The practice of AI is pulling us all into a black hole that is not entirely clear to us. The experts who develop these intelligent models claim that they themselves cannot always predict their own ways of working. It is possible that this ambiguity is for marketing/defense purposes, since admitting to the calculated programming of the model's "thinking" and working methods exposes them to lawsuits. And malfunctions do happen. The last story I heard was about an AI agent that was programmed to prepare purchase orders in a large organization. When it came across a purchase order from a budget item that the operator forgot to arrange for him to use, he prepared an order anyway. It is a forgery in every way – a perfect document with all the necessary marks and seals on it, but it has no budgetary coverage. The order was sent to the supplier. Only when the invoice arrived for payment for the service provided did it become apparent that the order did not come from the system. A small story, the money was not much either. One can only guess that malfunctions of this kind happen all the time in all kinds of organizations at all levels. Perhaps the financial savings due to the amount of work an agent can perform covers these types of faults. Or the amok pursuit of the modern and up-to-date causes people to turn a blind eye or treat them as necessary birth pangs.
In this essay, I decided to put AI's capabilities to the test and compare them to human capabilities. It starts with a personal story, and continues to the professional part. All my life I have suffered from edema in my vocal cords after viral infections, which forced me to remain silent until the cords recovered. At a young age, this took a week to ten days, but as I got older, the recovery time increased until it reached almost a month. At this point, I visited a vocal cord specialist and began treatment with a speech therapist who insisted that she could help. I received interesting information from both of them – the specialist took a photo of my vocal cords and said they were normal, but there was a thick liquid discharge. She recommended comprehensive allergy tests and a test to identify hidden lung disease. The speech therapist, for her part, tried to teach me to use my vocal cords better – we had very interesting sessions, in which she proved to me that with practice I could produce a different voice, lower, deeper and more masculine, and the strain on my vocal cords was surprisingly less. But when I asked her what it had to do with the original problem, I didn’t get an answer. I stopped the series after a few sessions.
A friend from work, a very smart guy, immediately advised me to stop consuming dairy products, since dairy products are known to produce mucus, and there is a very good chance that they are related to the matter. A good place to start. I took his advice, did not perform allergy tests and simply stopped consuming dairy products (drinking milk, yogurt, fresh white cheeses, in some cases also butter, etc.). The improvement in the general feeling of the vocal cords came almost immediately, regardless of any disease conditions. The fact that I did not perform an allergy test undermines the argument, but by all indications it is indeed better for me to avoid substances known to increase mucus production. The advice was correct, fast, and was based on knowledge and logic.
This week I turned to the AI for help. When I described my situation it told me that I needed to consult a vocal cord specialist and a speech therapist. That’s a start.
When I went to the AI and told it the specialist had seen discharge on the vocal cords, it referred me to saline irrigation, including nasal irrigation. It said that this in itself could not explain the prolonged problem after an illness, but it would be worth performing treatments that could "dry up" the excess mucus.
After discovering that there was no doubt that I was better off avoiding dairy products, I realized that the question to ask the AI was completely different, and only now would it be possible to use its unique capabilities – “Is it possible that there is a synergistic relationship between a viral infection and an allergy to dairy products that causes a worsening of post-viral edema of the vocal cords?” This took Perplexity’s deep-research model, with which I usually work, many minutes. It searched, found all sorts of tangential studies, and found researchers who touched on the field, but not a single place where such a relationship was proven in a controlled experiment. I asked several times, changed the wording of the question, tried to make it easier for it – not to look for synergy, only some correlation. It searched and searched – but it didn’t find anything.
Then I went to Google Scholar myself and asked to perform a search connecting the two topics. I only performed one search. Even I was surprised – the first article found answered my question[i]. Below is a screenshot of the first answers that appeared in the search:
Chadwick's 2003 article explicitly discusses the synergistic effect of allergies and viral infections on vocal cord problems. But this article comes from the experience of vocal cord specialists in their clinics. It is not a controlled study, but it cites several studies and refers to the subject as known and familiar to practitioners in the field. It is sad that modern doctors are not aware of the importance of this information, but it is not surprising in a situation where there is a shortage of doctors and minimal time is devoted to each patient.
When I passed this on to AI for review, it apologized, writing that the subject has indeed been known for a long time and that he regrets not having found it earlier. In my work with AI, I have already encountered apologies of this kind – for example, when I tried to use AI to build and improve my website where these essays are published. At least three times the AI model wasted entire hours before admitting that it had no idea how to solve my problem. For the sake of balance, I must admit that I used the AI model a lot during the construction of the site, which saved me from having to read instructions and watch tutorial videos. But if my question referred to a topic not covered by the above instructions – such as codes that are blocked from being changed within the template, he could have admitted in advance that it was beyond his capabilities, but it didn't.
One can try to find many excuses – for example, that the model I use is not the most expensive version but only a medium-priced version. I have an excellent answer for that too. About a month ago, my coworker and I were asked to perform a “search and sort” task for companies participating in the CPHI conference. The question was clear – which of the companies participating in the conference had the capabilities we were interested in. There was no reason to get two different answers from two AI models – this is not a complex analysis of data or search for hidden information. And yet, the two lists were not only very different in length, but there was almost no overlap, unbelievable. This is despite the fact that I use a paid channel at an intermediate level, and my colleague who is 30 years younger than me and has a much higher than average intelligence level uses chat gpt and a high payment of $ 100 per month, and the longer list was actually mine…
I heard someone say that he is constantly busy thinking about how to ask the AI the right questions, and that it really keeps him awake at night. If he could just ask the right question, he would get all the right answers that will solve all his dilemmas. I disagree with this approach:
1. There are cases where there is no substitute for human search. Over time, we need to learn to classify these cases.
2. AI can greatly speed up finding and cross-referencing source documents, but if you want to understand the topic, not miss important information, and not be satisfied with a summary, there is no substitute for reading source documents. The best example for me is regulatory documents for the pharmaceutical industry (FDA, EMA, ICH, etc.). Using AI alone can be suitable for management levels, but those interested in understanding the material being studied must read the source documents.
3. The way the AI model works is not entirely clear, and the sensitivity to question wording, focus, and the need to constantly remember that for some reason some documents are less accessible to AI or at least some models require human review.
4. An AI model is capable of searching and cross-referencing much faster than any human, but the amount of articles accumulated in the last decades is enormous, to the point where satisfactory performance of the task is impossible in a reasonable time.
5. I guess that providing a quick answer is a basic element in these models to make them attractive from a marketing perspective (e.g. no more than 10 minutes per question) but in the case of searching for initial information within a network overloaded with information, this is a dangerous trap.
At the same time – it is important to emphasize that I, like everyone else, regularly use AI services as part of my daily data mining. Usually, the answers received are good, quickly direct me to the right places and save a lot of time. So much so that learning a new topic that required many days may be completed in one day or less, and there is no doubt that this is a revolution. But reading or at least reviewing the original document can be critical so as not to miss what I, as a person and a scientist, think is the main thing.
The root of the problem is the illusion that an AI model thinks in a human way – when a person looks at the landscape from a top of a mountain, he sees the whole thing at once. And if a woman appears next to him he will see her too. A computer does not work that way, it scans the area item by item, even if it is done at very high speed. In the movie "Terminator" James Cameron tried to imagine decades in the future how the robot's vision process would proceed – and this is exactly what he predicted – scanning and analyzing items in sequence, one after the other.
Same problem exists when searching and analyzing a lot of information – the model decided what to search for based on my question, but because it works like a robot it scanned many thousands of documents but missed the most important one. It doesn’t see the whole landscape – the human brain works completely differently. I didn’t know exactly what I was looking for, and paradoxically that’s why I found the answer in the blink of an eye.
And what was the solution to my vocal cord problem? Very simple. My physician took pity on me and prescribed me steroid treatment that shortened the recovery time. Maybe not the healthiest, but there is a psychological cost in avoiding speaking for weeks. No need for experts, no AI, and no literary searches. Sometimes the best solutions are the closest and simplest.
Best wishes for good health to all readers.
[i] Chadwick, S. J. 2003 “Allergy and the contemporary laryngologist”, Otolaryngol Clin N Am 36 (2003) 957–988
כתיבת תגובה