use of AI tools to accelerate NDT research. “I presented ChatGPT with a request to develop a Python script that would tackle an old challenge in automated geometry generation for finite element modeling. It generated a script that, with some minor tweaks, was functional but it, unfortunately, did not actually solve the fundamental problem. However, it did make me aware of a library that I was previously unfamiliar. I later tested ChatGPT with a more general task—plot the dispersion curves for a generic isotropic material. It set up the script reasonably well, produced material property values for the material I requested, and generated functional code producing the correct kind of plot. However, the actual calculation, and thus the result- ing curves themselves, were complete nonsense. ChatGPT can, however, be ‘coached.’ So, with a knowledge of how to properly solve the problem, you can provide it with that explanation, and it will ultimately generate the right output. While I think there is real potential value in accelerating the more trivial parts of code development for NDE applica- tions, and it may be useful in providing new per- spectives or potential solution paths for more chal- lenging kinds of problems, it performs best with a human in the loop.” Risk and Concerns Let’s now ask ChatGPT about risks and concerns with NDT professionals using it. ChatGPT Q&A Part II What are the risks or potential issues using emerging AI chatbots with work in nondestructive testing? While there are potential benefits to using emerging AI chatbots like ChatGPT in nondestruc- tive testing (NDT), there are also some risks and potential issues that need to be considered. Some of these include: 1. Lack of human judgment: AI chatbots may not have the same level of judgment or intuition as a human worker, which could lead to errors or misinterpretations of data. 2. Limited flexibility: AI chatbots are only able to perform tasks that they have been specifically programmed to do, which could limit their useful- ness in certain NDT applications. 3. Data bias: AI chatbots may be trained on biased or incomplete data, which could lead to inaccu- rate or unreliable test results. 4. Cybersecurity risks: AI chatbots are vulnerable to cybersecurity threats like hacking or data breaches, which could compromise sensitive NDT data. 5. Ethical concerns: The use of AI chatbots in NDT raises ethical questions about the role of tech- nology in the workplace, and the potential impact on human workers. It is important to carefully evaluate these risks and potential issues when considering the use of AI chatbots in NDT applications, and to take steps to mitigate these risks through careful planning, rigorous testing, and ongoing monitoring and maintenance. What sources did you use to provide these responses on the benefits and risk for using AI for NDT? As an AI language model, I have been trained on a large corpus of text data that includes a wide range of sources on the topic of artificial intelligence and nondestructive testing. Some of the sources that may have contributed to my response include: 1. Research articles and publications from academic journals, conferences, and industry experts on AI and NDT. 2. Online articles and reports from reputable sources such as industry publications, news outlets, and research organizations. 3. White papers and case studies from companies and organizations that are involved in developing or using AI for NDT applications. 4. Technical documentation and specifications for AI technologies and tools that are relevant to NDT. It is important to note that my responses are generated based on patterns and associations in the data I have been trained on, and may not always reflect the most current or accurate information on the topic. It is always a good idea to consult multiple sources and consult with subject matter experts when making decisions about the use of AI in NDT. Humans are required to cite their specific sources when writing. It is a common courtesy to recognize one’s peers. But more important, citation of sources is a critical part of referencing where ideas come from and validating the foundations of their work. Can you please provide articles you are leveraging to respond to my questions on artificial intelligence and nondestructive testing above? As an AI language model, I don’t have access to a list of specific articles or sources that have contrib- uted to my response. This is because my responses are generated based on patterns and associations in the data I have been trained on, which includes a vast amount of text data from a wide range of sources. J U L Y 2 0 2 3 • M A T E R I A L S E V A L U A T I O N 31 2307 ME July dup.indd 31 6/19/23 3:41 PM

However, I can suggest some relevant articles and sources related to artificial intelligence and nondestructive testing that you may find useful: 1. “Applications of artificial intelligence and machine learning in non-destructive testing” by A. T. Arulrajah, P. R. W. Nadeeka, and M. A. Ali. This is an academic research paper that provides an overview of the applications of AI and machine learning in NDT. 2. “Non-Destructive Testing and Inspection Market -Growth, Trends, COVID-19 Impact, and Forecasts (2021 -2026)” by Mordor Intelligence. This is a market research report that provides insights into the trends and drivers in the NDT market, including the role of AI and machine learning. 3. “Artificial Intelligence in Nondestructive Testing: A Review” by M. Elsafi, S. Saeed, and S. M. Zubair. This is another academic research paper that provides a comprehensive review of the use of AI in NDT, including the benefits and challenges. 4. “Nondestructive Testing Using Artificial Intelligence: Trends and Opportunities” by S. K. Tiwari and R. K. Jain. This is a research article that discusses the use of AI in NDT, with a focus on the potential benefits and challenges. I hope you find these resources helpful in your research on AI and NDT. At first, this series of responses appears appro- priate and honestly impressive. However, there is one major issue if you dig deeper none of these suggested references in the last response are actually real publications. They are all fabrications. Why does this happen? GPT-3 is a language model that uses deep neural networks to predict the next word in a sequence of text based on the context provided by the previous words. Sometimes, when the predicted prompt responses are not precisely matched, the LLM will still do its best to string together the most plausible text response. Unfortunately, for some instances where precise responses are expected, sometimes they turn out to be nonsense. Besides fake references, poor logic and unex- pected responses are an issue for AI agents in 2023. The term hallucinations has been given to such fabricated responses (Alkaissi and McFarlane 2023). This issue becomes more critical when such tools are applied to technical and scientific work, like NDT, where specific details are of the utmost impor- tance and where we can’t afford to have such errors and fabrications. On NDT.net, there is a thread highlighting the bad idea of “ChatGPT3 writing your inspection procedure” (Bisle 2023). Clearly, AI agents are not ready to be given large complicated technical writing tasks and be expected to produce error-free content. From my perspective, this is OK. NDT technicians, engineers, and researchers should be leading and responsible for the quality of written procedures, reports, and scientific publications. The other issue that the aforementioned inter- change tried to highlight is the lack of any means to reference and verify where such content originated. I’ve written about the benefits and risks of AI for NDT in the past and was curious where ChatGPT was getting its material. While it is impressive that such AI agents can generate articulate responses to such questions, I do see an ethical issue. If these language models are being trained using material on the order of the content of the Library of Congress, shouldn’t they do a better job of pro- viding the source material for their response? To some degree, the current versions of these AI tools operate like efficient plagiarism agents, which is the antithesis of quality technical and scientific writing that depends on collegial citation. The Future These tools have come a long way in recent years and will only get better. While ChatGPT is based on GPT-3, OpenAI recently released GPT-4, which has received many positive reviews (Metz and Collins 2023). (While there is a monthly charge to access GPT-4 directly, Microsoft Bing Chat does provide free limited access to GPT-4 today.) There are also a number of other promising AI tools to explore today like Google’s Bard, DeepMind Sparrow, and Amazon Titan. In terms of knowledge capability, GPT-4 has been trained to be more precise and OpenAI claims it can score a 1300 (out of 1600) on the SAT. So, training on a wider depth of material and taking more care with the content selection will help. But, to some degree, if these AI agents are trained using the broad history of human writing, all of the positives and the negatives of our writing will be baked into these algorithms. The current black-box architecture will make it challenging to eliminate false or offensive responses. Going forward, the most effective way of using such tools will be in a collaborative way. This will follow our general experience with the applica- tion of AI/ML for evaluating NDT data, where maintaining a critical role for human inspectors ensures NDT data quality and helps compensate for instances of poor AI performance. (See Lindgren 2023 on page 35 in this issue for more discussion on this topic.) Workers are already finding ways to leverage these tools effectively while doing their job. In a recent survey, over 40% of Americans said they were using generative AI technology at work FEATURE |AI/ML 32 M A T E R I A L S E V A L U A T I O N • J U L Y 2 0 2 3 2307 ME July dup.indd 32 6/19/23 3:41 PM