Conjecture 1: The smartest chess playing system is the one that wins the tournament.

Conjecture 2: The computer system that imitates human dialog is as smart as the human whose dialog was imitated.

These are legacy views from the twentieth century.

Are they correct?

Response to Comments

The question is not about agents. It is about indicators, or metrics if you wish. The question is whether the established game championship and Turing's thought experiment, the dialog centered imitation game, are correctly designed metrics to establish the extent of intellectual ability.

That someone voted to close this question on the basis of broadness is irrelevant. This question and ones like it are important questions for AI researchers and enthusiasts to ask about the set of capabilities we call intelligence, which we attempt to realize in artificial systems.

There may also be a tendency to think of questions like this as seeking an opinion. It is certainly true that our opinions are not useful to the furtherance of AI, and opinion is not what this question invite when it asks, "Is this [pair of conjectures that guides much AI research] correct?"

Correctness must draw on mathematical rigor leading to a yes or a no. Some reasoning leading to the conclusion of correctness or incorrectness must be given. The answer may be yes in one context and no in another, but this is not because of the broadness of the question but because of the broadness of the many useless opinions and possibly counterproductive conjecture that remains from early thinking about what intelligence is and how to gauge its presence in computing machinery or in people.

Questions about intelligence from this author are intended to narrow the terms and concepts so that they can be expressed mathematically. Only with a mathematically terse and elegant set of fundamentals can AI progress as a field in the sciences. Some interesting opinion will probably need to be discarded and a logically constructed consensus must replace broadness and ambiguity for this to occur.

  • 1
    $\begingroup$ I think you should have just asked: "Is winning some fundamental property that an intelligent agent must pursue/have?". $\endgroup$ – nbro Mar 5 '19 at 9:28
  • $\begingroup$ @nbro Although your take is valid, here I think the question is about two key modern notions of AI, utility (which, though relative is objective) and reproducibility of less easily quantifiable intelligent behavior (purely subjective.) $\endgroup$ – DukeZhou Mar 5 '19 at 18:42


My feeling is that utility is the standard of determining intelligence, and merely connotes the strength of a decision making algorithm.

Winning is a metric involving strength in relation to other agents, but may not always result in optimal global utility. (Here I'm thinking of a game AI that always wins. In a game like Chess, a subset of serious players may appreciate this, but it's unclear that this would be any kind of fun for the average player. Similarly, an unbeatable AI in a game like Civilization or Stellaris would be very poorly received by most players.)

In the paper Depth In Strategic Games they formalize the term "strategy ladder". A simple definition of this concept is a series of AIs that are "meaningfully weak". Specifically, players learn to beat them and in so doing learn different levels of strategy, ascending up ladder.

In this example, user engagement is the ultimate goal, not the AI winning every game!


Turing makes a compelling case that imitation is an indication of intelligence (self-awareness and "true comprehension" as strictly corollary.)

I like to reduce this idea to Captcha, because it involves utility. At some point Captcha may no longer work because the average bot is a "smart" in solving them as the lowest-common denominator human.

The same goes for chat bot--it they can successfully imitate humans, they can be considered as smart as those humans in regard to chatting.

Conclusion: Both winning and imitation are indicators of intelligence, whether machine-based of biological.


To some extend I would say yes they are true and there is not that much more to say, but true intelligence and thereby the thing we call smart can only be achieved in a full extend if the machine learns to think and behave not only by imitation but by true learning. A really intelligent Software might me capable of winning every Go game (Go is more difficult than chess as probably everybody knows by now). Would you say this machine is more intelligent than the one that beats you in chess? I'd personally say no. Both are only designed to perform one task and that one alone. But what would happen if the chess computer is somehow learning (without human interference) to play Go? Would we call it intelligent or at least more intelligent than the one only playing Go? Why would we call it more intelligent. Same is for the imitation. I would say that most machines are not capable of imitating a human interaction to the spot, if not all, but lets start with such a computer. One that can overcome the touring test by imitating a human.

Pretty smart, I guess. But: What if an other machine is capable of deliberately failing the touring test? it would not have learned this by the hand of a human, since every human would design it to complete the test. Imagine a machine that can tell it is better to not be identified as intelligent / smart on its own. It might be an advantage to do so. True intelligence can not be obtained by just imitating. Every child and every chessplayer knows that. For example: if you play a game of chess and black moves just as white does, white will always win.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.