LCGC Blog: Artificial Intelligence: The Good, The Challenging, and The Terrifying

Artificial intelligence (AI) has become a mainstay in our daily lives. The term invokes a wide range of thoughts in the minds of different people, from those who do not use it, to those that use it for various everyday tasks, to those who have integrated its use into their work world. Succinctly, AI is about building systems that can perceive, learn, reason, and act in ways that would be considered intelligent if done by humans. The term encompasses a vast range of concepts. Most are familiar with terms like machine learning, deep learning, neural networks, chatbots, and large language models (ChatGPT). AI has permeated virtually any industry you can think of—from healthcare to pharma, education, military, transportation, and beyond. For the past several years, I have begun to integrate AI into my own group’s research. I have also developed a new course for non-science majors that I am teaching this semester, called Sex, Drugs, and Artificial Intelligence. Researching this course has brought me to the regular use of generative AI tools and it has given me a better appreciation for the scope of AI and its future in our society and technology.

AI has exploded onto the scene in analytical chemistry in the past several years. While the overlap of some efforts with traditional chemometrics cannot be denied, AI tools have expanded capabilities for more efficiently solving complex problems, such as chromatographic retention modeling, spectral prediction, and data mining. I have been tracking publications in this area using my Google Scholar alerts for some time,¹ and it is becoming nearly impossible to keep up with the constant stream of developments that combine AI in some form with analytical measurements and measurement data.

Our group has worked closely with data scientists in our industrial engineering department to implement more AI methodology in our research. I do not possess the aptitude to code Python, nor do I understand the intricate ins-and-outs of applying AI to solve problems. It took our groups several months of regular meeting to begin to communicate effectively and understand the boundaries and possible interfaces between data science and analytical chemistry. Since then, we have made some significant progress and contributed our own developments to the scientific literature.

We began working together to develop machine learning-based spectral prediction methods for vacuum ultraviolet absorption spectroscopy that utilize molecular encoding strategies.^2,3 Molecular encoding appears to me to be the essential language to develop further for advancing chemistry using data science techniques. Molecular encoding involves the combination of a variety of descriptors to represent molecules in vector formats, between which quantitative representations of molecular similarity and diversity can be derived to guide and interpret various AI processes but also to enable prediction of new molecules and materials. The choice of molecular descriptors is a key challenge; the descriptors must match the physics and chemistry of the process desired to be modeled.⁴

Molecular descriptors and encoding has been a key feature in our recent efforts to develop surrogate optimization routines for on-line supercritical fluid extraction–supercritical fluid chromatography (SFE-SFC) method development. As a powerful technique that can likely be used to extract and analyze any small molecule from any solid sample material, SFE-SFC still has a multitude of variables and interacting variables that need to be optimized. We had previously demonstrated that more traditional chemometric and multivariate optimization tools were insufficient to evaluate the large numbers of variables, as well as to model the complex response surfaces that resulted from extracting different analytes from different sample types.⁵ Surrogate optimization is a more flexible metamodeling approach that can handle systems with large numbers of variables and interactions, as well as model complex response functions, all while minimizing the number of expensive experimental runs needed. Our initial efforts have demonstrated the power of surrogate optimization for SFE-SFC method development,⁶ and we are now working on further expansion of the evaluated variable space and mapping changes in optimal extraction conditions from different sample matrices. We are using molecular encoding to assess the similarity amongst different analytes and are developing a framework for studying how molecular similarity varies with optimal extraction conditions for different analytes. After all, it is rare that one would just target a single analyte; it will be desirable to understand what variations of, or compromises in, extraction and chromatography conditions will be necessary to determine multiple analytes. Molecular encoding can provide a means for quantifying this effect.

It is exhilarating to delve into the many ways that AI tools can be used to advance analytical science but it also daunting. As little more than a relative novice researcher in AI and data science, I find it challenging to interpret and evaluate the myriad new applications I see being published essentially daily. Additionally, because I have now published research combining AI and data science, I find myself getting a significant number of requests to review such research. I feel quite competent reviewing analytical science but often less so, the details of data science. In that case, in my review comments, I try to make it clear to the editor what are the limits of my ability to judge the validity of some of the work.

I am also an associate editor for two journals. My own trepidations as a reviewer are magnified when I must adjudicate as an editor making decisions for such works. Of course, I can choose reviewers, but there are few who are well versed in both analytical science and data science. I also do not have a rolodex of data scientists to whom I can make consistent requests for reviews. I only see this as becoming an ever-pressing problem and I have rarely had reviewers confide in me that they do not feel comfortable judging some topics in a manuscript relative to others.

While the number of analytical scientists that are proficient in data science concepts will continue to grow as the integration of analytical and data sciences grows, I have a general suggestion that I think could help the peer review process. Each manuscript is supposed to be submitted with a set of keywords that generally capture the scope of the work. I do not think it would be overly onerous to ask reviewers to self-rate their own perceived proficiency associated with each keyword. This would be an enormous help to editors, as they could better ensure that manuscripts were being reviewed by reviewers adequately proficient in diverse subject matters. Furthermore, journal publishers could aggregate this information to provide improved reviewer recommendations to the editors over time.

Another imperative is that we need to broaden the scope of our traditional analytical chemistry education. Increasingly complex instruments capable of complex measurement tasks will continue to develop, and they will generate increasingly complex data. Analytical scientists need to be more intentionally trained in data science techniques. This will improve the efficiency of data analysis, and it will improve the proficiency of analytical scientists to use more data science. While I appreciate the ability to collaborate with data scientists on the cutting edge of their discipline, I can become a more effective collaborator and researcher if I gain better proficiency in data science. We need to seed our students with better proficiency in data science now, so that they are equipped to enter a work world filled with AI capabilities.

It is well known that the pace of technological advancement and the capabilities of AI are increasing exponentially. In my own research seminars where I highlight our use of advanced optimization techniques to develop methods for complicated on-line multidimensional analytical systems, I have also tried to predict what this will look like in the future. I have suggested that in five to ten years, we will not need to perform these complicated optimization routines. I believe we will have learned how to effectively encode chemical systems and model different extraction, separation, and detection systems. I believe, for most new applications using established measurement systems, we will be able to input our problem into some kind of smart database, and it will tell us how to set the parameters of our system to gain acceptable results. We may need to seed such models with results from one or two experiments, but it will not take nearly as many experiments as it does today to arrive at a workable solution for new applications.

Outside of analytical science, as I have prepared to teach my new course, I have read many works on the past, present, and future of AI and its place in our world. Perhaps some of the most awe-inspiring works I have read are those by futurist Ray Kurzweil, where he discusses the possibilities beyond the “singularity,” when biological and nonbiological intelligence merge.^7,8 Kurzweil paints a very compelling and exciting future with many benefits to humankind and he does an exceptional job addressing both potential positive and negative consequences of this merger. I cannot recommend the cited texts enough to those interested in the topic.

In my Sex, Drugs, and Artificial Intelligence class, I have strived to take a balanced look at various topics, including things like the pros and cons of psychedelic-assisted psychotherapy and advanced chemical recycling of plastics. In Kurzweil’s books, I found a great deal of material associated with the benefits of an exponentially expanding AI, including seemingly unbounded health, longevity, creativity, and discovery. And while I felt he effectively addressed how negative consequences could and should be mitigated as technology advances, I felt compelled to find another voice that emphasized the potential doomsday scenarios associated with the development of artificial superintelligence (ASI) vis a vis SkyNet in the Terminator movie series.

One does not have to look far to find severe warnings about the existential threat of increasingly capable AI. I currently find myself reading a rather macabre book entitled, “If Anyone Builds It, Everyone Dies”.⁹ This book, written by authorities on AI, paints a terrifying picture of the impending end of humanity if anyone succeeds in developing ASI. This viewpoint is more succinctly described in a problem statement issued by the Machine Intelligence Research Institute.¹⁰ Briefly, their main points are centered on the arguments that: 1) There is not a ceiling at human level capabilities—once AI reaches and surpasses human level capabilities, it will not stop; 2) ASI is very likely to exhibit goal-oriented behavior; 3) ASI is very likely to pursue the wrong goals; and 4) it would be lethally dangerous to build ASIs that have the wrong goals. When you consider the virtual arms race of AI that companies are waging in today’s society, it does not look to me like anyone is thinking about applying the brakes—they appear to be more concerned with making money. As with much of what I observe happening in the world today, I feel relatively helpless. As someone who has done a substantial amount of research and reading on the topic, I figure the one thing I can do is to at least spread the word and share awareness of this situation.

It is not my intention to convey negative sentiments in this article. I am trying to share my thoughts in the hope that it somehow inspires thought and action in whatever way the reader is compelled. Having experienced the introduction of the home computer, the internet, and the cellphone in my own life, I am excited and inspired by the capabilities that AI and data science bring to analytical chemistry, the discipline I love. I also wonder what my children will experience in their adult lives, given the current rate of technological advancement. I remain optimistic that humans will find a way to install proper guardrails on AI, but I am not certain it is being given the priority it needs and deserves.

References

1. Schug, K.A. Unanticipated Benefits of Keyword Searching the Scientific Literature. The LCGC Blog. July 10, 2014. http://www.chromatographyonline.com/lcgc/Blog/The-LCGC-Blog-Unanticipated-Benefits-of-Keyword-Se/ArticleStandard/Article/detail/847956?ref=25
2. Ho Manh, L.; Chen, V.; Rosenberger, J.; et al. Prediction of Vacuum Ultraviolet/Ultraviolet Gas Phase Absorption Spectra using Molecular Feature Representations and Machine Learning. J Chem Inf Model 2024, 64, 5547-5556.
3. Ghasemloo, M.; Chen, V. C. P.; Rosenberger, J.; Schug, K. A. A Principal Component Analysis-Integrated Machine Learning Approach for Predicting Gas Phase VUV/UV Absorption Spectra of Molecular Compounds. J Chem Inf Model 2026, 66, 299-309.
4. Yang, Y.; Chen, V. C. P.; Kan, C.; et al. A Comparative Exploration on Quantifying Molecular Diversity. J Chem Inf Model 2026, 66, 371–386.
5. Wicker, A. P.; Tanaka, K.; Nishimura, M.; et al. Multivariate Approach to On-line Supercritical Fluid Extraction – Supercritical Fluid Chromatography - Mass Spectrometry Method Development. Anal Chim Acta 2020, 1127, 282–294.
6. Bhakta, N.; Sood, J.; Yang, Y.; et al. Surrogate Optimization with Multivariate Adaptive Regression Splines for Supercritical Fluid Extraction-Supercritical Fluid Chromatography Hyphenated to Tandem Mass Spectrometry. J Chromatogr A 2026, 1767, 466624.
7. Kurzweil, R. The Singularity Is Near: When Humans Transcend Biology; Penguin Books, 2005.
8. Kurzweil, R. The Singularity Is Nearer: When We Merge with AI; Penguin Books, 2024.
9. Yudkowsky, E.; Soares, N. If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All. Little, Brown and Company, 2025.
10. The Problem. MIRI. https://intelligence.org/the-problem/