to identify and mitigate potential risks in the future.

• (I) Definition of participant representativeness The Ideathon platform initiated the two Democratizing AI Futures Deliberative Workshop sessions and launched the Democratizing AI Futures dialogue concept. Most participants in this event have contributed to Ideathon activities and volunteered to attend the workshop. The remaining participants are experts and scholars who were invited to attend. The source of participants in this event is specific target groups solicited via the voluntary and invitation systems. Not all event participants are front-line users. It is also clear that the proportion of people working in the design field is significantly higher than that of other fields, and there is a gap in their understanding, ideas and performance in technology applications. The appropriate source, number and representativeness of participant invitations depend on the purpose of the event design because one of the topics to be discussed in these two sessions is revised and used as the submission method for the next Ideathon. The composition of the participants in this event is connected and aligned with the purpose of the discussion. The event's design attempts to set the agenda by focusing on the impacts of AI on society. Various professionals can provide perspectives to create a preliminary blueprint in this direction. If there are plans to hold future discussions on similar topics, it would be helpful to consider making more diverse designs that consider the participants' representativeness. This could involve discussions in separate groups, with an invitation-only system for some sessions and a voluntary participation system for others. Such an approach can assist the event organizer in achieving the event's purpose and reaching the necessary opinion sources.
• (II)  Agenda setting for deliberative workshops The two Democratizing AI Futures Deliberative Workshop sessions started by collecting opinions from netizens. These opinions were summarized using polis.tw and supplemented with domestic and foreign news, research trends and other relevant information. The opinions were then categorized into six major themes (internal government governance applications, emerging law-supervision mechanism rule, information identification, data openness, intellectual property, education and human cognitive impacts) for on-site discussion. As the majority of participants in the Ideathon are volunteers, with some invited experts and scholars, the executive team assigns experts and scholars to groups based on discussion topics that align with professional fields. Since the voluntary contributors came from various backgrounds, they were randomly assigned to the topic groups. As a result, the opinions expressed in each group range from academic and practical to imaginative; the participants may not have a complete grasp of the assigned discussion topics. The practical operation of AI is limited by the technology's immaturity and lack of understanding among most people. As a result, many lack a comprehensive understanding or sufficient experience in dealing with the limitations and dilemmas of practical AI operations. The issues raised by AI applications are not about the knowledge or ability that can be manipulated through citizen consensus. It takes people with actual contact, training, or application experience to conduct a more in-depth analysis and propose prospective recommendations. If the opportunity to discuss similar topics arises in the future, we can prioritize the discussion topics, choose a singular topic or correspond to a specific occupation or industry category and conduct a more in-depth discussion. Only then can we gain a more accurate grasp of AI's practical applications or impacts on society.
• (III) Policy recommendation tracking and implementation Since citizen participation activities require significant pre-launch preparation and on-site communication, participants who are willing to spend time participating will likely have a certain degree of participation efficacy. There is a sense of efficacy or a certain expectation that the insight offered through discussions will be adopted. Even if the deliberation team has prior knowledge and on-site leadership, a good deliberation event can focus everyone's opinions on several meeting conclusions that most participants present are interested in and agree with. It may also be able to produce relevant policies through guidance. Even if the discussion during a meeting is engaging and effective in participation, the implementation of actual policy agendas and the impact of meeting conclusions on policies must still be executed by a civil service system with the appropriate powers and responsibilities. Common sense suggests that it is difficult for participants to continue tracking the subsequent policy implementation effects after similar discussion activities. Most policy development and promotion take time and are unlikely to be immediately enforceable, which has policy-oriented or social implications. The complex nature of such development issues makes it harder to confirm subsequent effects. Whether the outcomes of discussions held in event venues will have a tangible impact on policies, or if the content of the discussion aligns with the intended purpose is yet to be determined. Ultimately, this can only be tested through practical implementation. The results of these docking policies and practices are beyond the scope of what can be determined or guaranteed by current deliberation activities.

CIDS and NCCU advocate a social dialogue method that begins with analyzing the issues, inventorying the issue's current status on the network and identifying the challenges or problems that must be addressed. We then enter the stakeholder identification phase to consider who the stakeholders are and the purpose of each stakeholder's role in the context of this issue. Next, we can develop and customize operational methods based on practical needs during the issue diffusion phase through various levels of dialogue frameworks. The center's social dialogue design maintains deliberation and speculation, as well as emphasizes the importance of empathy and putting oneself in the shoes of others. We believe it is vital for every discussion to focus on theory and incorporate physical sensations in relating the issues to everyday life.

This workshop adopted the center's established methods to inventory and analyze the current state of generative AI issues. Additionally, opinions collected from polis.tw were examined to identify what the people think about the generative AI scenario. Finally, we investigated the causes of debates on various issues through the design and operation of the two deliberative workshop sessions, which can be used as a reference for future policy planning. The following is a brief description of the design concept and operation plan for this event:

• (I) Scenario setting

  We employed action drama performances at the start of the sessions to help participants easily understand the objectives of the subsequent topic discussion. The scenario setting of the action drama is as follows:

  How will the multiverse of generative AI change our current lifestyle?

  •  Scenario - a certain company
  •  Scenario - school

  

  "A" owns a consulting company that assists the private sector in developing generative AI services. Generative AI has advanced on a large scale over the years. In addition, generative AI development requires high professionalism. With abundant resources, "A's" capital has also increased significantly in the past few years. Recently, "A" helped company "B" to develop an image design robot. Employee "C" of company "B" is thrilled that generative AI can help to generate ideas, which has dramatically improved his work efficiency and eliminated the dark period of the past when he had to work overtime.

  Employee "D" stumbled upon a photo competition advertisement while wandering around the company after lunch. He was eager to participate because of the attractive prize, but the deadline was less than 10 hours. He had no time to create his own images due to his busy work schedule. So, he decided to use generative AI to help generate and modify the images. Sure enough, it took employee "D" less than 6 hours to revise and submit the images.

  Unexpectedly, reviewer "E" recognized Employee "D's" work as co-generated using generative AI. Reviewer "E" further discovered that part of the content generated by employee "D" came from the works of others, but failed to clearly indicate the source, which may violate the publishing work ethics. So, reviewer E objected and requested the organizer to disqualify employee D.

  

  In the multiverse of junior high schools, teachers use generative AI as teaching materials and tools, and students are also accustomed to using generative AI to learn and complete homework. For example, student "F" stated that whenever he had a question that he could not figure out, he would open ChatGPT for an answer. Teacher "G's" challenge was how to keep students from overusing generative AI – a common issue for all teachers in the multiverse.

  One day, student "F" announced in class he had written a code using generative AI that could change the teacher's photo to look like Lin Chi-ling, a Taiwan actress and model. Student "F" was very proud of developing this easy-to-use software and shared it with his classmates. One day, classmate "H" got bored and began to modify images using this software. He created a variety of materials for fun. One was to use classmate "F's" portrait to create a crying image. Classmates passed it around and unexpectedly reached the mobile phone of classmate F's Parent I, causing a commotion.

  Parent "I" even complained to the school's academic affairs office, assuming "F" was bullied by his classmates. Parent "I" walked into the academic affairs office and demanded dean "J" rectify the situation. At this time, dean "J" was using generative AI to create office documents. The multiverse's public service system has already created a generative adversarial network capable of checking whether application information is correct and completing the tasks quickly with technical assistance. After hearing the entire story, dean "J" realized that the students had gone too far with generative AI. Dean "J" began to consider ways to prevent students' perceptions from deviating from reality and whether it was time to revamp the teaching model.

• (II) Table assigning and topic setting

  After the action drama, the participants were divided into six tables, each covering one of the following topics: governance (internal applications), emerging rule of law (supervisory mechanism), information identification, data openness, intellectual property and education and human cognitive impact. The Q&A process at each table first defined each topic's challenges and issues. Participants first discussed and identified the benefits, challenges and risks associated with increased affinity and trust in generative AI applications, and then explored the possible solutions. The issues to be discussed at each table are first explained as follows.

  Topics	Dispute Scenario	Risks
  Government Governance - Internal Applications	How will the public service system be affected if Al is used to collect public questions and respond to public needs? If Al is used to simplify official document processes and accelerate office efficiency, what impact will it have on the office system? If AI is trained by high-level civil servants and used as an internal training mechanism for new civil servants, what impact will it have on the long-term development of the public service system?	If the information provided by Al is incorrect, a verification mechanism is required, or the government must provide accurate information to train the Al.
  Emerging Rule of Law - Supervisory Mechanism	The government should propose regulatory plans for AI applications, but given that Al technology professionals are primarily in the industrial chain, how will this affect the intensity of government supervision? If we establish a system to connect central and local governments, enterprises, and social networks for intelligence sharing and collaboration to create an early warning system for false information, how will it impact the government's regulatory capabilities?	The current regulatory plans for Al are primarily based on basic regulation or industrial internal control mechanisms, making it difficult to establish a strong social participation network or increase the density of supervision.
  Information Identification	If the Al computing black box is the key to the uncertainty of technology application, how will the widespread use of Al- generated data affect social development? If Al-generated products must include the source generation information via Al calculations to improve indexability, what impact does it have on improving the public's awareness of AI information identification?	There is no strict ban from using personal public data to train generative Al models. If information identification is difficult to trace, it can increase the risk of false information transmission.
  Data Openness	If Al can compile data, generate research papers or abstracts automatically, and improve the work efficiency of professional staff, what impact will long-term Al applications have on social development? Because the black boxof calculation is not transparent and it is difficult to access the data calculation process, humans are prone to distrust Al. What impact will long- term Al applications have on social development?	We must ensure that the Al system is subject to human supervision and is safe, transparent, traceable, non-discriminatory, and friendly to the social environment for different application risks.
  Intellectual Property	If Al creation is introduced during the early stages of work creation, can it be included as the copyright owner? Is there a risk of violating creative ethics if Al training is executed using works published by other creators on the Internet? Generally, creators are inspired by observing other people's work. Do we need to determine whether such cases are illegal?	Generative AI has superb creative capabilities through learning. With continued development, intellectual property that once belonged to humans may no longer have any protection value. On the contrary, we must actively discuss whether non-human works must be protected.
  Impact on Education and Human Cognition	If Al absorbs biased data and produces discriminatory or malicious information, and this negative information spreads due to its widespread use, what impact will it have on the education system? Al applications may cause students to become overly reliant on Al to provide instant answers and solutions while learning. What impact may this reliance have on students' cognition in the long run?	Al's lack of creativity and ingenuity compared to human educators may impede the students' critical thinking and problem- solving skills.

• (III) Discussion model and tool sheet design

  For this session, we designed the discussion process based on the preceding topic framework and discussion concept, as well as incorporated elements such as topic knowledge, empathy and emotions. Given that the majority of the participants are AI experts or members of relevant communities, most have professional research or empirical insights into the issues. Compared to the people's discussion model, we expect more participants for this session, and the discussions will be more focused and actionable.

  We focused the discussion model on specific audiences and maintained an open discussion style. Instead of an open discussion with no limits on the topic or argument, groups should conduct case studies on various topics. After the case studies, the table leader (facilitator) organized the members' views to ensure that everyone understood the topic and the discussion process could continue smoothly. Participant discussions were used to collect valuable opinions and perspectives on the topics.

  The composition of participants in each group is roughly summarized in the figure below. It includes the polis.tw community; information tool users, experts and scholars in related fields, as well as dialogue based on public governance perspectives. This discussion aims to envision new public governance methods and reach a consensus. To achieve this goal, we emphasized the importance of having a public governance perspective in addition to the participants' perspectives, like experts or community members. This approach can help achieve the desired discussion outcomes and enable the application of results effectively. The design for this part was also used and implemented for the discussion tool sheet and facilitator leadership skill guidance.

  Participants in this discussion were divided into six groups/tables based on the topic. Ten people were assigned to each table, including six representatives from relevant communities, two experts and scholars, and two facilitators to lead the discussion flow and guide the intervention from a public governance perspective. There were 60 viewpoint contributors in a single session, as shown in the figure below:

  

  The structure of this group discussion is roughly divided into two major parts. The first round "inventoried the current situation, advantages, and challenges of AI's impact on the society" while the second round focused on "how to use policy tools to promote coexistence between AI and humans." At the deliberation and design level, the first round mainly involves presenting questions and inviting the participants to consider the relationship between themselves and the topic. It is also the warm-up for the topic discussion process, which creates opportunities for mutual understanding and prepares each group to discuss and analyze the assigned topics further. Group sharing is held at the end of each discussion round to allow participants to understand cross-group opinions and construct a more complete imagination of the topic.

  During the second round, the facilitator guided the formation of the most basic classification based on public policies according to the views shared across the groups in the morning. The morning session addressed how policy tools can respond to risks and challenges by adopting a top-down public governance and bottom-up citizen participation model. After analyzing and sequencing the conditions, we formulated action plans that are both executable and innovative to usher in a new era of coexistence between humans and AI.

  Based on the session's purpose and discussion plan, our team created a tool sheet to help advance the on-site discussion. Participants are asked to share their positive and negative thoughts on AI applications using indicators. The participants of each group can observe each other and understand AI's preferences and application boundaries by reviewing the records and summaries on the record sheet.

  Then, we entered the topic discussion phase primarily associated with AI's impact on individual aspects, including advantages and challenges. We used the social structure classification framework to guide participants in sharing perspectives on personal cognition, stakeholder communities, social influence and national governance. The goal is to estimate the different needs and challenges faced by agencies at various levels of society. During this phase, we went beyond the question-asking level by guiding the participants to expand their original thinking framework and imaginations about the issues.

  The ultimate goal for the second round of discussions is to propose possible co-creation solutions. After discussing the topic in the morning, each group shared their understanding and the various problems and challenges that must be addressed. Before proposing solutions, we reset our expectations to re-energize the discussion. Although we all know there are many challenges and problems to solve, it is important to consider why humans still aspire to build a future society where AI coexists. What policy tools are currently available to help achieve this ideal? At this level, we categorized policy tools into three major categories: laws/rules, strategies/plans and competition projects. The goal is to assist participants in organizing solutions into specific categories, allowing them to effectively link to the corresponding policy development stage during policy implementation.

  Prior to implementation, our team usually conducts internal drills via the facilitator's workshop to establish the design framework and identify issues that can meet the original design vision while meeting the operation units' expectations. Drills also ensure facilitators can implement the framework and connect the issues and practices. The facilitators will guide the participants step by step to share opinions during the on-site discussion. Table records typically include writing keywords from the participants' opinions on N stickers or colored pen drawings, allowing participants to look back and construct their own views through opinion stacking. Finally, the facilitator summarized and organized the opinions to form specific suggestions within the group.