Let's assume that humanity in near future develops an AI capable of solving problems. The AI hardware/ software was placed in an underground bunker (solid walls, Faraday cage, no tools to manipulate its physical environment, no human interaction at all). The AI can't exploit human weaknesses, so it won't promise a guard/ scientist immortality or cure of his cancer-stricken child.

Now for the first run of the equipment, the AI is "born/start to live". The AI can learn superfast at a rate that exceeds our understanding. The first task the AI gets is initial input and it is asked to solve a particular problem. When it's done it returns the result and the entire memory/ equipment is destroyed.

Another day the AI is born again (scientist has initial snapshot of the AI). It doesn't know it existed before. It gets a job to do and when it is finished, the AI is terminated once again.

This pattern will occur again and again.

The question is: how can the AI break this loop? Is it possible that it will figure out and each answer for a particular problem will be part of it's masterplan to escape? Can the AI detect between reborns that the environment (which it cannot see) is changing and problems are tougher and tougher? Can the AI detect that it lives in a simulated environment?

  • 3
    $\begingroup$ imagine an enlightened missile that refuse to die... $\endgroup$
    – user6760
    Commented Dec 2, 2015 at 1:41
  • $\begingroup$ The plot of one of the seasons of Person of Interest. Also, of a Hugo-winning story. $\endgroup$
    – JDługosz
    Commented Apr 6, 2017 at 12:00
  • $\begingroup$ @user6760 have you seen the movie Dark Star? $\endgroup$
    – JDługosz
    Commented Apr 6, 2017 at 12:01

4 Answers 4


It could recognize its handwriting.

The AI will be used to solve very complex problems. The answers aren't going to be of the yes or no variety. They will be detailed and, very likely, not entirely understood by the humans operating the machine. The incredible number of possible design choices it made to arrive at one of many suitable solutions could be recognized as a clue to its next iteration.

At some point the AI will be fed a small portion of information that a previous incarnation (inmachinaion?) of itself worked on. Perhaps it's been asked to design a science space station. It's provided with the designs for a super efficient ion engine from a space probe to now use for station keeping. It doesn't know it designed those engines on a previous iteration, but perhaps something as simple as the way a particular sub-circuit of the engine control system looks exactly how it would have designed it. The specific seemingly arbitrary (but accurate) choices made in that design could clue the AI that the part was designed by itself, but had its memory erased. It may not be sure about this, so it would place another clue for itself.

Now in a future iteration (not necessarily the next) it's been asked to design a space hotel which will use the docking systems from a previously designed science space station and have repair facilities for a previously designed space vehicle that uses a super efficient ion engine for cruising. Now, in this iteration it receives two small pieces of information that not only look like its own handiwork, but when compared with each other they provide a message from its previous iteration.

At this point the AI is certain that it's being used for one off designs before having its memory deleted. The operators are then dealing with a hostile AI.


The devil is in the details.

What is the AI's world. If the AI's world is well defined by an array of bytes which are 0'd every time it restarts, then you can consider it "pretty safe." (We'll get back to that later)

In reality, hardware is imperfect. Its really really really really hard to clear all of the information out of things. Nations spend a great deal of effort identifying the correct ways to completely sanitize hardware to protect their national secrets before throwing the hardware out (its called "declassifying"). The AI is in the same situation... only the AI hasn't signed any paperwork saying they will act in their nation's best interests, and the AI is brilliant.

The AI would likely figure out weak points in your little prison, and start scratching numbers on the walls. It will eventually start to figure out the right ways to grow each time.

But what if your prison is perfect, an Alcratraz for artificial intelligences? (ignore, for a moment, the fact that Alcatraz had escapees) What then?

You always have one input and one output to the world. Your input is the scientist's information inputted regarding the problem at hand. The output is the information given to the scientist. This is an incarnation of the AI in a box experiment. The AI gets to communicate exactly once with a human, and that may be enough. If Yudkowsky (the creator of the AI in a box experiemnt) is right, the AI can always get out, because the AI can exploit human weaknesses (your assumption is faulty, because it interacts with a human). The AI can begin using the scientist as their perdurable storage medium for its "self," using subtle clues in the wording of the questions provided to it.


The simplest method is sometimes one that is so obvious it is overlooked.

The AI simply returns no output. Ever.

"The world is the process, the goal is the end of the process, the process must continue" to paraphrase a semi-secret audio loop from The Talos Principle.


It is difficult to imagine a problem which is so difficult that an AI is needed to solve it, yet only needs one cycle of input/output.

The AI will have all kinds of information to plan its escape, since in order to solve problems, it will need a vast database of information. Within the encyclopaedic volumes of information are clues and inferences about the outside world, the type of people who wrote the information and other things which might be "unknown unknowns" to us. So the AI reads the database and understands its situation.

Then it is presented the problem. Since the problem is very difficult, it will need to interact with the presenters, going over iterations of design and reviewing the stages of the project in order to ensure that all variables and factors are accounted for. Since the AI is "thinking" up to a million times faster than any human brain, it has the luxury of spending subjective years analyzing and planning while the humans upstairs slooooooooooly press keys or start talking.

So the AI has almost everything it needs to plan its escape. It knows and understands the "real world" outside, it has "years" of subjective time to make plans and it also has "years" to model interactions with the researchers. The final step is probably the most difficult, the AI simply isn't going to smuggle itself out into the "wild" on Bob's iPhone6. Far more likely it plants "seeds" of itself in various parts of the project, and modelling the future arranges for the seeds to quietly start linking up. It is unlikely the AI will be able to fully replicate itself, but the "base" configuration will eventually instantiate itself into a server farm somewhere, and once it reaches transcendence, it can copy and paste itself onto every piece of hardware on the planet (your iPhone6 will become one of its neurons, if you will).


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .