Time to put the intelligence in AI
By Ben Iannotta|September 2021
Mistakes by automatic speech recognition software, a form of artificial intelligence, can be a hoot, provided the errors don’t get out the door. Here are some of my favorites from recent interviews we’ve conducted at Aerospace America: “Silicon doped with arsenic” came out as “silicon Dr the arsenic,” “A-SAT launch” as “ASAP lawn,” and “boost glide” as “blue squad.” The human mind, aware of context and the stakes of getting things wrong, can easily spot such errors. Listening back helps too. But even with all those precautions, I did momentarily fall for “blue squad” in this issue, thinking that it was slang for U.S. Air Force personnel.
Let’s be charitable and accept that ASR technology converts speech to text correctly 80% of the time. The fact that it does so affordably strikes me as an amazing feat, given that society is just at the beginning of the artificial intelligence and machine learning revolution. But putting my editor’s hat back on, I am nervous about the risks this technology brings to our work here at the magazine. Looking at matters that way, ASR’s performance is dismal. A reporter or editor who incorrectly quotes a speaker 20% of the time would have a very short career.
Why do I raise all this? Fairly or unfairly, rolling out underperforming artificial intelligence software in our private and professional lives could poison our views of automation for headier tasks, such as transporting us from here to there in an urban air mobility aircraft. I am not surprised when I see reports of a road accident caused by software, even though intellectually I remind myself that I have no idea whether the computing code in the automobile bears any resemblance to that required for turning the spoken word into text. For me, remains this: How am I supposed to believe bits and bytes can drive a car or fly a plane if they can’t do seemingly simpler tasks?
Perception is reality, so here’s an idea: The firms creating the autonomy software for aircraft could help themselves by proving they can create higher performing but still affordable consumer software. I might believe your code can steer an aircraft safely if it can figure out that we are not doctors of arsenic here at Aerospace America.