Use ML & AI with a pinch of salt

A few days ago I was typing a LinkedIn message to a friend by name Dabita (name changed not because of confidentiality, but to avoid embarrassment to anyone). As soon as I typed the name, the machine learning (ML) or the artificial intelligence (AI) algorithm of the LinkedIn Message utility prompted me to consider picking one of the three other possible names: Kabita, Davita, Babita.?

But I wanted to send the message to Dabita! So I overrode the prompt and obstinately stuck to my intended addressee, Dabita; else the message would have been delivered to either Kabita, or Davita, or Babita. Just imagine, upon receiving the message what would have been the reaction from Dabita!?

In this example applying a judgement based on human learning (HL) or human intelligence (HI) overriding an action prompted by the ML or AI algorithm was not a complicated matter. However, it may not be as simple an exercise as in many situations.?

The Maneuvering Characteristics Augmentation System (MCAS) malfunctioning of Boeing 737 MAX airplanes causing a series of automated nosedives, thereby causing two plane crashes (Lion Air Flight 610 and Ethiopian Airlines Flight 302) killing 346 people onboard is still fresh in mind. Fortunately, In December 2018, the FAA privately predicted that MCAS could cause 15 more crashes over 30 years, and subsequently all the 387 Boeing 737 MAX aircraft were grounded; else, who knows, how many more airplanes would have crashed??

MCAS on the 737 MAX was designed to mimic the pitching behaviour of the previous generation of the series, the Boeing 737 NG, by pushing down the aircraft nose from an elevated angle of attack (AoA) by automatically adjusting the horizontal stabilizer and trim tab? (https://en.wikipedia.org/wiki/Boeing_737_MAX_groundings), fundamentally an ML followed by forced AI action, oblivious to pilots flying the aircrafts.?

A simple example of a message getting delivered to Babita instead of Dabita, or a more complex example of an aircraft crash, application of ML recommendations or AI actions must be validated by HL and HI, as no matter how sophisticated the machine learning and artificial intelligence algorithms are, they are always going to be a subset of human learning and human intelligence.?

Two fundamental equations will hold good for generations to come:?

1. Human Learning = Machine Learning + Human Interpretation
2. Human Intelligence = Artificial Intelligence + Human Judgement

Humans create algorithms, not the other way around.?

Just for fun, extending the Dabita example a bit further you may notice that the alphabet ‘r’ and ‘d’ on a typing keyboard are separated just by one diagonal space. So, if you are a bit careless while typing or have a fat finger, instead of typing ‘dear’ Dabita you may end up typing ‘dead’ Davita!?

If ML and AI algorithms have the power of human interpretation and human judgement, they would realize that one cannot send a message to a dead person and would therefore recommend to the user – would you like to use the word ‘dear’ in place of the word ‘dead’? But no! Today’s ML or AI algorithms would at best perform a spellcheck exercise and accept the word dead before Davita. If you do not apply your judgement, you would end up sending the message to ‘dead Davita’, and you will soon be dead; just like the pilots of the two Boeing 737 MAX airplanes did not know how to react to the series of automated nosedives triggered by the MCAS leaving 346 innocent passengers onboard dead… ????????