A couple of months ago, I read an editorial in a cruising boat magazine. One of the editors owned a boat in the NYC area and made a couple of hundred mile trip every summer, every year. He had a list of courses and waypoints saved which worked well every year. He was testing a navigational AI system for accuracy (a system trained and optimized supposedly for this very purpose. He wanted to fine tune it to reflect the effect of some fast tidal flows at certain times of day and the month, variations of weather, latest posted navigational notices posted etc. He input his normal course into the system and asked it to optimize. The result was garbage. It showed the route travellingove land for more than half the trip.
In order to give the program a better chance he input a huge amount of data on weather records of winds, tides and weather along with moon data and historic information that might affect the work. It never gave a meaningful result. Continued to run the boat aground. You could claim it was a highly technical program which worked to an amazint level of pricision, but it gave useless answers.
I worked for a major oil company in engineering, designing the processes used in our refineries, which used the best process simulators on the market. Part of the people worked on the chemical design and processing simulations and another worked on the planning of the ongoing operations to maximize profits and product slates and specifications. AI system developments starting appearing on the horizon in the 80's as near as I can remember, and we regularly tested the systems as they were proposed and beta versions of the commercial work. Progress was very slow for the first couple of decades. Meanwhile we developed our internal models running on commercial modeling software and had some of the best custom models in existence. What set our company apart was the great detail we added to the predictive sections of the model. The models differed from AI in that you put custom predictive capabilities calibrated to our exact situation, based on real operational data.
The trouble I have with AI is that it started as an intentional fuzzy logic system and it still is. It is not a rigorous mathematical or scientific system but a system that looks at everything you decide to supply to it and let it come up with the best "most probable" answer. This is all well and good but if you train it on limited data, the answer is often wrong. If you turn it loose on the entire internet, it has to either include the incredible amount of bad information there or make ranking decisions at some point.
I think it is interesting that one of the obvious major players in the field was Google. They obviously have a lot of real vested interest because of the massive global search work they do. Work like this is totally related to the systems that correct grammer and improve your writing techniques. However, I notice occasionally Google blows it. I did a search as to the most common value of some parameter (I don't even remember the specific question) and I got a labeled AI response that had two paragraphs of background info and values for typical situations which was cogent and patently reasonable. However, the final summarization was diametrically opposed to all the background info and the final answer was obviously a reversal of every background discussion point. It was an obvious black is white answer.
To date it seems that the most work has been done to date on systems referred to euphemisticly as "Plagiarism Systems". They, relatively, should be simple as a punctuation and spell check system. The style issue has been carried pretty far but I notice a rather boring similarity from some authors that I have to attribute to an overuse and standardization of their work. It leads to a lot of tiresome reading at times. The other thing I have noticed in fictional literature is the somewhat regular inclusion of modifiers which are strictly grammatically correct but the case or sense is not something normally used in writing or conversation by most people. The AI correction inserts the strictly proper word but the writing as a whole becomes stilted and legalesque. The same sort of situation arises when these systems change synonyms. This is a problem that has always existed with authors trying to use the most catching phrasiology in their fiction work. It leads to wording that is simply wrong and questions the authors credibility on the subject. I am a huge fan of mystery works and read it constantly. My all time peeve was when one of the most famous mytery writers in the world wrote that the hero walked into a room where a murder by pistol had occured and was confronted by the intense smell of cordite. This was in the 1980's and within the next two years, every murder mystery on the market was exuding cordite smoke on the store shelves. Cordite, of course, was a specialty British propellant used in the late 19th to mid 20th century exclusively in cannon ammunition and in a limited number of large bore high volume large game cartridges. It wasn't a powder but was produced in a long stick resembling thin sphagetti. It also had no distinctive smell different than normal gunpowders. It was obvious that the original quote was the result of a simple thesauris check because the first research search would have produced a page of references showing the written passage would not be true as to application or smell.
The authors of the cordite era used the same techniques as bad AI systems, to little deep research and validation.
