Taiwan’s approach to defense AI suffers from the same issue plaguing its overall defense establishment: A legacy from an authoritarian past, which disconnects the civilian government and defense establishment. This inhibits effective strategic alignment to array the country’s operational and tactical approach toward the Chinese threat in an asymmetric manner agreed upon by the country’s political establishment and its primary security guarantor, the United States.

Such a disconnect also prevents the defense establishment from effectively leveraging Taiwan’s impressive civilian technology innovation sector. Existing defense-AI programs tend to be piecemeal and unfocused and mostly aimed at “doing things cheaper and faster” instead of exploring new ways of doing things.

Even the rare attempt by the Taiwan Air Force to advance the future concept of operations (CONOP) incorporating AI as the centerpiece proved to be technically ambitious, organizationally far-reaching, yet conceptually ill-formed.

However, there is no shortage of grassroots ideas, within and without Taiwan’s defense establishment. If the country finds a way to funnel grassroots talent and solutions more effectively into an open defense innovation ecosystem, Taiwan stands to benefit significantly from the adaptation of defense AI in contending with China.

1 Thinking About Defense AI

1.1 Taiwan’s Unique Security Situation

Since the Nationalist Kuomintang (KMT) government retreated to Taiwan, Taiwan’s military has been, much like their Communist cousins on the mainland, loyal to the Party and its leader. Thus, instead of a dialectic process informing a national defense strategy, it was often decided within a relatively small circle surrounding the leadership. In turn, this created a situation where the armed forces were operationally and tactically competent yet lacked the capacity and culture to conduct strategic planning and foresight.

Today, Taiwan’s defense still suffers mismatched strategic, operational, and tactical objectives for Taiwan’s defense strategy from this authoritarian legacy. This further influences force structure, training, research, and development, and acquisition. And the advent of defense artificial intelligence (AI) as an issue of significance cannot escape this trend.

The situation is further exacerbated by Taiwan’s lack of international support due to China’s ongoing efforts to prevent all other countries from maintaining official relations with Taiwan, let alone arms sales. Thus, with few exceptions, the US has remained the sole source of defense systems acquisition for the island nation.

These arms sales, along with the Taiwan Relations Act, in which the U.S. promises to maintain its ability but not intent to intervene should China invade, constitute a rather shaky promise toward Taiwan’s security from its largest security partner in the world.Footnote 1

This means Taiwan is a country that does not get to decide its own future. Instead, the largest piece of its security strategy is in a constant state of flux, aptly termed “strategic ambiguity,” where the US attempts to simultaneously deter China from invading Taiwan, and Taiwan from declaring independence. This creates a powerful sense of contradiction within Taiwan’s defense establishment in both that Taiwan needs to independently fight against a Chinese invasion, and regardless of what Taiwan does, the fate of Taiwan might still be decided in Washington DC.

1.2 The Overall Defense Concept and Asymmetric Defense

Facing such uncertainty in allied support, along with decades of force reduction following Taiwan’s democratization and a lull in Cross-Strait tension in the late 90 s, leads to a military plagued by low morale and low investment. This situation is further exacerbated by the increasing shift of military balance across the Taiwan Strait, with the People’s Liberation Army launching its largest ever comprehensive reform in 2015. And as the operational approach against enemy forces enshrined within the country’s biannual National Defense Review gradually shifted from destroying invading forces outside of Taiwan’s borders to ensuring their destruction on the beach, the country gradually came to terms with the imbalance and attempted to exploit various strategic and operational asymmetries to its advantage, culminating in the Overall Defense Concept (ODC) (Lee et al. 2020). The document is aptly characterized with the slogan “a large number of small but lethal things.”

Yet the disconnect between the civilian government and defense establishment still remains, with the directives issued by the political institutions often failing to translate into operational or tactical realities. Instead, the official capstone documents such as the National Defense Review (NDR) and the Quadrennial Defense Review (QDR) merely serve to pay lip service to the concept promulgated by the civilian government. An example of this can be gleaned from the 2023 QDR, where strategic goals were logically placed, with the preservation of the democratic system placed above the preservation of life and property. However, when it comes to operational level discussions concerning the Overall Defense Concept, traditional platforms from capital ships to fighter jets were somehow characterized as fulfilling the asymmetric mandate without much explanation.

The convergence of these factors creates powerful incentives to search for a miracle solution to solve Taiwan’s defense woes. In this regard, the utilization of AI in defense is only the latest miracle, following autonomous drones and satellite communications. But the force reduction is only the symptom of another societal issue,Footnote 2 as Taiwan’s birthrate gradually declined over the decade, reaching its nadir in 2023 as the country with the lowest projected birth rate in world at 1.09 children per woman (台灣全球倒數第一 無解的難題 2023, CIA F). And for many in the defense establishment, adaptation of defense AI also represents an opportunity to ameliorate some of the pressures brought upon the existing force structure by the low rate of birth (陳建源 2018).

1.3 Defense AI under the Taiwan Context: Do it Cheaper and Do it Faster

Perhaps the most prominent public-facing conceptualization of Taiwan’s defense AI is nested within the larger context of the “Intelligent National Defense Program,” which began in 2019. Then Deputy Chief of General Staff General Li Ting-sheng, a proponent of the approach, would later disclose that the 10-year program was divided into six sequential stages (賴品瑀 2023):

  • Integration of various technologies with AI

  • Integration of Internet of Things (IoT) and 5G technologies with Taiwan’s unique strength in Information & communication technologies (ICT)

  • Support the use of AI-integrated IoT with big data

  • Battlefield situational awareness

  • Intelligent decision making

  • Cyberwarfare.

The National Chungshan Institute of Science and Technology (NCSIST), the top state-run defense research institution, also set up a parallel “Ten Year Intelligent Defense Plan” aimed at making the intelligent National Defense program a reality. Details of NCSIST’s plan were never made public, however anecdotal evidence suggests that these developments match closely with the Ministry of National Defense’ vision of a six-phased approach. Orchestrated by the information and communication research division, the plan includes a potential long range anti-submarine warfare (ASW) early warning and surveillance system that incorporates autonomous sensing, learning, identification, and intelligent remote guidance and control to enable autonomous unmanned aerial vehicles (UAV) and potentially unmanned teaming through the Artificial Intelligence of Things (AIoT) (涂鉅旻 2019). The essence of Intelligent National Defense, as then director of NCSIST’s information and communication research center, Lin Gao-chau stated, is the collection, integration, fusion and comparison of data to predict future situations, assist commander’s decision making (涂鉅旻 2019).

The program launch coincides with the Ministry of National Defense (MND) framing 2019 as the “Year Zero” for Taiwan’s intelligent defense, mirroring the “AI Year Zero” on the civilian side by the then Ministry of Science and Technology (朱泓任 2017). This highlighted the MND’s desire to take advantage of Taiwan’s robust tech ecosystem.

Many of these described efforts demonstrate three things concerning the Ministry of National Defense’s approach toward defense AI:

  • a propensity to favor the kinetic over the digital;

  • a conceptual origin mixed in the context of digitization and threats posed by such, e.g. the penetration of digitized network by PLA cyber actors;

  • and as a consequence of the first two, a very limited conceptualization of AI applications, with most applications filed under ICT and Command, Control, Communications, Computers, Intelligence, Surveillance, Reconnaissance (C4ISR) contexts.

It is worth noting that the development and standing up of Taiwan’s Information Communication Electronic Force (ICEF), colloquially known as the Cyber Force, also suffers a similar identity crisis.

1.4 Regulating Defense AI

As of the end of 2023, no specific code of ethics governing the use of defense AI exists in Taiwan. The development of defense AI would have to observe the AI Technology R&D Guidelines issued by the Ministry of Science and Technology (later National Science Technology Council, NSTC) in 2019 (科技部 2019). The Guidelines stems from three core values the ministry has identified as conforming to universal expectations on the use of AI in the global community:

  • Human-centric Values

  • Sustainable Development

  • Diversity and Inclusion

These three core values are further expanded into eight guidelines concerning the research and development of AI:

  • Common Good and Well-being

  • Fairness and Non-discrimination

  • Autonomy and Control

  • Safety

  • Privacy and Data Governance

  • Transparency and Traceability

  • Explainability

  • Accountability and Communication

Ethics governing specific aspects of AI applications also exists, and even predates the R&D guideline discussed above; the 2018 Unmanned Vehicles Technology Innovative Experimentation Act (Ministry of Economic Affairs 2018), where elements such as the protection of privacy and personal information, ensuring the safety of human beings, and the transparency of AI, are all present in some form.

Taiwan’s comprehensive answer to the question of ethics governing AI, the Artificial Intelligence Fundamental Act, is expected to be introduced sometime in 2024. The overall regulatory approach will be similar to that of the European Union, through amending specific regulations like, for example, the Personal Data Protection Act. The Fundamental Act itself will classify AI systems into various levels of risk and allow individual competent agencies to come up with specific regulations, e.g. the Financial Security Commission will regulate intelligence finance, and the Ministry of Transportation will draft detailed codes governing intelligent transportation (王儷華 2023).

1.5 Defense AI According to Taiwan’s Defense Documents

While exact details of the 10-year AI plan have not been publicly disclosed, parts of the vision can be gleaned from other official documents published by the MND. Taiwan’s primary external-facing capstone defense documents comprise the biannual National Defense Review (NDR), and the Quadrennial Defense Review (QDR). Under the general direction of building an asymmetric force largely following the ODC, 2021’s releases represented the first prominent mention of defense AI as an integrated component of Taiwan’s defense concept. The 2021 NDR identified several technologies that have implications for warfighting, as well as influencing the modalities of war. Chief among them AI, with specific mentions on the development of disruptive AI such as swarm and human-machine cooperative UAVs in the context of the US Third Offset Strategy, as well as the PRC’s current focus of AI application in unmanned systems.

Other disruptive technologies listed include precision strike, wargaming, modeling and simulation, and deepfakes (Ministry of National Defense 2021a, b). Judging that these would significantly improve the PLA’s joint operational capability and pose a significant threat to the Taiwan Strait and surrounding areas, the 2021 QDR stresses the need to integrate these technological trends into the development of Taiwan’s C4ISR. Specifically to integrate them into intelligent command and control (C2) systems in order to achieve two major objectives: first, improving situational awareness for the battlefield commander, and secondly, to provide battlefield assessment and decision-making aid (Ministry of National Defense 2021a, b). Additionally, the QDR also emphasized developing AI applications in prosecuting offensive and defensive computer network operations (CNO).

The following 2022 NDR, 2022 QDR, and the 2023 NDR all presented similar characterizations for the future direction of Taiwan’s armed forces as a combination of the asymmetric approach with the implementation of defense AI: “Mobile, small, man-portable & AI integrated.” This is further reflected in Taiwan’s latest 5-Year Force Planning Document submitted for review to the parliament in 2023 (陳鈺馥 2023), which echoed the characterization by stating the aim of developing a military that conforms to the asymmetric approach to defense as outlined in the ODC. Therefore, force planning should be geared towards improving the military’s capacity to take advantage of “Long range [fire], Precision [fire], Unmanned [systems], and AI [systems].”

Curiously, in the discussion for another major pillar of defense strategy as outlined in the 2021 QDR, that of establishing a self-sufficient defense industrial base (國防自主), AI was not identified as one of the major categories to be developed in accordance with the need for defensive combat operations, but was instead given only a cursory mention in the development of next-generation intelligent unmanned aerial and underwater vehicles (Ministry of National Defense 2021a, b). This stands in stark contrast to the positions taken by the government during the 2017 establishment of the then brand new ICEF, where one of the three major directives for the new branch issued by President Tsai was to pioneer the development of an academic and industrial base for Taiwan’s cybersecurity scene (楊孟臻 2017).

1.6 Origin of the Defense AI Civil-Military Ecosystem

On the civilian side, the “the Taiwan AI Action Plan” launched by the Ministry of Science and Technology pushes for the establishment of a civilian AI development ecosystem that leverages Taiwan’s unique position in the global semiconductor supply chain (行政院 2019). The origin of parallel civilian and defense technology development ecosystems can at least be traced back to the National Science and Technology Development Plan for 2013–2016 (行政院國家科學委員會 2013). Since then, one of the primary missions of the self-sufficient defense industry has been to spearhead technology and industry development, acting as the driving force for industry. This was further cemented in 2020 during President Tsai’s speech, in which the government centered the civil-military integration of defense industry as one of the country’s six major core strategic industry goals (National Development Council 2021). The details of such plans are further laid out in the Defense Technology Development Blueprint and White Paper, which also encompasses the plans to set up defense research centers within university campuses (Ministry of Science and Technology 2020), and the creation of the Defense Advanced Technology Development Program (國防部軍備局 2023).

2 Organizing Defense AI

There is no dedicated defense AI organization, from policy to research to production and employment, within Taiwan’s defense innovation ecosystem. Instead, various directives and experimental programs are nestled within different tiers of Taiwan’s Defense Technology Development Mechanism (DTDM), with the majority of disclosed research programs belonging to the Defense S&T Research Center initiatives located in various universities.

2.1 The Defense Technology Development Mechanism

The DTDM can be roughly separated into three tiers (Fig. 1); policy guidance and coordination, research and development, and production and manufacture. At the top is the Defense Technology Development and Implementation Committee, chaired by the Deputy Minister of Defense, the Vice Minister of Economic Affairs, and the Vice Chair of the National Science and Technology Council. The committee coordinates myriad government organizations and provides policy guidance for the research and development tier of the system (林柏州 2019).

Fig. 1
An illustration of Taiwan's defense technology development has 3 tiers. Guidance and coordination of Defense Technology Development and Implementation Committee, research and development involving universities and S O I C, and production and manufacture involving armed forces and the private sector.

Taiwan’s defense technology development mechanism. Source: 林柏州 2019, partially updated by Author, 2020, “Defense Industry Development Act, Article 11,” https://law.moj.gov.tw/ENG/LawClass/LawAll.aspx?pcode=F0110024

Full size image

The research and development tier of the DTDM spans both public and private sectors. A series of armed forces organizations, including the Navy’s Naval Ship Development Center (NSDC), the Army’s Ordnance Readiness Development Center (ORDC), the Air Force’s Aerospace Technology Research and Development Center (ATRDC), and the Armaments Bureau’s various arsenals occupy both the R&D tier as well as the production and manufacturing tier, as they sometimes take part in limited production for more specialized projects, either due to sensitivity or lacking economy of scale for serial production.

Outside the armed forces, the National Chungshan Institute of Science and Technology (NCSIST) leads research and development on sensitive defense projects. The projects that require security clearance unavailable to civilian institutions are exclusively conducted within the confines of the defense industry development institutions (國防工業發展機構),Footnote 3 directly under the auspices of the MND. This means that most publicized research projects are equal or below technology readiness level (TRL) 5.Footnote 4

The Ship and Ocean Industries R&D Center (SOIC) (Ship and Ocean Institutes R&D Center n.d.) and the National Defense Industry Development Foundation (NDIDF) (National Defense Industry Development Foundation n.d.) serve their respective niches as shipbuilding R&D centers and legacy funds for defense investment and tech transfers. Universities take center stage as the most vibrant defense AI element in the DTDM. The details and projects involved under the various universities’ Defense Technology Research Centers will be further elaborated in the next section.

2.2 Defense S&T Research Center and Corresponding Universities

The most active and public portion of the DTDM concerns the “academic research center (學研中心)” initiative. A joint venture between the Ministry of Science and Technology and the Ministry of National Defense started in 2020, the NTD5bn (USD167M) initiative aims to establish an initial three to six research centers in various universities in order to foster systematic and continuous research on defense applications for various technologies. The initial focus of the project consists of seven major areas of research:

  • Technology foresight research on sensing and precision manufacture

  • Advanced materials and analytical mechanics

  • Information, communication, electronics, and smart technologies

  • Critical systems analysis and integration

  • Cutting-edge power plants and aerial vehicles

  • Advanced surface vessels and underwater vehicles

  • Advanced systems engineering research

These research centers also serve as think tanks for defense technology development, advising on various topics including the defense application of artificial intelligence for the Ministry of National Defense (張玲玲 2023).

By 2023, there are at least seven defense S&T research centers located in various universities around Taiwan (科技部 2022), each specializing in one of the seven areas of focus (Table 1). These research centers are coordinated through the Defense S&T Research Center under the competence of the Defense Technology Development Implementation Committee (Fig. 1) located under the National Defense University’s Chung Cheng Institute of Technology (CCIT).

Table 1 University Defense S&T Research Centers and Specializations in Taiwan
Full size table

2.3 Defense Technology Research Programs for Academia

There are two major defense technology research programs with three distinct horizons for maturity under Taiwan’s defense innovation ecosystem; the Defense Technology Exploration Program (國防科技探索專案計畫, DTEP), which focuses on projects with a long maturity timeline (科技部 2022); and the Defense Advanced Technology Development Program (國防先進科技研究計畫, DATDP), which fund projects with maturity horizons of either 3 years, or 5–8 years depending on the category (蘇思云 2020).

While these programs seemed to be tailor-made for defense S&T research centers coordinated through NDU CCIT’s Defense S&T Research Center, application and participation in these research programs are not limited to the seven centers but are open to general academic institutions as well.

2.3.1 Defense Technology Exploration Program (DTEP)

DTEP, orchestrated by the Ministry of Science and Technology/NSTC,Footnote 5 is patterned after the Artificial Intelligence Exploration (AIE) and Microsystems Exploration (μE) programs conducted by the US Defense Advanced Research Projects Agency (DARPA). The DTEP aims to focus on dual-use technologies concerning small to medium unmanned vehicles in the air, ground, surface, and underwater terrains, as well as information, communication, electronics (ICE), and cybersecurity technologies (科技部 2022). The program invites submissions based on an envisioned scenario and uses a case in the 10–30-year horizon. Proposed projects will be evaluated first on the scenario set before technical evaluation. Successful pitches are given 18 months and a maximum of NTD8M (USD26,6700) per year to develop a proof of concept or a prototype, with the end results evaluated and rolled into consideration for approving future pitches by the same investigator (科技部工程司. 2022).

While DTEP has a stated goal of encouraging innovative thinking into future scenarios, the request for the proposal listed out extremely detailed guidelines taken directly from the year’s NDR on Taiwan’s current operational threat environment and current defense tactical needs. This somewhat defeats the purpose of an exploratory program with a long-term horizon, and to a certain extent, also explains the limited and scattered nature of the project requirements issued by different stakeholders within the defense innovation ecosystem.

2.3.2 Defense Advance Technology Research Program (DATRP)

DATRP orchestrated under the Ministry of National Defense covers a wide range of topics, where individual organizations under the ministry submit concept notes outlining interested projects and the timeframe of execution for academic institutions to apply. Generally, new projects are unveiled every year and can be divided into two categories: foundational research, with a max timeframe of 3 years, and breakthrough research, with a max timeframe of up to 8 years. The submitted projects follow the seven focus areas outlined for the academic research centers (Table 2). Submission agencies vary and include Navy HQ, Air Force HQ, ICEF HQ, and more. Some projects even originated from within DTDM’s research and development tier, such as NCSIST and the Navy’s. The number of projects submitted every year is also quite large. For 2023 alone, there were over 140 approved projects of varying sizes and length (國防部 2022a, b, c).

Table 2 Taiwan Technology Development Budget (in NTD100M)
Full size table

3 Developing Defense AI

This section will examine a selection of projects under the Defense Advance Technology Research Program (DATRP) from FY2022 and FY2023 involving the use of defense AI. Evaluation of technological maturity conforms to the technology readiness level (TRL) system first developed by NASA (國防部 2022a, b, c). Most projects were aimed at raising the underlying application’s TRL to 4 or 5 for further development.

3.1 Developing Artificial Intelligence Applications for Close Quarter Air Combat

The Air Force has always been Taiwan’s first line of defense against a Chinese invasion, and it has traditionally enjoyed political favoritism since the time of Chiang Kai-shek (許劍虹 2023). Thus, the attempt of the Air Force Technology Research Development Center (ATRDC) to develop an AI-pilot within a simulated environment (國防部軍備局 2023) deserves special mention as this project is unique as it is technically ambitious, organizationally far-reaching, yet conceptually ill-formed.

Inspired by DARPA’s Alpha Dogfight project, the goal of the 3-year project is threefold. First, to develop an intelligent platform capable of simulating operational scenarios involving AI pilots for tactics development and validation. Secondly, develop AI pilots that optimizes decision-making based on different scenarios, environments, and mission objectives. Thirdly, develop metrics for evaluating AI pilots’ performance and feasibility for different missions.

Conceptually similar to many Taiwan armed forces projects, its stated origin is to follow the “trend set by major military powers,” achieving a vision of “mosaic warfare” as set out by DARPA, where distributed shooters and sensors allowed the massing of firepower without having to mass forces (DAPRA 2018). Upon closer examination of the ATRDC’s project proposal language however, it is uncertain whether the Air Force had sufficiently considered its adaption under Taiwan’s operational context. In describing the future concept of operation, the proposal stated that the system should be able to achieve the following:

In a simulated environment, validate [the feasibility of] UAV wingman and manned lead aircraft can conduct missions together, that the human pilot can communicate and control the AI pilot to execute its assigned mission, with orders relayed through datalinks to the UAV wingman, where UAV wingman can achieve a high degree of autonomy through developed algorithm, and complete the mission assigned by the human lead, thus enhancing the Air Force’s capability to conduct manned/unmanned operations (國防部 2022a, b, c).

This is almost an exact copy of the US Loyal Wingman project’s conception of operation, aimed at operating against a near-peer power under a contested environment facing complex integrated air defense systems (IADS) (Losey 2022). However, the often-unstated assumption here is that this is used offensively during first days of a future war to conduct either deep penetration or suppression or destruction of enemy air defense (SEAD/DEAD) operations into enemy territory.

This is, however, far from the operational environment facing the Taiwan Air Force (TAF). The role of TAF is defensive, with contesting control of the air as its primary mission (中華民國空軍 2018). A mission Loyal Wingman would be ill-suited for this since it is principally developed, at least currently, for air-to-ground operations (Fish 2022). Furthermore, the extremely limited airspace above Taiwan is already heavily monitored, with the second highest density in radar stations and air defense missiles in the world (朱明 2021), controlled through four Regional Operation Control Centers (ROCCs), each capable of independently direct air defense of the entire island (徐葳倫 et al. 2021). The surveillance network is further augmented by airborne E-2 T AWACS. This comprehensive air defense network calls into question the utility of a loyal wingman-like platform to act as an independent sensor platform.

It is also interesting to examine the project’s focus on within visual range (WVR) engagements, an understandable focus given the limited airspace above the Taiwan Strait and the relative proximity between Chinese and Taiwan airbases. Any beyond visual range (BVR) engagement would very quickly devolve into a dogfight, and most engagements would offer little chance for coordinated and organized flow that a more offensively postured air force such as the US Air Force would be accustomed to.

Considering the defensive posture, and air-to-air mission focus of the force, another concept of operations (CONOPS) in the proposal entitled “maximizing UAV package recon/attack range” is even more puzzling:

In a simulated environment, when a UAV package is conducting recon/attack missions, it can utilize the minimum number of UAVs in achieving maximum effect. And although the goal is to minimize the number of UAVs engaged, redundancy should still be considered; beware of the tradeoff between reliability and UAV numbers to maximize operational range (國防部軍備局 2023).

It is hard to conceive that—under the current strategic guidance and operational doctrine –such an eventuality would manifest itself as a cost-effective measure for Taiwan’s armed forces. It is perhaps worth further careful examination as to what asymmetric advantage such an unmanned, AI-agent-based UCAV could actually offer under Taiwan’s strategic and operational context, especially considering the significant technical and organizational challenges to overcome.

A second category of issues stems from the ambition of the program in overcoming technical challenges in an impossibly short timeframe. Not only does the project attempt to replicate the results of Heron’s winning bid for DAPRA’s AlphaDogfight challenge (Tucker 2020), but it also further seeks to develop and implement human-machine teaming (HMT), which currently lacks any reliable framework for testing, evaluation, validation, and verification (TEV&V) (Motley 2022) within a yet to be developed simulated environment. The project further aims to involve not only academic research centers but also the defense industry development institutions, ATRDC, and active-duty pilots. All these efforts are to be accomplished, as defined by advancing development to TRL 5, within a mere 3 years.

Given the conceptual ambiguity and technical challenges facing such an effort, it is uncertain whether this particular effort represents an exploitation of Taiwan’s asymmetric advantage in defense AI.

3.2 Additional Defense AI Projects

In addition to ATRDC’s development efforts, Taiwan has also launched additional defense AI projects for support tasks, to advance predictive maintenance, for IT network management, and to assist space-based assets (機構系統開發 n.d.):

  • Optimizing Air Defense Launchers and Supporting Mechanisms with Artificial Intelligence (運用 AI 智慧技術優化防空裝備發射暨支撐機構系統開發)

Issued by the Army’s Missile and Electro-Optics Depot for FY2022–2023 (陸軍 2022), budgeted at USD1.5 M, the project aims to develop a series of autonomous stabilization and preventative maintenance systems using AI. Significant technical challenges are projected for creating the intelligent diagnostic system for motors, bearings, and thread rods loads; the digitized monitoring system for the support struts motivator; and the intelligent monitoring system for the missile’s box launcher motivator. All developments are judged to be a leap from TRL 1 to 4, as most of the prerequisite digitization to allow AI adaptations on machinery are non-existent.

  • Intelligent Balancing and Calibration of Vessel Propeller Shaft(艦艇動力旋轉機構智動平衡校正之研究)

A project issued by Navy Headquarters for FY2022–2024 (國防部 2022a, b, c), budgeted at USD100,000 for the first year. The project seeks to utilize AI in the total life cycle management of a naval vessel’s main screw(s), to facilitate preventative maintenance to increase mean time between repairs (MTTR) and estimate spare parts requirements. The project is divided into two major capstones: first, design a simulation of vessel propulsion system referencing MIL-STD-2189 (design methods for naval shipboard systems), validate with real-world data, and conduct dynamic analysis. Secondly, the project will develop a monitoring mechanism for the propulsion system’s vibrations. This would allow real-time monitoring of the fatigue developed through normal wear and tear and enable automatic stabilization when unexpected vibrations occur.

  • AI Trained SDN Network Orchestration for Network Management and Security Detection (以人工智慧導入SDN網路編排 管理與安全檢測之研究)

Issued by the ICEF Headquarters for FY2021–2023 (國防部 2022a, b, c), the project intends to utilize an AI algorithm to design a software-defined networking (SDN) architecture for Taiwan’s armed forces and reorient the armed forces’ various networked modules and platforms under this newly established architecture. The project is divided into 3 year-long phases with three respective capstones: establishing a network detection platform to analyze behaviors of individual packets and collecting information on SDN network and services for the first year, budgeted at USD60,000. The second year is spent on establishing an appropriate AI training and detection model to develop network orchestration techniques suitable for Taiwan’s armed forces, budgeted at USD83,333. These efforts would culminate in the integration and evaluation of the Armed Forces’ various platforms and modules subsumed under an SDN architecture in the final project year, with an additional budget of USD83,333.

  • AI-Assisted Next Generation Satellite Point Cloud Matching and Object-Oriented 3D Model Construction From Satellite Images (AI輔助新世代衛星點雲密匹配及物件導向三維建模研究以衛星影像為例)

A follow-on to a previous project establishing AI-mapped terrain for a common operational picture (COP) issued by Arsenal 401 for FY 2022–2024 (國防部 2022a, b, c). Budgeted at USD66,666 for the first year, the project aims to construct ultra-high-resolution 3D object-oriented satellite images from vector models. Using clustering techniques and AI/ML algorithms on existing point clouds produced from satellite remote sensing, a vector model can be identified as an object such as buildings. Sufficient mapping of these objects can then allow the detection and prediction of changing geological and terrain features, such as rivers and marine abnormalities.

3.3 Discussion

Additional defense AI research projects vary, from using AI to optimize UAV flight paths, to identify and map dangerous obstructions around airfields, to intelligent target identification for night-time urban warfare, and intelligent algorithms for tactical simulators. However, with the curious exception of the Air Force’s AI dogfight project, most of these revolve around niche applications that endeavor to enhance, rather than revolutionize, the way business is done within the Taiwan Armed Forces. While there seemed to be some doctrinal attempts at narrowing down use-case and scenarios, the basic understanding of defense AI application does not exceed data-based optimization on “how to do things cheaper and faster.”

4 Funding Defense AI

No dedicated budget categories for dedicated defense AI exist. Financial resources for projects reviewed above came from two sources: the academic research center initiative totaling NTD5bn (USD167M) over the course of 5 years, establishing seven defense research centers (蘇思云 2020), and the overall defense technology budget (Table 2).

Taiwan’s defense technology budget has steadily increased over the past decade, consistent with the government’s push to establish a self-sufficient defense industrial base through primarily public sector investment. In 2023, such expenditure accounted for around 3.16% of the total defense budget,Footnote 6 as compared to the United States, where research, development, test, and evaluation (RDT&E) accounted for about 17% of the total defense budget in the same year (Thomas 2023).

The annual budget for NCSIST, which conducts most application research once these development programs reach a certain level of maturity or sensitivity, has increased significantly over the past few years. While it is tempting to attribute these as pure R&D budget, it is important to note that NCSIST also conducts limited production of indigenous weapon systems. A non-trivial portion of its recent budget increase can be accounted for as part of the 5-year NTD240bn (USD8bn) contract NCSIST signed with the Ministry of National Defense on the indigenous production of eight major weapons systems (范正祥 2023).

Taiwan’s overall government technology budget as a percentage of GDP in 2023 to an impressive 4.5% (行政院主計總處 2023), whereas OECD countries average around 2.3–2.4% in this category (OECD 2023). After Taiwan’s election on 13 January 2024, the incoming Lai administration indicated their focus on five major categories of industry, which included both AI and defense as two out of the five (呂雪彗 2023).

5 Fielding and Operating Defense AI

While there have been several research projects on defense AI over the past few years, public disclosure of the Armed Forces’ operational defense AI applications has been relatively limited. A few scattered examples are outlined below to provide a feel of Taiwan’s sensitivity in disclosing its defense AI applications. Notably, a few of these examples even predate the armed forces’ focus on AI starting in 2019 (國防部 2017):

  • Intelligent ECG Analysis Platform

An advisory panel on the development of intelligent military medicine was convened in 2019 consisting of members from the MND Medical Affairs Bureau, physicians, and industry representatives, to promote the development of AI application in medical affairs. In cooperation with Quanta computer, the armed forces established an AI Lab in 2020 (國防部 2021a, b). The effort seemed to bear some fruit in 2023 with the successful development and tech transfer to Quanta of the AI-based electrocardiogram (ECG) analysis platform, enabling advanced diagnoses of multiple cardiovascular disease, especially for remote areas with few experienced medical personnel present (范瑜 2023).

  • Constructive Mixed Reality CPR and AED Training System

The Hualien Armed Forces General Hospital has developed a Mixed Reality training system for the instruction and training of cardiopulmonary resuscitation (CPR) techniques and the use of automated external defibrillator (AED). The system employs an AI-generated constructive simulation with a virtual patient and virtual AED, while allowing an on-site instructor and trainee to simultaneously interact with and be monitored by the simulation through linked cameras, receiving instantaneous feedback on accuracy of the technique and effects on the patient (國防部 2021a, b; 陳穎信 et al. 2020).

  • AI Performance Trend-based Engine Monitoring and Diagnostic System

In 2018 the Air Force Institute of Technology demonstrated to the press corps a predictive engine monitoring and diagnostic system. Based on the digitization of past engine maintenance records and designs for Taiwan’s next-generation jet engine for the Air Force’s future fighter program, they were able to construct a digital simulation of the engine to predict failure trends. The system won a gold medal and the special contribution award in Poland’s International Innovation Fair (徐振威 2018).

Perhaps the most damning testimony on the state of Taiwan armed forces’ fielding of defense AI comes from the original proponent of the Intelligent National Defense program, General Li Ting-sheng. In an op-ed published in August 2023, General Li, now retired, remarked that the Armed Forces’ application of AI to replace manpower in the spirit of asymmetric and innovative approach to defense has largely remained a matter of slogan, and rarely implemented (張延廷 2023).

6 Training for Defense AI

On top of establishing defense research centers, the NTD5bn academic research center initiative reviewed above also included provisions for training 150 graduate-level researchers for defense technology research. As early as 2016 defense AI training primarily focused on the employment of AI for computer network operations (CNO), setting up sections on information systems and network offense and defense under the department of information systems engineering of the National Defense University Chung Cheng Institute of Technology (國防大學理工學院 2023). Education on AI applications for defense logistics was introduced at Navy and Air Force Academies in 2019 largely through one man’s effort, former Navy engineer Dr. Wang Zhi-zhong and his company Xwin Prognostics Technology, in collaboration with Microsoft and Axiomtek (李恬郁 2019).

Systematic training on the conceptual employment of AI in the defense realm seems to be lacking, other than the occasional conferences and workshops, predominantly focused on the intersection between digitization, cyber, and AI, targeting senior-level officers (周昇煒 2019), with occasional forays into issue areas such as the application of AI in AR and VR training (中華民國陸軍 2020). However, there does not appear to be an overall initiative aimed at providing the force with a comprehensive understanding for the role defense AI may play in future conflicts.

The incorporation of AI into civilian defense education seems to have received a better hearing on the civil defense education side following the outbreak of the Ukraine war, where middle-school activities on defense AI included simulations of battlefield operations involving autonomous UAVs and ground vehicles (方志賢 2022).

Incorporation of defense AI into training scenarios for the armed forces also proves promising, specifically in using AI to assist analysis of big data accumulated in training to improve future training modalities. The most prominent example lies in the medical realm, specifically at the National Defense Medical Center. The center has been incorporating augmented reality (AR), mixed reality (MR) and smart glasses in the training of simulated tactical combat casualty care (TCCC), simulated mass casualty events, simulated patient care on med-evac vehicles, and simulated assessment and emergency care for CBRN events. Data collected from these training events are incorporated with questionnaires filled by trainees post-event and analyzed through statistical software such as SPSS before feeding into the database for processing by AI algorithms to improve future training events (楊策淳 et al. 2020).

Other experimental examples of AI-assisted training methodologies include the National Defense University Management College’s Military Human Factors Research Center, which began constructing databases of biometric data during the Marine Corps’ long endurance exercise in 2019 and has recently been experimenting with collecting biometric data from soldiers engaged in VR simulations mounted on top of a 6-axis motion control platform for analysis on how to improve future training regimes (蕭佳宜 2023).

7 Conclusion

Much of Taiwan’s approach to defense AI can be analogized as a microcosm of its approach to the defense of the nation, which can be summarized into three observations:

  • Observation 1: A propensity for “things done cheap” instead of “things done differently.”

A colloquial term often coined to describe Taiwan’s current industrial approach is the “cost-down” approach (Liu et al. 2012). From many of the projects examined in this chapter, we can see that the approach was not a comprehensive rethinking on how business can be done differently, but how business can be done “cheaply.”

  • Observation 2: A defense policy serving two masters—a government disconnected.

The legacy of an authoritarian government created a rather disjointed approach to defense policy, where higher-level strategic considerations are disconnected with the operational and tactical realities executed by the defense establishment. The contradictory approach to the definition of asymmetric warfare to defend the country is just one example.

  • Observation 3: a bottlenecked civil-military defense innovation ecosystem.

An obvious bottleneck, and potential explanation for Taiwan’s lack of fielded examples of defense AI lies with the NCSIST. Any development into an operational application that is deemed even remotely sensitive would eventually have to pass through its doors. Yet this is an institution plagued with corruption scandals (楊國文 2022), ineffective auditing mechanisms that result in significant delays in the delivery of major weapons systems, and an inability to prevent revolving doors from taking place between the institute and the defense establishment (楊丞彧 et al. 2023).

Would a revised, more comprehensive, yet top-down approach on defense innovation, focusing on a select few applications and approaches, with more clear communications to stakeholders, be a more profitable approach?

The answer may lie in a more grass-roots and less restrictive approach to the development and adaptation of defense AI, especially from the civilian realm of applications. There is certainly no shortage of talents from the civilian side, with the Minister of Digital Affairs Audrey Tang among TIME magazine’s 100 most influential AI figures of the year (Serhan 2023), and innovators winning awards on AI applications all over the world (張溎壕 2023). The challenge seems to be setting up the appropriate infrastructure where Taiwan’s defense establishment can properly benefit and harness from grassroots civilian efforts, without a preconceived notion of what the developmental pathway should be. To this end, a few policy recommendations may be beneficial in accelerating the process:

  • Stop leveraging defense innovation and industry for economic gains.

The government should recognize the urgency and priority of Taiwan’s defense needs and focus both the defense innovation system and defense industry at large on the overriding goal of ensuring Taiwan’s security against the threat posed by China. This would mean a comprehensive assessment of what sort of defense industry Taiwan needs to maximize its resilience and self-sufficiency under various Chinese coercive scenarios, instead of attempting to devote government subsidies and issue policy directives based on an unobtainable win-win scenario where defense innovation and industry can enhance the economy while fulfilling defense needs. The relevant economic trade-offs of such an assessment should be made clear to the public before implementation. It also means that the government needs to proactively leverage civilian industries to fulfill defense needs by going out to the civilian industries and leveraging existing government bureaucracies outside of the defense establishment to do so.

  • Flatten defense RDTE organization and security considerations.

The bottleneck represented by NCSIST and associated organizations within the defense industry development institutions must be alleviated. Consequently, a conscious, well-researched, well-informed trade-off between operational security and the ability of the defense establishment to benefit from civilian innovation must be made. Instead of a centralized and siloed approach based on clearance levels, and instead of issuing problems seeking solutions, a redesigned ecosystem should attempt to let the solutions be presented seeking potential applications from a democratized defense innovation ecosystem involving multiple operational stakeholders.

  • Establish an experimental unit for “roadshow.”

The previous two recommendations are really aimed at fostering an environment that would be more conducive to a bottom-up approach to defense innovation. But in order to spark such transformation, a small and nimble unit that can go around in a “roadshow” fashion, with both the authority and budget to experiment on solutions to various defense challenges and collect results for evaluation by both the defense and civilian establishment, would be crucial. The Ministry of National Defense Department of Integrate Assessment, a direct counterpart to the US Department of Defense’s Office of Cost Assessment and Program Evaluation (CAPE) office, is ideally placed to take advantage of this approach, should the relevant budgetary authority be granted.

  • Emphasize potentially profitable approaches for Taiwan’s defense AI.

AI-enhanced real-time translation using large language models, capable of enhancing joint training and operation between US and Taiwanese forces, would also be an initiative that leverages the asymmetric advantage between the US and its allies against the solitary nature of Chinese forces. Additionally, applications that can leverage Taiwan’s decades of consistent data collection of its surrounding operational environment, such as algorithmic optimization of decoys and smart sea mines, could also prove to be a profitable approach (Mitre et al. 2023).