I recently realised that many people don’t understand how an AI has to behave.
There seems to be a common misconception that it should never give you the wrong answer, whatever that might mean.
An AI of any use will always give wrong answers.
It is easy to build an AI that is never wrong.
It can simply say “I don’t know”, or fail to answer, every question.
Of course, the problem is that you would tell me that such an AI is no use, and you would be right.
The issue is that you want some answers.
So my AI can start giving you answers; but now, some of those answers will be wrong.
In this respect, AIs are very like humans. The person sitting silently in the corner of the room may well be very knowledgeable and intelligent, but you have no way of knowing, and they are not much help in solving your problems.
On the other hand, the person who seems to know everything may well be a lot of help, but is likely to have remembered things wrongly, or made some incorrect deductions based on what they thought they knew.
The more helpful a person is, in terms of answering your questions, the more likely they are to sometimes get it wrong.
In fact, if you insist they always give an answer, then they will definitely get things wrong.
Just like a school exam – if you are not required to answer all the questions, you can restrict yourself to the things you are confident of, and get most things right.
But if I make you answer all the questions, or mark empty answers as wrong, you are likely to have a significant number of wrong answers, if the questions were of any sort of interesting challenge to you.
This is true no matter how careful you are – there will be things that you think are right, but are in fact not.
So, if you want an AI to answer a good range of questions, you have to accept that it will give undesirable answers sometimes.
In the world of AI and Machine Learning, this is called precision and recall.
Precision is a measure of the proportion of answers it gives that are considered to be correct.
Recall is a measure of the proportion of answers it gives where it is expected to be able to answer.
And the bottom line is that they can never both be 100%.
As one climbs, the other is bound to fall, and the challenge is to get them both as high as possible, and then get the right balance.
The wonder of ChatGPT 3.5 and others of their time was that they seemed to get the recall high enough without dragging the precision down too far.
That is, it gave enough good answers while not giving too many bad and hallucinatory answers; this had been the big problem for general AI & Machine Learning up to that point.
An interesting aspect to all this is how the currently discussed Turing Test is now viewed.
In this construction, someone talks to an AI or a person, and has to work out which it is.
What would the test do if the AI or person simply didn’t answer?
It is arguable that this is the sensible strategy for at least the AI to take, and possibly the person.
Certainly if the AI’s objective is to not be detected to be the AI.
In Turing’s original Imitation Game construction it was different, and this was part of his genius.
He understood that all the participants in the Game needed to have objectives, so he made them play a game.
He then cast the question of Intelligence in terms of the statistical outcome of the game, no matter how good or bad the participants were.
A final comment:
It is a bit like road safety.
I can make the roads perfectly safe, without any deaths or injuries.
I would simply ban all traffic.
And that is actually the only way.
But of course the cost in terms of industry and starvation would be enormous.
OK, impose a 5mph speed limit? The whole society still wouldn’t work smoothly, and in fact there would till be the occasional death and injury.
So the question you have to ask about road accidents is “What is your road death target?” And it probably shouldn’t be zero.
Because the more you want to reduce deaths, the higher the costs to the society.
And, shockingly, if you reduce deaths below the target, then it is possible there will be negative effects on society, including deaths, which will be greater than you wanted to incur.
And in fact recently, I have seen “Vision Zero” for road deaths from Leeds, Oxfordshire, Kent, and Essex, to name a few.
