Let's suppose an artificial intelligence developed a level of intelligence which we humans can not understand/handle.

will the program/robot be able to answer every possible question?

My argument: Even with all the science which the program knows, it wouldn't be able to answer questions which we humans couldn't think of (think of it as, humans can't create that which is perfect, by that logic the program can't answer all the questions) and with the more knowledge you have, integration between fields will generate more questions, and answers will generate more questions meaning the program is in a paradox (kinda)

Keep in mind:

1.This robot can make up questions which we possibly never answered or even thought of

2.This robot has to answer this question!

  • 6
    An AI operating according to the known laws of computing can't solve the Halting problem or any other problem known to be noncomputable. On the other hand if someone claims that an AI can surpass the known laws of computing, the burden is on them to explain how that would work. en.wikipedia.org/wiki/Halting_problem. Note that parallelism doesn't help, "deep learning" techniques don't help, and even quantum computing doesn't help. All are subject to the limitations of Turing machines.
    – user4894
    Apr 26 '19 at 18:31
  • AI must learn like a human. It could only answer questions as a human does, based on it's learning. It is true that an AI intelligence has the advantage of being easily connected to other systems. Where a human has an arm, an AI could have a floating point calculator, so it could answer maths questions, and do maths far more efficiently than a human. And in fact, current AI has the whole internet as a limb. It can answer, where was person X at 3.15 last saturday, who was near them, what were they talking about. Does person X like strawberries? Etc.
    – Richard
    Apr 26 '19 at 23:02
  • @FrankHubeny ok
    – user4894
    Apr 26 '19 at 23:24
  • 1
    Not even God can answer "every possible question", many of them simply make no sense, "a fool can ask more questions than seven wise men can answer". And even to simply outperform Turing machines AI'll need hypercomputing capabilities, currently physically impossible. But keep in mind that Turing machines can "outperform" themselves with a bit of luck, if random generation is allowed, and on many tasks they can already outperform humans.
    – Conifold
    Apr 26 '19 at 23:57
  • 1
    Interesting question. First you'd need to clarify what is a "possible question". If it means "grammaticaly correct question", then clearly no AI can answer something like "Is beauty more comparable than beetroots ?". If restricted to only questions that makes sense, you stumble into the very interesting but difficult philosophic problem of making sense of the word "sense".
    – armand
    Apr 27 '19 at 1:45

In 1936, Alan Turing gave the definition of computation which is still in use today. His model of computation has come to be known as a Turing machine (TM).

Turing proved that there is a problem, namely the Halting problem, that no TM can possibly solve.

Any AI using current technology is a practical implementation of a Turing machine; and is subject to the same limitations as explained in Turing's original paper. Even the fanciest, most advanced AI in use or even envisioned, is subject to the limitations Turing outlined.

Some approaches that don't help include:

  • Massive parallelism. Parallel computing is still subject to the limitations of Turing machines. You just execute each logic thread round-robin, exactly as conventional computers run many programs in parallel.

  • Introducing randomness. Computer scientists model this idea as nondeterministic automata. These also have the same computational power as Turing machines. The idea is that the Turing machine just executes every possible branch at each step.

  • Quantum computers. It's true that for some specialized problems there are quantum algorithms that offer substantial speedup over conventional computers. But speed is not a factor in determining what a computer can compute. Quantum computers have the same computational power as Turing machines.

  • Machine learning and "deep learning" techniques. Even the most impressive deep learning machine, AlphaZero, is a conventional program running on conventional hardware. It has no more power than a Turing machine.

Turing even studied [in his doctoral thesis, after writing his famous 1936 paper] what happens when you go beyond Turing machines. Suppose you have a TM, which we know can't solve the Halting problem. We could then introduce a black box apparatus called an oracle that solves the Halting problem. Such an idea is purely theoretical at present.

Turing showed that there is now a new problem that the augmented machine can't solve. And if we add an oracle for that, there will still be yet another unsolvable problem. We would have an endless hierarchy of oracle machines such that no matter how many oracles we add, there will always be an unsolvable problem. This phenomenon is closely related to Gödel's result that any sufficiently interesting axiomatic system [omitting the technical definition here] is necessarily incomplete. And even if you add a new axiom to plug the incompleteness, the augmented system must still be incomplete.

Now if we knew that a human could solve the Halting problem, we'd know that there's a problem we can solve that a computer can't. This would be a great breakthrough. However it's unknown if humans can solve the Halting problem. To date, we know of no noncomputable problem that can be solved by humans. So none of this analysis bears on the question of whether we ourselves might be nothing more than Turing machines or computations. Roger Penrose has speculated along these lines.





  • Excellent answer. Apr 27 '19 at 0:02
  • "If we knew that a human could solve the Halting problem" - depending on how you define "solve," we know that a human cannot solve it. If you require that a "solution" include a proof that it is correct, then a proof would have to exist for every possible instance of the problem. But if a proof existed for every possible instance, then a TM could exhaustively search all possible proofs to find it, meaning a TM would be able to solve the Halting Problem, which is absurd. So there are TMs which do not halt, and cannot be proven to not halt.
    – Kevin
    Apr 27 '19 at 19:02
  • @Kevin I fail to see why the halting problem, or p vs np, or any other time based conjecture would prevent AI from functioning. It didn't stop it earlier today when I asked my phone what planet Thanos was born on (with my mouth), and it gave me the answer in less than half a second.
    – Richard
    Apr 28 '19 at 19:50
  • 1
    @Richard: The question was "Can AI answer every possible question?" Exhibiting a single counterexample, no matter how esoteric and disconnected from everyday human experience, is a sufficient answer to that specific question. AIs are Turing Machines, because computers are Turing Machines. Arguably, human brains are Turing Machines as well. So anything a Turing Machine cannot answer, is also inaccessible to AI, and perhaps even to humans.
    – Kevin
    Apr 28 '19 at 23:25
  • 1
    @Richard: Any computer program "implemented by a Turing Machine" is another Turing Machine. That is basic computation theory, and I would really expect you to familiarize yourself with it before making claims of this nature. Regardless: It is certainly the case that AIs cannot solve the halting problem any better than any other computer program. I believe it is also the case that humans can't do those things any better, because I don't believe there are physically realizable models of computation which are more powerful than TMs (so the brain must be bound by those limits). 1/2
    – Kevin
    Apr 29 '19 at 0:43

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