AI seems a formidable tool when you are building your company and need expertise in a field you are not familiar with. At the beginning of your entrepreneurial journey, you cannot have all skills in-house and you need to prioritize your costs. You generate prompts, you refine them with your constraints and wishes, and soon you get an insight into what you could do, and you can start implementing it. That sounds easy and will help you make early decisions.
AI gives you fast answers, at low or zero cost, and seems a tempting approach. However, there are limitations when it applies to Product Development, from my experience in developing Hardware products. Those limitations will cost you precious time when building your hardware product. You might make the wrong decisions early, decide to dive into a project without assessing its feasibility or risks properly and end up having lost months of time and work.
Let us examine together the limitations I encountered so far.
AI hallucinates.
I have tried several AI searches for literature to find research articles faster. I usually document my suggestions for product development based on results found in accessible science and research, which is published as scientific articles in specialized journals and thesis. At the beginning of a project, I might not know the relevant keywords I will need to refine my search. The AI searches are useful to acquire them. However, when looking directly at answers to my technical questions, sometimes AI suggested solutions based on no research at all, and no real reference on which its answer was based.
I decided to evaluate its abilities and overcome this hallucination problem by refining my prompt and asking it to mention which article it was based on and give its references so that I could find it online. It gave me the reference of an article that sounded promising, and would give me an interesting formulation basis I could work on; however I could not find the article later on, for the simple reason that AI had changed the name of the polymer the article was studying. The real article I found later dealt with another polymer, with completely different properties, and did not fit my needs. This phenomenon happened several times. So, I went back to my initial search, with more tedious but reliable tools. And I found another tool to search into existing literature faster than the manual way. It does not have to be AI and still answer your question.
AI is good for the initial keywords search, then there are other efficient tools to locate research articles.
AI does not have a critical approach to scientific articles.
Having a critical mind when reading a research article is necessary. The academic environment is known for its “publish or perish” characteristics: the more articles you publish, the more you can get funding and keep working as a researcher. The harshness of the competition has led to the development of “paper mills,” which generate completely fake articles based on no research results nor studies. Those papers sound relevant, logical, and reliable when you read them and you have little knowledge of the field about which you are reading.
However, according to an article published in 2023 in Science, these fake papers have become so numerous that 20 to 30 percent of papers would be fake nowadays (Brainard, 2023).
It takes experience of which characterization methods have been used, what their accuracy is, what you can deduce from the results, and which limits they have, to determine if a paper is a fake or not. AI is not trained well enough to notice an article from a predatory journal, it cannot assess if the experiments have been made properly nor if the results are reproducible.
Experienced researchers use their knowledge of the field daily to assess whether a paper is fake or not. It takes time and effort, but when you collaborate with an expert, you have a more accurate idea if your product development starts safely and reliably.
You can use AI to compare the properties and features in papers together, but only a trained expert can assess if the sources are reliable.
Risk assessment is not reliable.
When you ask technical questions to AI about which materials you could use for your desired application, it will provide you with the most efficient one, no matter what, based on the properties you require.
It can include cost, chemical safety, environment-friendliness if you add them in your prompt, but it will not be able to check if the component in question will react in a dangerous way with the rest of your formula. It will not mention in which conditions the compound needs to be stored, what happens if you do not store it accordingly, which regulations and safety measures you will have to comply with when you order, store or use it.
These factors add to the cost estimate when developing your product, in terms of safety measures, regulation compliances, and legal responsibility.
All this information is part of the training of a professional: reactions between different compounds come from chemistry courses and experience in formulation; we do not know necessarily all the hazards for each compound, but we have been trained to check them before using them, to use protective equipment and what to do in case of an accident. All this expert training makes it easier for you to build prototypes and/or a pilot line safely from the start.
Rely on professionals for assessing risks, especially with chemicals.
This said, AI is a tool, which used wisely, brings amazing outcomes. It is important to remember its limitations and when to involve experienced professionals in your project.
Bibliography
Brainard, J. (2023, May 9). Science, Science.org: https://www.science.org/content/article/fake-scientific-papers-are-alarmingly-common?ICID=ref_fark