Editorial: AI and Machine Learning: Proceed with Caution

One of my favorite movies growing up was Terminator 2: Judgment Day. For those unfamiliar with the Terminator movies, they center on humankind’s resistance to the rise of a malevolent artificial intelligence (AI) entity known as Skynet. Some of the largest pushback against the burgeoning use of AI is prompted by such media portrayals. But the reality is that AI has a long way to go before it can take over the world.

Compared to computers, humans excel at ignoring obviously misleading information when learning about new topics. For chatbots based on large-language models (LLMs), such distinctions are difficult. LLMs are only as reliable as the data they are fed. For example, in a popular Forbes article published in May, the author collated several examples of Google’s AI providing hilariously misleading results to simple queries. For one user who wanted to stop the cheese from sliding off of his pizza, Google’s AI Overview suggested adding “about 1/8 cup of non-toxic glue to the sauce to give it more tackiness,” drawing from a Reddit troll post from over a decade ago.

Google is not alone when it comes to their AI product generating absurd responses. I was playing around with OpenAI’s GPT-4o — the current default ChatGPT model for free users — and found that it can struggle with basic comprehension. For example, I asked it “Which is bigger? 4.9 or 4.11,” and it answered: “When comparing version-like numbers, 4.11 is bigger than 4.9.” I’m no math whiz, but even I know that 4.9 is greater than 4.11. However, by the way that the program responded, we can tell that it interpreted a simple math problem as a comparison of version-like numbers — for example, different versions of the same software. If I asked a human this (admittedly, poorly worded) question, they would likely answer 4.9 or request clarification on the context.

I asked Meta AI (another chatbot) a similar question and received a correct answer of “4.9 is larger than 4.11.” Because the model didn’t initially give me any of the rationale behind the answer, I asked a follow-up question of “Why?” for more insight. The model replied, “In this case, 4.9 is larger than 4.11 by 0.78.” In reality, we know that 4.9 is larger than 4.11 by 0.79.

While it is quite easy to break out a calculator and verify the math in this case, I wonder if simple errors like this can bleed into other areas of the LLMs, which can color responses and lead to further inaccuracies. For instance, I asked ChatGPT, “What has a greater atomic mass: Cerium or Praseodymium?” The answer it gave, Cerium, was incorrect — although it did tell me the correct atomic masses for both elements (140.12 amu for Ce and 140.91 amu for Pr).

For now, relying on LLMs to do your job will almost certainly lead to problems. Still, these tools can be extremely helpful for specific tasks, like finding related references for a niche topic of interest or summarizing action items from a meeting transcript. Our Special Section on AI and Digitalization this month examines some of the more practical aspects of using AI. Articles review how and where chemical engineers can use LLMs effectively (pp. 19–25), how AI is being used for HAZOP revalidation studies (pp. 27–34), and how AI-driven digital twins represent a first foray into the industrial metaverse (pp. 36–41).

Although the AI systems of today are not yet ready to lead a robot uprising against humankind, their use in engineering processes and our day-to-day lives is very much inevitable in the coming decade.

Emily Petruzzelli, Editor-in-Chief