Available online at www.sciencedirect.

com

ScienceDirect
Procedia CIRP 119 (2023) 7–14

33rd CIRP Design Conference

ChatGPT for design, manufacturing, and education

Xingzhi Wanga*, Nabil Anwerb, Yun Daic, Ang Liua
a
School of Mechanical and Manufacturing Engineering, UNSW Sydney, NSW 2052, Australia
b
Université Paris-Salay, ENS Paris Saclay, LURPA, 4 Av. des Sciences, 91190 Gif-sur-Yvette, France
c
Department of Curriculum and Instruction, Chinese University of Hong Kong, Hong Kong SAR, China

* Corresponding author. Tel.: +61 4 1125 6306; fax: +61 4 1125 6306. E-mail address: [email protected]

Abstract

The manufacturing industry involves innumerable complex tasks that require significant knowledge and experience to execute. With the rapid
development of artificial intelligence, particularly with the emergence of powerful large language models such as ChatGPT, new opportunities
have risen to provide knowledge through conversation. With its seemingly endless knowledge base and highly organized response style, ChatGPT
is expected to revolutionize every aspect of the industry. However, the extent of ChatGPT’s capabilities and how they could contribute to the
industry’s future revolution remains unclear. In light of this, this paper performed a systematic testing of ChatGPT to uncover its advantages and
limitations. Based on the testing results, the authors provided some prospects and critical research questions of ChatGPT from a manufacturing
perspective. Furthermore, the authors recommended a technology development roadmap to successfully integrate ChatGPT into the
manufacturing industry.
© 2023 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer review under the responsibility of the scientific committee of the 33rd CIRP Design Conference
Keywords: Artificial Intelligence, Engineering Design, Product Development, Smart Manufacturing, ChatGPT, AI-generated content, Education

1. Introduction
Nomenclature
ChatGPT is a large language model (LLM) that can generate
human-like text responses to user queries based on knowledge AI Artificial Intelligence
gained from massive datasets. It has gained significant attention LLM Large Language Models
from researchers across various domains due to its exceptional KBS Knowledge-Based Systems
performance for complex text generation (e.g., document COT Chain of Thought
writing, code debugging, question-answering) and knowledge
acquisition. Recently, by answering prepared questions,
ChatGPT has passed professional exams in law, business, and In the context of the manufacturing industry, ChatGPT is
intuitively associated with knowledge-based systems (KBS), as
medical licenses. This breakthrough has sparked people's
imagination. On the one hand, some believe that ChatGPT will both systems share similar components, as shown in Figure 1.
boost human creativity and productivity. On the other hand, These include an interface for user queries, an inference engine
others express concerns about ChatGPT’s potential impact on to interpret user queries and generate solutions, and a
the labor market. This controversy highlights the need to knowledge base to store and supply knowledge [2].
rethink how humans and artificial intelligence (AI) can Considering KBS has been extensively used to support various
collaborate in this rapidly changing society [1]. design and manufacturing tasks or serve as learning system for
engineering education and industrial training [3-5]. Thus,

2212-8271 © 2023 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer review under the responsibility of the scientific committee of the 33rd CIRP Design Conference
10.1016/j.procir.2023.04.001
8 Xingzhi Wang et al. / Procedia CIRP 119 (2023) 7–14

ChatGPT is expected to have a similar role in the industry. technical design is the process of transforming design
However, ChatGPT outperforms KBS in two crucial aspects. parameters into process variables. The manufacturing topics
Firstly, the knowledge base of ChatGPT is seemingly limitless, include manufacturing processes such as casting, machining,
allowing users to ask almost any questions they want to know. assembly, etc. Manufacturing systems encompass all
KBS, on the other hand, typically only stores knowledge from equipment, people, information, and other system-level
specialized domains. Second, ChatGPT can generate long-form functions. And production management includes the activities
and highly organized responses, while KBS responses tend to across various manufacturing stakeholders such as production
be less detailed. In this regard, ChatGPT has the potential to go planning, supply chain management, quality control, etc.
beyond the capabilities of KBS and act as an AI expert to assist
with various design, manufacturing, and engineering education
tasks.

Figure 1. The components of ChatGPT based on KBS

However, the manufacturing industry currently has limited

knowledge about the capabilities and limitations of ChatGPT,
and its potential for industrial applications remains unclear. To
address this gap, this paper aims to answer the following
questions:
(1) Can ChatGPT replace industrial KBS, and how close is
it to human experts?
(2) What are the potential roles of ChatGPT in design and Figure 2. Testing methodology
manufacturing, and how they can be achieved in practice?
The rest of this paper is organized as follows. Section 2 The performance of ChatGPT is evaluated by answering 100
elaborates on the methodology to test ChatGPT’s performance. selected questions from course materials (e.g., textbooks,
Section 3 summarizes the advantages and limitations of quizzes, student projects) and industrial documents (e.g., ISO
ChatGPT from a manufacturing perspective. Section 4 standards, SOP documents, case studies). Besides, by
proposes a roadmap to guide the manufacturing industry in modifying some standard questions, several trap questions are
integrating ChatGPT with existing manufacturing systems. also generated to test its cognitive flexibility. Based on the
Some key research questions are also recommended. Section 5 revised Bloom's taxonomy [6], questions are categorized into
concludes the article and highlights some future research four difficulty levels:
directions for advancing the use of ChatGPT in industrial Level 1 - Basic knowledge: Understand and explain design
applications. and manufacturing concepts.
Level 2 - Application of knowledge: Choose, employ,
2. Testing methodology illustrate, and demonstrate concepts to solve specific problems
properly.
Figure 2 shows the methodology used to test ChatGPT by Level 3 - Critical analysis: Break down information into
answering prepared questions in six design and manufacturing components to understand its nature or determine essential
topics. Design topics cover functional, conceptual, and features.
technical design. Functional design is a process that converts Level 4 - Extension of knowledge: Synthesis of new
ill-defined customer needs into well-defined functional knowledge, making inferences, transferring knowledge to
requirements. Conceptual design is the process of transforming usefulness in new areas.
functional requirements into a set of design parameters. And
 Xingzhi Wang et al. / Procedia CIRP 119 (2023) 7–14 9

The answers generated by ChatGPT are evaluated against Based on the assessment results, four critical conclusions can
four qualitative characteristics: correctness, relevance, clarity, be made to describe ChatGPT’s capabilities.
and comparability [7]. Among them, correctness and relevance
are two basic requirements. Clarity and comparability are 3.1. ChatGPT can explain the most well-known knowledge in
enhancing requirements. If ChatGPT fails to meet basic an organized and detailed manner, but it may not always be
requirements, its performance is considered low. Low accurate in its explanations.
performance indicates ChatGPT is still not ready to assist with
relevant human work. If ChatGPT meets all four requirements, ChatGPT utilizes a chain-of-thought (COT) approach to
its performance is considered high, meaning its performance is enhance the reasoning of its responses, as depicted in Figure 4.
comparable to that of humans. Otherwise, its performance is With COT, ChatGPT can generate a series of intermediary
considered medium, meaning it achieved acceptable steps to lead up to its final answer, or it can provide detailed
performance but still has room for improvement. rationales and justifications for its responses. Additionally,
ChatGPT can highlight the limitations of its answers to alert
3. Results and discussions human experts. As a result, compared to KBS, ChatGPT’s
answers are more thorough and detailed.

Figure 4. An example of COT

However, ChatGPT may not always provide accurate

knowledge explanations. Firstly, ChatGPT may give answers
that are irrelevant to the manufacturing industry. For example,
when we asked ChatGPT about the process of developing
product specifications, its answer was more relevant to
software products than to manufactured products. Secondly,
Figure 3. Overview of testing results ChatGPT could confidently assert false information. For
example, in one question shown in Figure 5, ChatGPT provided
Figure 3 shows the overview of the testing results. Given an incorrect answer, although the correct answer was easily
questions across four categories in six topics, ChatGPT retrievable through search engines.
achieved high performance in 16 out of 24 categories and Our tests indicate that ChatGPT was not specifically trained
achieved medium performance in 6 categories. Based on the with technical knowledge from the manufacturing industry. If
difficulty level of questions, ChatGPT had an extraordinary considering testing questions within the social-technical
performance in answering extension of knowledge questions as paradigm, ChatGPT can achieve higher accuracy when
it achieved high performance in all 6 topics. It also performed questions are closer to the social end, where manufacturers
well in answering basic knowledge and knowledge application have to face uncertainties of customers, markets, new
questions, although it was unable to achieve high performance technologies, and so on. However, its accuracy drops
consistently across all topics. However, the performance of significantly when it comes to specialized questions on the
ChatGPT was notably low when answering critical analysis technical end, such as detailed material processing techniques,
questions. In terms of question topics, ChatGPT performed machining parameter settings, and so on. One possible
well in functional design, conceptual design, and explanation for the problem is that ChatGPT was trained using
manufacturing process questions. However, its performance crowdsourced online data. Due to the nature of the data source,
decreased significantly when facing technical design, it may be inherently difficult to obtain high-quality design and
manufacturing system, and production management questions. manufacturing knowledge.
10 Xingzhi Wang et al. / Procedia CIRP 119 (2023) 7–14

engineers must figure out how to pose questions to best obtain

the knowledge they want. Furthermore, engineers should
possess certain verification skills, or else they may be misled
by erroneous information, potentially leading to flaws in their
design and manufacturing practices.

3.3. ChatGPT has some creativity and doesn't have many

cognitive limitations, but its creativity may not exceed
experienced engineers.

Figure 5. An example of obviously incorrect answer The creativity of ChatGPT was evaluated by answering
extension of knowledge questions, which are often open-ended
3.2. ChatGPT could fabricate facts in order to give an and lack definitive answers. In such cases, we placed greater
answer, which is related to user prompts. emphasis on the relevance, clarity, and comparability of
responses. From the logical reasoning aspect, creativity is
As a probabilistic language model, ChatGPT's primary regarded as the process of producing new ideas through
focus is to generate coherent and natural-sounding text by abduction [9]. To test the creativity of ChatGPT, we
leveraging statistical associations between words in training decomposed abduction into two sequential questions: analytic
data. In that sense, although it can generate highly organized proposition questions, and synthetic proposition questions [10].
and detailed text, it doesn't have deep understanding to the The analytic proposition questions evaluate whether ChatGPT
related concepts [8]. Our test reveals that a significant can decompose a general subject into smaller predicate using
drawback of ChatGPT is its tendency to fabricate facts when “knowing-that” knowledge. Synthetic proposition questions
faced with questions without an existing answer. To provide an test whether ChatGPT can generate predicate of subject based
answer, ChatGPT may create non-existent and sounding on “know-how” knowledge. For instance, we posed a design
plausible evidence, axiom, equations, and rules. This is very concept ideation question in Figure 7, asking ChatGPT to
different from KBS. In most of the current KBS, knowledge is generate design concept based on the functional requirement
encoded in complex ontologies, rules, logic, or cases. “enable human fly”. We asked ChatGPT to try functional
Consequently, any answers provided by KBS are referenced decomposition as well as decomposition based on user
with factual evidence in the knowledge base and can be traced interactions. The testing results show that ChatGPT can
back to the source for verification. Therefore, compared to creatively synthesize concepts in many exciting ways. In
KBS, answers from ChatGPT are not easily verifiable and may addition, based on the authors' years of design education
not always be reliable. There are two possible causes for experience, the performance of ChatGPT is comparable to the
ChatGPT generating false knowledge. Firstly, as a generative best students in the class.
model, the nature of ChatGPT is to predict the most likely next
word based on the input it receives, rather than focusing on
logical consistency and factual correctness. Secondly, the data
used to train ChatGPT may already contain false knowledge,
which can propagate to the model and lead to fabricated facts.
Our testing also shows that the reliability of answers is
closely related to user prompts. For the same question, different
answers are obtained if the user interacts with ChatGPT
differently. For instance, as shown in Figure 6, the user posed
the first query in a direct command tone, which caused
ChatGPT to attempt to force a synthesis of two irrelevant
concepts (i.e., snake and coffee machine), resulting in obvious
commonsense errors and implausible answers. However, when
users ask the same question in a more request tone, the answer
appears more reasonable and accurate. And ChatGPT will point
out that 'snake' is not the most obvious source of inspiration.

Figure 6. An example of different user prompts

Additionally, the quality of answers is also influenced by the

query frequency and detailedness. Our testing has revealed that Figure 7. An example of abduction questions to test creativity
 Xingzhi Wang et al. / Procedia CIRP 119 (2023) 7–14 11

Additionally, ChatGPT is not subject to many of the However, in the manufacturing industry, many tasks require
cognitive limitations that affect humans, as increased demands high standards for factual accuracy. Performing relevant tasks
of operations in question did not appear to increase its cognitive remains a challenge to ChatGPT. Considering that big data
load. For instance, when posed with a question in Figure 8, the analytics and machine learning-based analytics are becoming
model generates 20 immediately-sound answers. Even when popular in manufacturing industry [4, 12], integrating relevant
we increased the number of concepts from 20 to 40, ChatGPT analytics could offset deficiencies of ChatGPT.
could still produce the required concepts promptly without
compromising quality. As human creativity is often 4. Future prospects of ChatGPT in the manufacturing
constrained by knowledge base and cognitive capabilities, industry, integration roadmap, and key questions
ChatGPT could be an invaluable tool for engineers to stimulate
creativity. 4.1. Future prospects

In general, ChatGPT demonstrates several promising

capabilities to greatly improve efficiencies and potentially lead
to breakthroughs across various manufacturing industries.
Figure 9 highlights the potential prospects of ChatGPT, which
could include improving human-machine collaboration,
knowledge management, design innovation, and engineering
education within the manufacturing industry.

Figure 8. A question to examine the creativity and cognitive

load of ChatGPT

However, it is important to note that ChatGPT is trained

using massive amounts of online data with a varying degree of
quality. Consequently, the solutions provided by ChatGPT are
unlikely to exceed the average quality of human solutions.
Furthermore, unlike experienced experts who can proactively
connect cross-domain concepts through divergent thinking,
ChatGPT is not capable of doing so. Its responses are Figure 9. Future prospects of ChatGPT
coherently aligned with the topic unless users have provided a
list of related concepts for inspiration. Considering that the First, the combination of ChatGPT and current
creativity of ChatGPT may not surpass that of experienced manufacturing systems could greatly enhance human-machine
engineers, it could be especially helpful to train novice collaboration. Manufacturing systems today are equipped with
engineers in generating creative ideas. various interfaces and communication technologies for
interaction. Those diversified interfaces are overwhelming to
3.4. ChatGPT lacks high-level analytical skills and operates engineers as they require much effort to be used proficiently
as a black box when making decisions. [13]. ChatGPT could provide a simple and unified interface for
engineers to interact with different systems simultaneously.
According to our test, ChatGPT lacks advanced analytical Due to its extraordinary situational understanding of human
skills required for the manufacturing industry. When presented text, ChatGPT could significantly reduce the amount of effort
with 26 critical analysis questions, ChatGPT could only required. If the manufacturing industry can implement
provide correct answers for 46% of them. It struggles with both ChatGPT on top of other systems, users can simply use natural
qualitative and quantitative analysis and cannot use task- language to interact with ChatGPT, and ChatGPT will convert
specific analytical tools or methods. For qualitative analysis, natural language into required text commands, voice
ChatGPT tends to rely on inductive arguments, exemplifying commands, or programming commands, to control other
several observations to make broader generalizations. While manufacturing systems. This approach will help engineers be
many of its results are not obviously false, its reasonings were more efficient in coordinating various systems simultaneously.
easily seen with multiple mistakes. Moreover, ChatGPT makes ChatGPT also has the potential to improve knowledge
decisions in a black box manner. When performing qualitative management efficiency. Currently, design and manufacturing
analysis, it is difficult for users to see how it prioritizes involve enormous amounts of knowledge across various
competing interests. For example, when evaluating design domains, such as material science, physics, computer science,
concepts, it's unclear if it prioritizes cost, efficacy, or cost- management, etc. Relevant knowledge is stored in multiple
effectiveness. Regarding quantitative analysis, ChatGPT formats, such as natural languages, images, video, audio,
frequently employs irrelevant formulas to solve the problem geometric models, etc. However, there lacks a universal access
and may skip intermediate steps for complex calculations that method to identify and extract multi-modal data. Moreover,
require multiple sequential operations [11]. there is a lack of universal methods to summarize and represent
In summary, ChatGPT in its current version can assist with different types of knowledge, which hinders fast and flexible
some commonsense reasoning and inductive reasoning tasks. retrieval of solutions to engineer's queries. ChatGPT could
provide a solution to these problems by changing the way
12 Xingzhi Wang et al. / Procedia CIRP 119 (2023) 7–14

knowledge is searched and retrieved. With its strong text also be developed to enable ChatGPT to perform more complex
synthesis capability, ChatGPT can effectively summarize tasks. At the outer layer, the manufacturing industry must
multi-modal knowledge, making stored knowledge more restructure engineer skills and operation procedures to enable
accessible for fast and flexible retrieval. more effective collaboration with ChatGPT. In each layer,
ChatGPT also has the potential to facilitate design specific tasks should be performed accordingly, based on task
innovation. Traditionally, due to cognitive limitations, complexity or priority.
designers must frame the situation of a design problem. In the knowledge layer, ChatGPT should integrate with
However, designers should also reframe design problems more specialized technical knowledge to be more reliable. By
constantly to adapt to social-technical changes and uncover assorting knowledge complexity from low to high, knowledge
overlooked information. This reframing process often leads to can be classified into task-specific knowledge, product/process
significant leaps in the innovation process. ChatGPT provides knowledge, and enterprise-level knowledge. Task-specific
an efficient way to help designers reframe problems. ChatGPT knowledge is considered the least complex for two reasons.
is trained using crowdsourced online data, and its answers to First, task-specific knowledge is typically well-documented
the same question are revolutionizing in accordance with social and easily accessible, and many have already been codified and
dynamics. By employing lateral thinking techniques (e.g., embedded in KBS. Previous research has successfully fine-
5W1H) and routinely querying ChatGPT, designers can obtain tuned a general GPT based on task-specific knowledge for
immediate feedback to see if new context and opportunities concept generation [18]. However, product/process knowledge
have emerged. This approach has the potential to revolutionize and enterprise-level knowledge are more difficult to obtain as
the design process and help designers overcome several grand they often span multiple domains. Some of this knowledge,
challenges they face [14-17], ultimately enabling them to create such as societal and business knowledge are tacit or
innovative solutions for complex design problems. experiential, making it challenging to identify and represent.
ChatGPT can also support engineering skill training and
education. The most exciting feature of ChatGPT is its endless
knowledge base and text generation capabilities, which can
help engineers dive into new knowledge domain more quickly.
Suppose an engineer wants to acquire a new skill. In that case,
ChatGPT can provide basic knowledge structure, summarize
relevant concepts, and offer detailed explanations of
terminologies. Additionally, ChatGPT can serve as an AI coach
to help engineers refine their skills by providing instant
feedback. For instance, after completing a task, ChatGPT can
provide diverse perspectives, highlight apparent mistakes, and
suggest directions for improvement in a timely manner. With
its ability to provide personalized, on-demand training and
feedback, ChatGPT has the potential to revolutionize
engineering education and training programs.
Despite its potential benefits, ChatGPT has some limitations
that need to be considered. Firstly, not all knowledge generated
by ChatGPT is reliable or appropriate. As previously
explained, this is partially due to the lack of high-quality design
and manufacturing knowledge in the training dataset.
Additionally, ChatGPT’s reliance on logical reasoning alone Figure 10. A three-layer model for ChatGPT integration
makes it insufficient to perform advanced analysis required by
the manufacturing industry. Due to the inherent limitations of In the reasoning layer, as a probabilistic model for
ChatGPT, it is important to use it with caution and to verify the knowledge, ChatGPT needs more advanced reasoning
accuracy of its responses. Furthermore, the manufacturing capabilities to be reliable. Based on priority, ChatGPT should
industry should carefully consider which tasks to outsource to first provide original sources for its responses. This way,
ChatGPT. Users may need to develop new skills to ensure that engineers can verify solutions generated by ChatGPT by
ChatGPT is used appropriately as well. tracing back to original files and resources. This approach
could help engineers identify fabricated knowledge, making
4.2. A three-layer model for ChatGPT integration them more confident with ChatGPT. Second, ChatGPT could
further integrate with existing KBS to enhance its analytical
To fully leverage the benefits of ChatGPT while mitigating skills. KBS encode numerous rules, logic, formulas, and cases,
its limitations, a triple-layer model for ChatGPT integration is enabling ChatGPT to interpret user intentions and call relevant
proposed, as illustrated in Figure 10. The knowledge layer is analysis functions in KBS to produce more concrete solutions.
the core of the model, emphasizing the need for an enhanced Additionally, KBS solutions can be enriched by ChatGPT as
knowledge base to enable the application of ChatGPT in the results are more detailed in information. Third, ChatGPT could
manufacturing industry. In addition to improving the be integrated with knowledge graphs. By doing this, ChatGPT
knowledge base, more advanced reasoning capabilities should is logically connected with cross-domain concepts, making it
 Xingzhi Wang et al. / Procedia CIRP 119 (2023) 7–14 13

possible to perform divergent thinking and more advanced graphs to improve its reasoning and analytical skills. In the
reasoning. collaboration layer, manufacturing industry professionals
In the collaboration layer, engineers must acquire new skills should reconsider the way engineers collaborate with ChatGPT
to interact with ChatGPT more effectively. First, the answers and develop new skills to use ChatGPT appropriately.
provided by ChatGPT are not definitive, and therefore an Regarding future research, we will propose a ChatGPT-assisted
expert-driven verification process is required. Such a process engineering design method to help novice designers quickly
would help filter out faulty knowledge, thereby avoiding acquire design skills and get desirable design outcomes.
misguided decisions. Second, prompt engineering should be
performed to enable users to understand how to pose questions Acknowledgment
that elicit specific knowledge required for design and
manufacturing tasks. They should also learn how to ask a chain This work is supported by an Australian Government
of questions to drill down to the details they need. Third, a new Research Training Program (RTP) Scholarship.
collaborative design and manufacturing procedure should be
proposed to clarify which tasks are outsourced and what skills References
and characteristics engineers should possess.
[1] E. A. van Dis, J. Bollen, W. Zuidema, R. van Rooij, and
4.3. Key research questions C. L. Bockting, "ChatGPT: five priorities for research,"
Nature, vol. 614, no. 7947, pp. 224-226, 2023.
Here are some key research questions on ChatGPT that [2] B. M. Li, S. Q. Xie, and X. Xu, "Recent development of
might be of interest to manufacturing researchers and knowledge-based systems, methods and tools for One-of-
practitioners. a-Kind Production," Knowledge-Based Systems, vol. 24,
1. Can the limitations of ChatGPT be beneficial for the no. 7, pp. 1108-1119, 2011/10/01/ 2011, doi:
manufacturing industry? https://doi.org/10.1016/j.knosys.2011.05.005.
2. How to incorporate data from existing design and [3] J. Miranda et al., "The core components of education 4.0
manufacturing tools into ChatGPT? in higher education: Three case studies in engineering
education," Computers & Electrical Engineering, vol. 93,
3. What techniques can be used to improve the accuracy
p. 107278, 2021/07/01/ 2021, doi:
and reliability of ChatGPT in the manufacturing
https://doi.org/10.1016/j.compeleceng.2021.107278.
industry?
[4] X. Wang, A. Liu, and S. Kara, "Machine learning for
4. What are the most effective strategies for integrating engineering design toward smart customization: A
ChatGPT into the engineering design process? systematic review," Journal of Manufacturing Systems,
5. What are the potential risks associated with applying vol. 65, pp. 391-405, 2022.
ChatGPT in the manufacturing industry? [5] J. F. Arinez, Q. Chang, R. X. Gao, C. Xu, and J. Zhang,
"Artificial Intelligence in Advanced Manufacturing:
5. Conclusions and future works Current Status and Future Outlook," Journal of
Manufacturing Science and Engineering, vol. 142, no. 11,
In this paper, we conducted an evaluation of ChatGPT’s 2020, doi: 10.1115/1.4047855.
capabilities in supporting design and manufacturing tasks. Our [6] D. R. Krathwohl, "A revision of Bloom's taxonomy: An
results indicate that ChatGPT is impressive in providing overview," Theory into practice, vol. 41, no. 4, pp. 212-
information, generating coherent and structured content, and 218, 2002.
proposing initial solutions. As a generative model, ChatGPT [7] R. Schuette and T. Rotthowe, "The Guidelines of
provides promising creations in a comprehensive, creative, and Modeling-an approach to enhance the quality in
objective manner, therefore showing its potential to support information models," ER, vol. 98, pp. 240-254, 1998.
summarization, synthesis, and creation-related works. [8] L. Floridi and M. Chiriatti, "GPT-3: Its nature, scope,
However, ChatGPT’s limitations were also identified. It limits, and consequences," Minds and Machines, vol. 30,
struggles to understand questions and lacks the ability to pp. 681-694, 2020.
properly use knowledge to generate correct solutions. In some [9] C. J. O’Reilly, "Creative engineers: Is abductive reasoning
cases, ChatGPT can even fabricate non-existing rules or encouraged enough in degree project work?," Procedia
equations in order to generate solutions. Additionally, CIRP, vol. 50, pp. 547-552, 2016.
[10] S. C.-Y. Lu and A. Liu, "Abductive reasoning for design
ChatGPT doesn't have a deep understanding of the underlying
synthesis," CIRP annals, vol. 61, no. 1, pp. 143-146, 2012.
concepts of its responses, and its answers are not always
[11] S. Shahriar and K. Hayawi, "Let's have a chat! A
reliable, traceable, and verifiable.
Conversation with ChatGPT: Technology, Applications,
Based on the testing result, the authors highlighted the and Limitations," arXiv preprint arXiv:2302.13817, 2023.
potential for ChatGPT to facilitate human-machine [12] F. Tao, Q. Qi, A. Liu, and A. Kusiak, "Data-driven smart
collaboration, knowledge management, design innovation, and manufacturing," Journal of Manufacturing Systems, vol.
engineering skill education. Furthermore, we proposed a three- 48, pp. 157-169, 2018.
layer model to help integrate ChatGPT into the manufacturing [13] L. Wang, W. Shen, H. Xie, J. Neelamkavil, and A.
industry. In the knowledge layer, ChatGPT should be trained Pardasani, "Collaborative conceptual design—state of the
with more specialized technical knowledge. In the reasoning art and future trends," Computer-aided design, vol. 34, no.
layer, it should integrate with hard facts, KBS, and knowledge 13, pp. 981-996, 2002.
14 Xingzhi Wang et al. / Procedia CIRP 119 (2023) 7–14

[14] J. Jin, Y. Liu, P. Ji, and C. K. Kwong, "Review on product usage context modeling method for smart
Recent Advances in Information Mining From Big customization," Procedia CIRP, vol. 109, pp. 641-646,
Consumer Opinion Data for Product Design," Journal of 2022/01/01/ 2022, doi:
Computing and Information Science in Engineering, vol. https://doi.org/10.1016/j.procir.2022.05.307.
19, no. 1, 2018, doi: 10.1115/1.4041087. [17] Q. Zhu and J. Luo, "Toward Artificial Empathy for
[15] X. Wang, A. Liu, and S. Kara, "Constructing Product Human-Centered Design: A Framework," arXiv preprint
Usage Context Knowledge Graph Using User-Generated arXiv:2303.10583, 2023.
Content for User-Driven Customization," Journal of [18] Q. Zhu, X. Zhang, and J. Luo, "Biologically Inspired
Mechanical Design, vol. 145, no. 4, 2022, doi: Design Concept Generation Using Generative Pre-Trained
10.1115/1.4056321. Transformers," Journal of Mechanical Design, vol. 145,
[16] X. Wang, A. Liu, and S. Kara, "An ontology-based no. 4, 2023, doi: 10.1115/1.4056598.