Preamble

First Author: Jiyin Yang (Male, PhD candidate), primarily engaged in research on maritime navigational aids. Correspondence Address: No. 1 Lingshui Road, Dalian, Liaoning Province, China. Contact Number:

Application Prospects and Risk Challenges of Large Language Models in the Maritime Domain. School of Navigation, Dalian Maritime University, Liaoning.

摘要

Introduction

The emergence of Large Language Models (LLMs) represents a significant revolution in information technology, promising to breathe new life into various industries through their diverse applications. This paper focuses on the integration of LLMs within the maritime sector, exploring their potential prospects and the inherent risks and challenges they present. Our objective is to provide a reference framework for the intelligent and information-based transformation of maritime operations.

The Concept and Evolution of Large Language Models

Large Language Models are deep learning architectures trained on vast datasets to understand, generate, and manipulate human language with high proficiency. Their rapid evolution has transitioned from simple text prediction to complex reasoning and multi-tasking capabilities, setting the stage for specialized applications in niche domains such as maritime engineering and navigation.

Application Prospects in the Maritime Domain

The application of LLMs in the maritime field is extensive and multifaceted, offering transformative potential in several key areas:

Maritime Education and Training

LLMs can serve as intelligent tutors, providing personalized learning experiences for seafarers. By simulating complex scenarios and providing instant feedback, these models can enhance the efficacy of maritime education and help trainees master intricate navigation protocols and safety procedures.

Library and Documentation Management

The maritime industry generates vast amounts of technical documentation, regulations, and manuals. LLMs can significantly improve the efficiency of information retrieval and knowledge management by summarizing lengthy documents, categorizing technical data, and providing precise answers to queries regarding maritime laws and standards.

Assisted Ship Navigation

In the realm of navigation, LLMs can act as sophisticated decision-support systems. By processing real-time data and historical patterns, they can assist officers on watch in route planning, collision avoidance, and interpreting complex meteorological reports, thereby enhancing the safety and efficiency of maritime transit.

Reflection and Analysis of Maritime Accidents

LLMs can be utilized to analyze historical maritime accident reports to extract lessons learned and identify recurring risk factors. This automated synthesis of past failures allows for more robust safety management systems and the development of proactive measures to prevent future incidents.

Key Challenges and Research Directions

Despite their potential, the implementation of LLMs in the maritime sector faces several critical challenges that must be addressed:

• Medical Advice for Vessels: Providing accurate medical guidance in remote maritime environments requires high precision and reliability, as errors can have life-threatening consequences.
• Multimodal Data Integration: Maritime operations rely on diverse data sources, including radar images, sonar signals, and sensor data. Integrating this multimodal information with linguistic models remains a significant technical hurdle.
• Data Security and Privacy: Given the strategic importance of maritime logistics, ensuring high levels of data security and protecting sensitive information from cyber threats is paramount.
• Liability and Responsibility: Defining the scope of responsibility when an LLM-assisted system is involved in a maritime incident is a complex legal and ethical challenge that requires clear regulatory frameworks.
• Alignment with Practical Needs: Bridging the gap between theoretical model capabilities and the practical, rugged requirements of real-world maritime application scenarios is essential for successful deployment.

Conclusion

The rapid development of Large Language Models is poised to inject new vitality into the maritime industry. While the path toward full integration involves overcoming significant technical and regulatory obstacles, the potential benefits for maritime safety, education, and operational efficiency are profound. Future research should focus on addressing these challenges to ensure that LLMs contribute effectively to the modernization of the global shipping industry.

关键词

Large Language Models in the Field of Navigation: Application Prospects and Risk Challenges

Abstract: With the rapid development of artificial intelligence, Large Language Models (LLMs) have demonstrated exceptional capabilities in natural language processing, knowledge representation, and reasoning. This paper explores the application prospects of LLMs in the field of intelligent navigation, analyzing their potential roles in maritime communication, autonomous ship decision-making, maritime safety regulation, and crew training. Furthermore, the paper discusses the technical risks and practical challenges associated with deploying these models in maritime environments, providing a theoretical foundation for the integration of LLMs into the shipping industry.

Keywords: Large Language Models; Intelligent Navigation; Application Prospects; Risk Challenges

1. Introduction

The maritime industry is currently undergoing a significant digital and intelligent transformation. As the core of this evolution, intelligent navigation requires the processing of vast amounts of multi-modal data, including sensor inputs, maritime regulations, and complex communication protocols. Traditional rule-based systems and specialized machine learning models often struggle with the dynamic and unpredictable nature of the marine environment.

The emergence of Large Language Models (LLMs), such as the GPT series, has introduced a new paradigm for artificial intelligence. These models, trained on massive datasets, possess sophisticated linguistic understanding and logical reasoning capabilities. By integrating LLMs into the maritime sector, there is potential to bridge the gap between human-like decision-making and automated systems, ultimately enhancing the safety and efficiency of global shipping.

2. Application Prospects of LLMs in Intelligent Navigation

2.1 Maritime Communication and Information Interaction

Effective communication is the cornerstone of maritime safety. LLMs can serve as advanced intermediaries in maritime VHF (Very High Frequency) communications. By converting standardized maritime communication phrases into natural language and vice versa, LLMs can reduce misunderstandings between bridge teams of different nationalities. Furthermore, LLMs can automatically summarize Notice to Mariners (NtM), weather reports, and port information, providing officers with concise, actionable intelligence.

2.2 Decision Support for Autonomous Ships

Autonomous Surface Ships (MASS) require high-level situational awareness and decision-making capabilities. LLMs can be integrated into the decision-support architecture to interpret complex International Regulations for Preventing Collisions at Sea (COLREGs). Unlike traditional algorithms that may struggle with the "gray areas" of maritime law, LLMs can provide reasoned explanations for specific maneuvers, enhancing the transparency and reliability of autonomous navigation systems.

[FIGURE:1]

2.3 Maritime

(Navigation College, Dalian Maritime University, Dalian 116026 , China)

Abstract

advent large language models regarded information technology revolution, their application bring vitality industries. article focuses

application large models field navigation, presenting application prospects challenges large models navigation, aiming provide

references

intelligence informatization navigation field. Firstly, article introduces concept development trend large language models.

Secondly, explores broad application prospects large language models navigation, including navigation education training, library management, ship-assisted navigation, reflection

summary

maritime accident lessons, time, forward problems solved research large language models navigation, including providing medical advice ships, integrating multimodal data, meeting security requirements data, clarifying scope responsibility, difficulty connecting actual navigation needs application scenarios. rapid development large language models surely inject vitality field navigation. words Large Language Model (LLM) Intelligent navigation application prospect challenge Artificial General Intelligence (AGI)

21 世纪以来，人工智能技术取得了长足的发展。其中，深度学习方法在各个传统领域

Significant improvements in algorithmic performance, particularly in Natural Language Processing (NLP), have been achieved recently. OpenAI’s ChatGPT, a natural language processing tool based on the Large Language Model (LLM) architecture known as the Generative Pre-trained Transformer, has ignited a new wave of artificial intelligence. Compared to previous chatbots, ChatGPT demonstrates superior language understanding and text generation capabilities, greatly broadening its range of applications \cite{2,3}. In various scenarios such as text generation, knowledge retrieval, and polishing or translation, ChatGPT has reached a professional level of utility. Following the success of ChatGPT, other teams globally have explored different implementation methods and application domains for LLMs, developing chatbots such as Huawei’s Pangu series, ChatGLM, Baidu’s ERNIE Bot, Alibaba’s Tongyi Qianwen, and iFLYTEK’s Spark. These teams have delved into vertical fields such as office automation and healthcare, discovering even greater potential for LLMs. Large language models are already being widely applied across numerous industries. Seven departments, including the Ministry of Transport, jointly released the "Guiding Opinions on the Development of Intelligent Shipping." Intelligent shipping represents a modern shipping paradigm formed by the deep integration of traditional shipping elements with high-tech innovations such as modern information technology, communications, sensing, and artificial intelligence. To realize intelligent shipping, LLMs are an indispensable key technology. This paper focuses on the application of large models in the maritime industry across several scenarios, including maritime education and training, library and document management, and maritime collision avoidance decision support. This paper also emphasizes the risks and challenges brought by large models as a new technology, aiming to provide a reference for maritime peers researching the application of large models in the maritime field.

The Development History of Large Language Models: Large language models are considered one of the key technologies for future Artificial General Intelligence (AGI). A Language Model (LM) is a technique used to model the generation probability of word sequences. Its primary goal is to estimate the probability distribution of the next word or character given a specific context. Language models can measure the linguistic plausibility of a sentence and are applied to various Natural Language Processing (NLP) tasks. Language models play an important role in the following two application examples:

Natural Language Generation Evaluation: In natural language generation tasks, such as machine translation and text summarization, language models can be used to determine whether generated sentences conform to human linguistic patterns. A language model calculates the probability of a sentence; if a generated sentence has a high probability within the language model, it is considered more consistent with human linguistic habits, thereby improving the quality of the natural language generation system.

Corpus Statistical Analysis: Language models can also be used to analyze large-scale corpora. By performing statistical modeling on a corpus, language models can learn the relationships between words and contextual information. These statistical methods help us understand the patterns and characteristics of a language, subsequently providing linguistic knowledge for other natural language understanding models. We can use language models to calculate word co-occurrence frequencies and predict the next possible word, which assists in tasks such as information retrieval and text classification. Language models can be categorized into statistical language models and neural network language models. Statistical language models use n-gram statistical methods to improve performance in tasks like retrieval. Neural network language models can effectively utilize distributed word vectors to model contextual relationships within language.

Large-scale Pre-trained Language Models: In 2017, Google introduced a feature extractor called the Transformer, which adopted a self-attention mechanism. This innovation replaced the Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) previously common in language model research. The Transformer architecture improved natural language understanding capabilities and enabled a deeper representation of semantic information. By employing self-attention and a stacked encoder-decoder structure, this architecture significantly outperformed deep learning models like CNNs in semantic and task feature extraction. The advantage of this architecture lies in its ability to simultaneously consider global information and local dependencies within the text, thereby better capturing context and semantic connections.

The Transformer model consists of an encoder and a decoder; the encoder is used to encode input information, while the decoder decodes the encoded information into an output result. Compared to traditional methods, the Transformer model can focus on contextual information simultaneously and is more efficient during training. Based on the innovative ideas of the Transformer, a series of methods have been derived, such as Bidirectional Encoder Representations from Transformers (BERT) and the Text-to-Text Transfer Transformer (T5). BERT introduces noise to decompose the original training text, expanding the spatial range of model inference. Meanwhile, T5, as a pre-trained language model based on the Transformer, converts various problems into a text-to-text format.

There are two main technical paradigms for Transformer-based language models. Beyond ChatGPT, a series of new models using only partial structures of the Transformer have emerged, including encoder-based and decoder-based models. The development of these models has brought flourishing opportunities to the field of large-scale pre-trained language models. Their relationships are illustrated in [FIGURE:1]. Trends in Large Language Models: LLMs have begun to integrate multiple data types, including text, video, and images, gradually evolving into complex multimodal models. Consequently, they exhibit higher levels of intelligence.

Large language models initially improved performance by increasing model scale but are gradually shifting toward a balance between scale and economic cost. The focus of research is moving from model architecture to model application. LLMs are no longer confined to the foundational research stage; instead, there is active development of model ecosystems and expansion into application scenarios such as government affairs.

Applications of Large Language Models in the Maritime Domain: In Maritime Education and Training, ChatGPT can generate instructional content such as lecture notes, textbooks, and case studies to meet educational needs \cite{14,15}. It can also provide personalized learning resources based on students' needs and knowledge levels, helping them better understand and master maritime knowledge. Furthermore, it can automatically generate maritime examination questions, reducing the burden on instructors while ensuring the quality and accuracy of the assessments.

At the same time, it can recommend relevant learning resources based on students' learning progress and preferences, achieving personalized learning path recommendations. Integrating ChatGPT with technologies like Augmented Reality (AR) can drive innovation in maritime education. Students can simulate actual maritime operations and situational handling through virtual scenarios, interacting with virtual crew members for real-time maritime practice and decision-making drills.

In this process, ChatGPT acts as a virtual tutor for students, providing guidance, feedback, and answers to questions, thereby enhancing learning outcomes and practical abilities.

Maritime education involves a wide range of knowledge domains. Utilizing large language models can help students quickly acquire and understand complex maritime concepts. Through natural language processing technology, LLMs generate teaching materials and provide personalized learning, making maritime education more individualized and efficient. They also assist maritime instructors in teaching assessments and identifying key knowledge points. By analyzing students' learning status and knowledge levels, instructors can better guide students and improve the quality of maritime instruction. Regarding Ship Library and Document Management: Maritime literature is voluminous and complex. Effectively managing these documents and finding specific knowledge within extensive books is a challenge. For instance, finding the operational steps for a specific function in a maritime instrument manual can be simplified with ChatGPT. We can input the manual into the LLM and obtain knowledge by asking questions, which is both simple and fast.

Furthermore, ChatGPT can be utilized for the automated reading and knowledge extraction of maritime literature. It can read maritime documents and generate concise summaries, outlining the main content to help researchers quickly grasp key points. Based on user queries, ChatGPT can extract relevant knowledge from maritime literature and provide answers, assisting researchers in quickly locating and utilizing information. Large language models possess significant advantages in processing massive amounts of maritime literature and voyage reports. They can efficiently perform text summarization, entity recognition, and relationship extraction, helping maritime researchers quickly extract valuable information from large volumes of text. Such automated text processing methods accelerate the accumulation and exchange of maritime knowledge. These models also have potential advantages in constructing knowledge graphs and updating maritime knowledge bases. By processing multi-source maritime data, LLMs can help build more comprehensive and accurate knowledge graphs and databases, promoting deeper and more refined maritime research.

Ship-Assisted Navigation: ChatGPT and other LLMs can provide decision-making suggestions for ship navigation. Maritime navigation is highly complex, requiring captains to formulate personalized maneuvering and decision-making plans based on different situations. LLMs possess powerful prediction and classification capabilities. By integrating specific ship information with extensive navigation data, they can provide customized navigation decision support for captains, thereby improving operational efficiency. [FIGURE:2] shows a case where the author used ChatGPT to obtain collision avoidance decision-making suggestions.

Case Study of ChatGPT for Maritime Collision Avoidance Decision-Making: "Suppose you are a maritime expert. I will provide a scenario; please provide me with collision avoidance decisions. In a head-on situation, the primary goal is to ensure navigational safety. Based on the International Regulations for Preventing Collisions at Sea (COLREGs) and general maritime principles, I suggest the following collision avoidance decisions:"

1. Maintain Vigilance: Ensure all personnel on board are aware of the head-on situation and are prepared to take necessary action.
2. Determine the Situation: Assess the distance, speed, and heading of the vessel on the port side. Observe whether it is approaching, moving away, or maintaining a relatively stable position.

3. 判断优势：如果左侧船只在你的左前方（约为

If a vessel is located on your port side at the 10 o'clock position or earlier, according to the International Regulations for Preventing Collisions at Sea (COLREGs), that vessel has the responsibility to avoid your ship. However, if the vessel on your port side is located at your port quarter (8 o'clock or later), your ship may need to take action to avoid it. 4. Taking Action: Based on the assessed priority, select the appropriate maneuver to avoid a collision. Possible actions are listed below:

If the vessel on your port side is located off your port bow, you may maintain your course and speed, allowing the other vessel to pass around you. If the vessel is located off your port quarter, you may alter your course or speed to allow it to pass safely along your port side.

In critical situations, you should use onboard communication equipment to contact the other vessel and negotiate the optimal collision avoidance plan. 5. Monitor Dynamics: Continue to observe the actions and responses of the vessel on the port side. Ensure that your maneuvers do not create other hazardous situations.

Please note that the above suggestions are for reference only and may require adjustments based on specific circumstances. At all times, maritime safety must be the primary consideration, and strict adherence to COLREGs as well as applicable national and regional maritime regulations is essential.

In this scenario, your vessel is positioned directly behind the other vessel and is in the process of overtaking it. To avoid a collision and in accordance with international maritime standards, assume you are a maritime expert; I am providing a scenario for which you must provide collision avoidance decisions, including the display of navigation signals.

Explain how our vessel should act when we are positioned directly astern of another vessel and are currently overtaking it. Assuming you are a maritime expert, evaluate the following collision avoidance case according to the COLREGs and clarify the responsibilities of each party. Note that the SANCHI used UTC+9 time, while the CF CRYSTAL used UTC+8 (Beijing Time). The collision occurred at approximately 1950 LT (Beijing Time) in the northern East China Sea, about 160 nautical miles east of the Yangtze River estuary.

The SANCHI remained on fire until it exploded and sank at 1645 LT (Beijing Time) at coordinates $28^{\circ} 21' 48''$ N, $125^{\circ} 57' 55''$ E.

The SANCHI was heading north toward Korea at a speed of 10 knots; to its northeast, the CF CRYSTAL was heading toward the Taiwan Strait. The two vessels were in a crossing situation. According to the COLREGs, the CF CRYSTAL was the stand-on vessel, and the SANCHI was the give-way vessel. The SANCHI first detected the CF CRYSTAL on radar at 2024 LT (UTC+9 / 1124 UTC) and found the Closest Point of Approach (CPA) to be negative, indicating that the CF CRYSTAL would pass astern of the SANCHI.

The CF CRYSTAL was operating on UTC+8 time, and the Chief Officer was on watch at the time. Visibility was good; however, according to the crew, they did not detect the SANCHI. Based on the provided information and COLREGs, we can determine the responsibilities of the parties as follows: based on the heading, speed, and position of the vessels, the SANCHI was the give-way vessel and should have taken the necessary actions to avoid the collision.

When the SANCHI first noticed the CF CRYSTAL on radar, it identified a negative CPA, suggesting the CF CRYSTAL would pass astern. Between 2025 LT (1125 UTC) and 2044 LT (1144 UTC), the Third Officer of the SANCHI was preoccupied with avoiding Chinese fishing vessels and failed to notice the course adjustments and approach of the CF CRYSTAL.

The CF CRYSTAL was the stand-on vessel. In this situation, it had the right of way and was required to maintain its course and speed unless action was necessary to avoid an immediate danger or if the give-way vessel failed to take appropriate action.

The vessel followed its planned route, but at 1945 LT (1145 UTC), it failed to avoid the SANCHI. Based on the International Regulations for Preventing Collisions at Sea (COLREGs) and general maritime principles, I recommend the following maneuvers:

3. 给予足够空间：确保你船只与对方船只之间保持足够的安全距离，以避免任何潜在的碰撞风险。根

In accordance with the requirements of the International Regulations for Preventing Collisions at Sea (COLREGs), you must make every effort to avoid approaching other vessels and maintain a safe distance at all times.

4. 通知对方船只：如果你认为对方船只未能察觉你的存在或者没有采取适当的行动，你可以使用船上的呼

Use communication equipment (such as loudspeakers or radio) to communicate with the other party, alerting them to your presence and intentions. 5. Safe Overtaking: If you are certain that you can safely overtake the other vessel, you may take the following actions:

Confirm the reaction of the other vessel before overtaking and ensure there is sufficient space for passage. Determine that no other vessels are in your navigation channel and ensure that the maneuver will not interfere with the navigational safety of other ships.

Ensure that you maintain a stable course and speed after overtaking the other vessel. The above suggestions are for reference only, and specific situations may require adjustments based on the actual environment and regulations. Maritime safety is of paramount importance; please adhere to COLREGs as well as applicable national and regional maritime regulations.

In the context of maritime accident reconstruction and lesson summarization, ChatGPT, combined with speech recognition technology, can interact with crew members in a conversational format to obtain detailed information regarding the circumstances of a ship accident. Simultaneously, large language models (LLMs) can assist in tasks such as liability determination based on specific on-site conditions. Following a maritime accident, ChatGPT can converse with the crew, inquiring about personnel status and the traffic conditions encountered. It can collect, classify, and integrate accident information from multiple sources faster than humans. This allows regulatory authorities to gain a more comprehensive understanding of the accident, providing a foundation for subsequent summaries of lessons learned and the allocation of responsibility.

To the northeast of the SANCHI, the heading was southwest, to the left of the planned route. At 1943 LT (1143 UTC), to bring the ship's position closer to the planned route, the Chief Officer adjusted the course from 1945 LT (1145 UTC) and handed over the watch to the Third Officer. At 1949 LT (1149 UTC), when the Third Officer saw the approaching vessel on the radar at a distance of [X] nautical miles, he ordered right rudder, followed by the collision. The SANCHI used UTC+9 time. From 2025 LT (1125 UTC) to 2044 LT (1144 UTC), the Third Officer was preoccupied with avoiding a Chinese fishing vessel crossing from right to left that posed a collision risk; consequently, he did not notice the course adjustment made by the Chief Officer of the CRYSTAL. By 2044 LT (1144 UTC), the Third Officer of the SANCHI noticed the distance between the two vessels. He only ordered the helmsman on watch to signal the CRYSTAL with a light and took no further measures, hoping at that time that the other vessel would give way. It was not until 2048 LT (1148 UTC) that the Third Officer realized the danger and urgently called the Captain to the bridge. At 2050 LT (1150 UTC), the bow of the CF CRYSTAL struck the starboard mid-forward section of the SANCHI.

In summary, in this collision case, the SANCHI bears primary responsibility as it was the give-way vessel and failed to take timely and appropriate action to avoid the collision. The CRYSTAL also bears a certain degree of responsibility during the navigation process for failing to avoid the SANCHI. Please note that the above judgment is based solely on the information provided and general collision avoidance rules; the final determination of liability requires a detailed investigation and adherence to relevant regulations.

Providing medical advice to crew members is another application, as research into LLMs in the medical field is advancing rapidly. General medical models such as "Mingyi" can provide professional medical advice. Furthermore, there are LLMs specialized in vertical medical fields, such as MedChat, which focuses on the treatment of vertigo, and these models are also developing quickly.

LLMs can act as professional doctors to provide crew members with support for physical and mental health. For questions crew members may have regarding illnesses, treatments, and drug side effects, ChatGPT can answer and explain the meaning of medical terminology. This enables them to use medication accurately and rescue injured personnel in a timely manner when professional medical support is unavailable. ChatGPT provides health education and rehabilitation guidance to crew members, helping them better manage the recovery process and accelerating their return to health. In the field of maritime law, LLMs can combine maritime precedents with maritime law to analyze cases, organize case clues, and provide decision support for judicial rulings.

Regarding communication message comprehension, LLMs are not limited to text-based tasks. Through multimodal development, LLMs can be integrated with speech recognition tasks. By inputting VHF maritime communication audio, the model can organize the maritime situation, helping navigators enhance situational awareness and automatically generating VHF logs to provide further support for voyage records.

In maritime environmental perception, combined with image recognition technology, LLMs can analyze information such as sea waves, obstacles, and vessels to organize maritime traffic conditions and predict collision risks. Simultaneously, by incorporating the International Regulations for Preventing Collisions at Sea (COLREGs), they can provide reasonable navigational decisions. By integrating historical ship data and existing weather forecasts into text inputs for the model, the LLM can perform reasoning based on databases and historical routes to provide optimized navigation paths.

Issues to be addressed for maritime LLMs include the enrichment of multimodal data. Humans use multiple senses to receive multimodal information, which allows for the effective cognition and processing of various tasks; the same applies to LLMs. Currently, LLMs primarily rely on textual information, but multimodal input can supplement knowledge and enable more complex tasks.

While the application of LLMs in the maritime sector currently focuses on text-processing scenarios, maritime data is diverse.

Maritime data includes not only text but also various types such as video, images, and voice. Products like KOSMOS-1 and domestic alternatives support the processing of multimodal data. The maritime sector needs to conduct in-depth research into the application of LLMs for processing video, image, and voice data types.

Meeting high data security requirements is essential. Since maritime data involves personal privacy and commercial secrets, data privacy and security issues must be considered when applying maritime LLMs. These issues include data leaks, unauthorized access, and data misuse. Data contains sensitive information such as ship call signs and trajectory information; therefore, effective measures must be taken to ensure data security. Solutions include using data encryption technologies to protect data during transmission and storage, and establishing strict data access and permission control mechanisms to ensure only authorized personnel can access sensitive data. It is necessary to formulate appropriate data sharing policies and contracts, clarifying the purpose and limitations of data use in alignment with relevant laws and regulations. Furthermore, LLMs may generate large amounts of convincing but unsourced content in a short time, making responses appear highly credible while lacking a reliable basis. This can be extremely misleading for users with lower judgment capabilities. Therefore, it is also necessary to strengthen the monitoring and auditing of data security to promptly discover and respond to security vulnerabilities and threats.

Defining responsibility boundaries is critical. In ship management, the Captain, as the primary decision-maker, must take responsibility for the decisions made. When a Captain uses an LLM, they may face situations where responsibility is difficult to define. First, the output of an LLM is unpredictable, making it difficult for the Captain to make reasonable decisions based on that output. LLMs are based on existing data; essentially, they learn from existing knowledge and perform reasoning, making it difficult for them to make rational judgments in business scenarios such as emergency handling and future trend prediction in the maritime field. Ship decision-making is based on maximizing the ship's interests. LLMs represent "machine will," and whether they can reflect the Captain's will remains to be verified. According to the Interim Measures for the Management of Generative Artificial Intelligence Services, the accuracy and reliability of generated content must be improved, and effective measures should be taken to increase the transparency of generative AI services. In the practical use of LLMs, it is necessary to clarify responsibility boundaries, reasonably define the alignment between LLM-made decisions and human decisions, correctly distinguish between machine will and the Captain's will, and strive to ensure that LLMs can make decisions that align with the public interest of the vessel.

Researching "human-in-the-loop" decision-making processes is a sound approach to ensure that final decisions are made by humans, thereby avoiding errors and information discrepancies caused by hallucinations.

Disconnection from actual maritime needs and application scenarios can make implementation difficult. The maritime field is highly complex and professional, requiring deep maritime knowledge and professional judgment. While LLMs may perform excellently in general language understanding, their understanding of specific maritime knowledge and professionalism remains limited. This makes it difficult to apply LLMs for decision support and on-site guidance in maritime scenarios. Specific experiments can refer to [TABLE:N]. LLMs represented by ChatGPT and DeepSeek have application potential in many maritime-related fields, such as maritime education, maritime library and information management, auxiliary navigation, and maritime accident lesson summarization. They possess the capability to reshape the service models of the maritime industry. Although they face many challenges, including technical and social issues, the advent of LLMs has initiated a major transformation in the maritime sector. Only by following the trends of technological development can we seize the momentum and achieve success.

Large Language Models and Their Applications in the Field of Government Affairs. Journal of Tsinghua University, 2024, 64(04):

Artificial Intelligence Discusses Artificial Intelligence Translational Medicine[J].

JACC: Basic Translational Science, 2023, 8(2): comprehensive survey pretrained foundation models: history Chatgpt[J].

International Journal Machine Learning Cybernetics, 2024:

Survey Transformers[J]. CoRR, 2021, abs/2106.04554.

Dehghani Bahri Efficient Transformers: Survey[J]. CoRR, 2020, abs/2009.06732.

Radford Narasimhan Salimans Improving language understanding

generative Pre-Training[J]. 2018. San Francisco, CA, USA, 2018.

Radford Child Language models unsupervised multitask Learners[J].

OpenAI Blog, 2019, 1(8): Application Testing and Discussion of Large Language Models in Architectural Engineering 2023, 53(09):

Technical Application Prospects and Risk Challenges of Large Language Models. Journal of Computer Applications 2024, 44(06): A Review of the Application of Large Language Models Represented by ChatGPT in Clinical Medicine. Journal of Medical Informatics 2023, 44(07):

Research and Application Development of Large Language Models in the Medical Field 2023(04): Vaswani A, Shazeer N, Parmar N, et al. Attention is All You Need[C]//Proceedings of the International Conference on Neural Information Processing Systems.

Hook, Curran Associates Inc., 2017: Raffel Shazeer Roberts Exploring limits transfer learning unified Text-to-text Transformer[J].

Mach. Learn. Res., 2020, 21(1). doo-Anu Ansah Education Generative Artificial Intelligence (AI):

Understanding Potential Benefits ChatGPT Promoting Teaching Learning[J].

Journal date], 7(1): Opara Mfon-Ette Theresa Aduke ChatGPT teaching, learning research:

Prospects Challenges[J]. Opara Emmanuel Chinonso, Adalikwu Mfon-Ette Theresa,

Tolorunleke Caroline Aduke (2023). ChatGPT Teaching, Learning Research:

Prospects Challenges. Humanit 2023, Chatgpt changer Education[J].

Zhai, X.(2023). ChatGPT: Reforming Education Aspects.

Shanghai Education, 2023: Talan Kalinkara Artificial Intelligence Higher Education:

ChatGPT Assessment Anatomy Course[J]. Uluslararas netim Sistemleri Bilgisayar Bilimleri Dergisi, 2023, 7(1):

Paranjape Schinkel Panday Introducing Artificial Intelligence Training Medical Education[J].

Medical Education, 2019, 5(2): e16048. Seetharaman Revolutionizing Medical Education:

ChatGPT Boost Subjective

Learning and Expression?[J]. Journal of Medical Systems, 2023, 47(1): 61.

Munaf Ul-Haque future medical education research: ChatGPT blessing blight Disguise?[J].

Medical Education Online, 2023, 28(1): Vaishya Misra Vaish ChatGPT: version healthcare Research?[J].

Diabetes Metabolic Syndrome: Clinical Research Reviews, 2023, 17(4):

Howard Markov Comparing scientific

Abstract

generated ChatGPT original

Abstract

using artificial intelligence output detector, plagiarism detector, blinded human Reviewers[M].

Manning Human Language Understanding Reasoning[J].

Daedalus, 2022, Sallam utility ChatGPT example large language models healthcare education, research practice:

Systematic review future perspectives potential limitations[J].

MedRxiv, 2023: Javaid Haleem Singh ChatGPT healthcare services: emerging stage innovative Perspective[J].

BenchCouncil Transactions Benchmarks, Standards Evaluations, 2023, 3(1):

Zapadka Translating radiology reports plain language using ChatGPT GPT-4 prompt learning:

Results, limitations, Potential[J]. Visual Computing Industry, Biomedicine, 2023, 6(1):

Sallam Salim Barakat ChatGPT applications medical, dental, pharmacy, public health education: descriptive study highlighting advantages Limitations[J].

Narra 2023, 3(1): e103. Cheng potential impact ChatGPT/GPT-4 surgery: topple profession Surgeons?[J].

International Journal Surgery, 2023, 109(5): Singhal Azizi Large language models encode clinical Knowledge[J].

Nature, 2023, 620(7972): lsoylu Yildiz Image Fusion Technique Maritime Computer Vision Applications[C]//Proceedings IEEE/CVF Winter Conference

Applications of Computer Vision (WACV) Workshops. 2024: 859 – 868.