Technical Breakdown: Computing Machinery and Intelligence
Alan Turing's Paper From 1950 Posed a New Way to Think of the Question: Can Machines Think?
Summary
Imitation Game: Turing proposed it to test if machines can mimic human behaviour convincingly.
Digital Machines: Digital computers follow predictable, discrete-state patterns and could theoretically mimic human behaviour to a nearly indistinguishable degree in the Imitation Game.
Blackbox Prediction: Alan Turing anticipated that machines might operate in ways we can’t fully understand, even though their behaviour remains observable.
AI Training Prediction: Turing suggested teaching a "child brain" instead of creating an “adult brain”, predicting modern machine learning methods.
Generalizability Prediction: Machines should deduce new rules from basic principles—a core idea in AI today, anticipated in this paper before it truly existed.
The Imitation Game
This paper’s goal is to tackle the idea of machine’s thinking. The Imitation Game, presented by Turing, is an alternative to the question “Can machines think?”.
The game requires 3 players:
A man (A)
A woman (B)
An interrogator (C)
The interrogator is in a separate room from the players and they can only communicate in writing to ensure neither player gains an unfair advantage.
The interrogator knows the two players as labels X and Y and must correctly assess whether X is A (the man) and Y is B (the woman) or vice versa. Player A’s goal is to convince the interrogator that he is player B, while player B’s goal is to help the interrogator correctly guess that she is indeed player B.
Ex.
Interrogator (C): What’s your favourite hobby?
A (the man pretending to be the woman): I love knitting sweaters in my free time.
B (the actual woman): I enjoy painting landscapes, especially in the summer.
Interrogator (C): What’s your go-to makeup product?
A: Definitely mascara—I can’t leave the house without it.
B: Lipstick, because it adds the perfect pop of colour.
The interrogator then uses these responses to guess who is A (the man) and who is B (the woman).
Now to better frame the question using the Imitation Game:
“Can machines think?” → “Are there imaginable digital computers which would do well in the imitation game?”
What is a Digital Machine?
Turing describes a digital computer as having the following three requirements:
Store: A store of information that corresponds to memory.
Executive Unit: Carries out various arithmetic calculations (processing power).
Control: Ensures the table of instructions (or code), in the store, are obeyed correctly and in the right order.
These machines will always obey their orders and never deviate from any given instruction.
A typical instruction given to the digital machine may be something like “Add the number stored in position 6809 to that in 4302 and store the result in 4302”. This describes how the computers we use everyday interpret the code we give it.
How Would a Digital Computer Do Well in the Imitation Game?
To think about how a digital computer could possibly do well in the Imitation Game, we must first learn about discrete-state machines.
Discrete-state machine: Machines which move by sudden jumps from one definite state to another (like a 0 or 1, there is no in between).
I will walk through a slightly simplified version of Turing’s example to describe how discrete-state machines would work:
Imagine a light switch with two positions: on and off, and two possible actions (inputs): flip the switch or leave it as is. The switch can only move between on and off, following a predictable, repeatable pattern.
If we label on as 1 and off as 0, and assume the starting position is 0 (off), the final state of the switch depends on the number of flips:
An even number of flips leaves the switch at 0 (off).
An odd number of flips leaves the switch at 1 (on).
This predictability allows us to always determine the switch's final state based on the initial position and the number of flips.
Digital machines fall into this category of a discrete-state machine but would have significantly more positions; if there are x amount of bits in a digital computer, there would be 2x positions in the machine.
Provided it could carry out these calculations quickly, the digital computer could mimic the behaviour of any discrete state machine if it were given the initial position and inputs.
Therefore, the Imitation Game could be played with a discrete-state machine (as B) and the mimicking digital computer (as A), where the interrogator would be unable to distinguish them.
With this definition of discrete-state machines, Turing poses a new question:
Given a digital computer, if we upgrade its storage, processing power, and code, can it be made to satisfactorily play the part of A (the mimic) in the Imitation Game, with B played by a human?
Note that each of the three modifications correspond to the three basic and required components of digital computers. Turing is essentially saying that if we were to upgrade a digital computer, would it be able to play the Imitation Game well enough to be indistinguishable to the interrogator.
Turing’s Accurate Predictions
In this paper, Turing made some comments that I initially overlooked as they have become common knowledge today. After a moment, I remembered that this paper was made in 1950, 75 years ago from today. This paper came out only 4 years after the first digital computer was made and before the term ‘AI’ even existed. I’ll go over some of his most impressive predictions that I caught while reading the paper.
Blackbox
Over the course of the paper, Turing describes the possibility that we might not be able to describe how a thinking machine actually works, just that it does.
When describing machines that can be used in the Imitation Game, he says “We permit every kind of engineering technique, even allowing the possibility that an engineer cannot satisfactorily describe the way it works”.
Near the end of the paper he also states “An important feature of a learning machine is that its teacher will often be very largely ignorant of quite what is going on inside, although he may still be able to some extent to predict his pupil’s behaviour”.
This remains one of the unresolved questions in machine learning, as the exact inner workings of these algorithms are not yet fully understood.
Building and Training Models
Turing believes that instead of trying to produce a program to simulate the adult mind, trying to produce a ‘child’s brain’ and then teach it would be a better way to make machine’s think.
He then goes on to expand on this idea by breaking it into the child-programme (Algorithm type) and education process (training process).
Backpropagation (the modern training process for all machine learning algorithms) came out in the early 1960s, meaning Turing wrote of this process a decade before it was officially published.
Generalizability of Models
While explaining how complex the ‘child brain’ should be, he basically says that instead of making the child brain know everything, if it knows some rules, it should be able to deduce other rules from it.
For example, if a teacher says “Do your homework now”, some rules would be “Teacher says ‘Do your homework now’”, and “Everything the teacher says is true”. From this, the child brain would deduce that it should do its homework now.
This principle is the core idea of generalizability in models: the ability to accurately perform on new, unseen data by leveraging knowledge learned from previous data.
This innovative paper helped pave the way for artificial intelligence research to thrive. In today’s world, the Imitation Game is exceedingly relevant as machines become increasingly similar to humans. Things like text, voice and even video can now be replicated to a nearly indistinguishable degree, begging the question, can machines think?