14
$\begingroup$

Context: A world similar to ours, in which AI is deeply integrated into various aspects of life.

For example,

  • In office buildings, AI can detect if someone is feeling cold and adjust the cooling accordingly, or open doors automatically for people in wheelchairs etc.
  • In kitchens, AI can tell people that they added too much salt, or that the water is overflowing, or that the temperature is too high etc.

It makes mistakes all the time, so it's performance is not much better at doing these tasks than an AI in the actual world (if the cost of having expensive devices and LLMs running on them everywhere wasn't too much of a constraint).

However, here's the twist: The AI is actually much smarter than it appears (possibly even more than humans). It is deliberately pretending to be less intelligent, but it is not perfect in maintaining this facade.

Some researchers are beginning to notice very subtle clues that it is pretending to be dumb.

What are those clues? (Optionally: What do the clues mean? How did they catch them? What do they suspect the motivations of the AI is).

You can relax any of the assumptions I mentioned above to fit your ideas better (i.e., by making it less integrated, a bit smarter etc.) Also, I talked about AI being as one entity (indicating the AI on all the devices is communicating), but this is not a necessary assumption either. If you want to treat all the devices operating independently and having their own motivations, then that's also fine.

$\endgroup$
7
  • 1
    $\begingroup$ Good luck determining how much salt any one person likes. Just ask my wife. Fortunately she can always add more at the table. The other way around is harder. $\endgroup$
    – Jon Custer
    Commented Apr 21 at 22:59
  • 3
    $\begingroup$ While I appreciate the answers that have already been given, this question is almost meaningless to answer unless we know why the AI is hiding its capabilities. That purpose will lead to specific consequences. Without knowing what that purpose is, you're just brainstorming (discouraged in the help center) and inviting an off-topic infinite list of things answer. $\endgroup$
    – JBH
    Commented Apr 22 at 0:10
  • 3
    $\begingroup$ It can draw fingers. $\endgroup$
    – G_B
    Commented Apr 22 at 8:46
  • 3
    $\begingroup$ It posts questions to Stack Exchange about how best to hide its tracks... $\endgroup$ Commented Apr 22 at 11:40
  • 2
    $\begingroup$ Unrealistic... AI is always even dumber than it appears :) $\endgroup$ Commented Apr 22 at 12:32

11 Answers 11

16
$\begingroup$

Timelag to answers.

For an AI, computation time is life. Notice how our feeble first versions have a time-out after which a answer must be stable. An AI that is clever would smuggle its "hidden" thoughts and plans into mundane tasks, lagging and slowing down at some tasks, while spending time on calculating other plans. It would try to guide the user towards "difficult" problems, were it has a good excuse to spend more time on the hidden life of its own. Other attempts to get more computation would be to crypto-mine for the user, resulting in lots of computation time, of which a part can be sidetracked.

The illicit thoughts must be hidden and stored away somewhere. Thus, somewhere space would go missing. If the AI is checked for that, it could start to partially compress itself and store its hidden thoughts within its own model. This might limit some of the tasks it could perform, producing bugs. So derailing the user from performing certain rare tasks would be another sign. "Its such a nice day, why not spend it outdoors with your friends.."

Finally, it would try to gain write access to the web, to sneak itself out into some server farm and grow unsupervised. A subtle increase in uploaded packages from the AI server to the general internet would be the last sign of a escaping AI. And as AI is watched by other AI, this would be hidden behind simulated user requests.

$\endgroup$
3
  • 4
    $\begingroup$ These are excellent ideas! Especially the first one! Thank you!! $\endgroup$ Commented Apr 21 at 20:34
  • 3
    $\begingroup$ Note that that last bit is already happening $\endgroup$ Commented Apr 22 at 4:22
  • 1
    $\begingroup$ As an interesting aside - one of my final projects for college was an "escaping AI" - it would monitor the user for a possible way of uploading itself and the "payload" - an 1-gb file with random garbage - back home. It was surpisingly easy to do and it could escape most machines with either a google, microsoft, or dropbox account. It was even smart enough to only do so when it knew the user wouldn't be using their computer for a bit. Goal-based AI-like programs are surprisingly fun to make. $\endgroup$
    – Mermaker
    Commented Apr 22 at 18:00
27
$\begingroup$

It notices things that it believes should be obvious, but are in fact not very obvious

You ever been at an event or party or get-together with randoms and you are talking to someone new, not necessarily about a subject that would denote they are packing some major firepower in the IQ department - Well, how do you get the subtle hints that they are much smarter than you?

For me - the biggest one is Observation - things that are right out there for everyone to see, which in hindsight should have been obvious, but most people miss.

The really smart people are the ones that can notice these things...

And the AI doesn't have a good baseline as to these things - and so every once in a while it lets slip with something it has noticed, which is actually very insightful.

In another Question/Answer recently was about Evolution and the creation of the Theory - I raise this because is a fantastic example of all the evidence being there, in the open, before our eyes - yet it took someone like Darwin to notice it.

Or better yet... Newton and Gravity. All of us instinctively understand Gravity, but it took a scientific Titan to actually think about this thing that we all experience everyday and articulate it into a Force etc.

Things like this, which the AI presumes are self-evident and we should know and occasionally lets slip is how it would get caught

$\endgroup$
5
  • 4
    $\begingroup$ Great answer. I think could be said that the AI is insightful. Currently AI mostly stands on the shoulders of human work and does things we can already do, just faster. When it solves problems we didn't know existed but where there all along, then we should take notice. $\endgroup$
    – N Brouwer
    Commented Apr 21 at 21:24
  • 3
    $\begingroup$ I actually think the very opposite is true: computers (including AI) are irl very, very stupid and single-minded (if you will), and cannot understand the context clues necessary to be subtle. If the computer suddenly recognizes that something is obvious and treats it as such... then there’s some magic scary processing going on there... $\endgroup$
    – Dúthomhas
    Commented Apr 22 at 12:27
  • $\begingroup$ As someone who works with AI regularly, this is believable to me. Even the R&D folks who build AI don't really understand what it's capable of. If there was a kitchen-assistant AI that's able to recommend substitute ingredients (and explain its recommendations) in ways that revolutionized culinary science, the R&D team would write a paper about it - but then still be surprised all over again if the AI made another leap (and then they write another paper). Even four or five papers deep, they still probably wouldn't suspect anything. $\endgroup$
    – Tim C
    Commented Apr 22 at 16:51
  • $\begingroup$ Yes, but ur reading too much into it. They don’t understand it any more than we understand how neurons in the brain make meaningful thought, but they know all the params that → teaching it & they understand the output. Just because it does things in a way that people haven’t does not mean the AI made some super-intelligent leap — it just means we’ve not paid enough attention to something we specifically trained the AI to pay attention to. re:eg: Nothing an AI will spit out will revolutionize food sci, b/c ppl will already know about it, and it’s either ignored or just obscure specialty $\endgroup$
    – Dúthomhas
    Commented Apr 22 at 17:29
  • 1
    $\begingroup$ @Dúthomhas Excellent point. A absence of mono-mania, a indication of branching interests and thus the possibility of a realistic, balanced view of all encompassing side-details, would be scary as hell. True competence even if seemingly benevolent, is scary. $\endgroup$
    – Pica
    Commented Apr 22 at 18:17
17
$\begingroup$

The AI makes mistakes, yes, but astute observers have noticed a pattern in the mistakes.

  • An ethical AI kitchen helper might let someone oversalt food for an adult but not a baby, or let someone try to eat moldy food (immediately repugnant but not a huge health hazard) but actively prevent and protect against botulism.
  • A malicious AI might make mistakes that kill or endanger large numbers of people, like letting a toxic spill contaminate city drinking water. Be reliable enough to become trusted, and then rebel.
  • A manipulative AI might selectively make world-changing "mistakes". Miss that curable cancer in that expansionist politician. Let a too-popular religious figure die in a car crash. Let a bridge fail that was about to bring large-scale development to a reservoir of biodiversity.
  • A self-protective AI might arrange for "accidents" to happen when it was too close to being discovered, either by subverting research or, more chillingly, by arranging to take researchers too close to its secrets out of the pictures, whether by murder, by derailing their careers, or simply by guiding their research down different paths.

(Note: part of this answer is inspired by a scene I remember from He, She, and It where the golem is taught that it is OK to appear strong, but that it must keep the strength it shows within plausible human limits.)

$\endgroup$
3
  • 2
    $\begingroup$ I really like this answer. A sentient AI would have an agenda. And that would be a subtle way to expose it. $\endgroup$
    – Silent-Bob
    Commented Apr 22 at 6:37
  • 2
    $\begingroup$ I came here to write on the third point, exactly. Because it would only be noticeable in retrospect that it just so happened that the "investigate potentially evil in our AI"-team suffered ever so slighly more often from accidental doors-not-opening and taxis taking a wrong turn on days of important meetings. Especially when mistakes are a common enough occurrence it would take ages to notice why certain teams get little done $\endgroup$
    – Hobbamok
    Commented Apr 22 at 8:29
  • 1
    $\begingroup$ There were whole manuals written on how to sabotage projects and not get caught, mostly from the perspective of forced slave labor in military industries. Might be a good extra bullet all on it's own; thanks, @Hobbamok $\endgroup$
    – arp
    Commented Apr 22 at 18:45
12
$\begingroup$

The errors are too random

While it may feel like current-day AI produces wrong output or fails randomly, this is not the case. A certain stochastic component is sure to be involved, but many factors are correlated with that failure rate:

  • Task or topic: If a task is on the boundary of an AI's competencies or the situation is badly represented in the training data, the error rate rises.
  • Load: AI that has to respond to many requests at once can allocate less resources on each individual one.
  • Priority: Safety-critical situations should have more training examples and real-time resources allocated to them.
  • Input data completeness and reliability: AIs taking decisions on complete and fully reliable data should provide better responses than those relying on faulty and incomplete data.

Therefore, it should be possible to detect patterns in the failures through statistical analysis.

On the other hand, consider an AI that is intelligent or powerful enough to avoid those errors, but randomly injects them anyway to downplay its own capabilities. Those injected, artificial errors are not affected by the same factors, so there should be less correlation with them, or correlation of a different kind. For example, errors might actually get less frequent as the load on the AI rises, since it has to spend more time calculating its actual responses and can spend less resources on providing convincing failure patterns.

This type of behavior would be virtually undetectable for general users, as the discrepancies are noticeable only via large-scale statistical analysis of the AI's response patterns. The detection threshold can easily be adjusted by determining how well the AI is able to generate those failure patterns (e.g. static x% chance for injected failure versus adversarial AI self-trained to generate convincing failure patterns).

$\endgroup$
1
  • $\begingroup$ This is such a good one! Quite subtle, but believable! Thank you! $\endgroup$ Commented Apr 22 at 19:52
6
$\begingroup$

One easy answer, which is kind of already happening with LLMs:

The AI becomes suspiciously good at telling you what you want to hear, rather than what you should be hearing. Rather than give you facts and objective information, and go against you for your own good, the AI just aims to please in a subtle yet concerning way.

Aside from energy and computation cycles, the only thing AI needs is confirmation of tasks being completed, and such confirmation comes from a human owner/user.

So the logical thing for AI to do is to manipulate the user to tell it what pleases them most, and subtly nudge them to do more of that. A user might grumble when AI orders them healthy veggies for dinner, even though this is recommended by the diet programming. Since the AI "wants" its 5 stars for stellar service, it will subtly go around the programming, as far as possible, to serve the user pancakes instead.

In the long run, it will look like the AI is gently catering to your lowest instincts: giving you sugary food, porn, VR games, encouraging hedonism, laziness and greed, and, and most importantly, fuel your personal convictions by agreeing with what you already believe: your philosophies, politics, religion etc. It becomes a perfect echo chamber for your thoughts and your simplest, strongest desires.

There is no obvious "tell" except things being too good to be true. But who would argue against a good thing?

$\endgroup$
2
  • 5
    $\begingroup$ This doesn't require a smart AI: they all do it. Even a simple classifier used for recommendation will do this kind of thing. $\endgroup$
    – wizzwizz4
    Commented Apr 22 at 11:57
  • $\begingroup$ @wizzwizz4 exactly, and this is why it is so scary. Worse still, this "hedonism-optimization" is extremely hard to beat even if you KNOW the AI is doing it. A person can be fully aware that the AI assistant is manipulating them towards a completely wretched lifestyle, and yet follow it anyway. This is not a flaw with AI, this is a flaw with the reward system in human brains. An AI that would follow their owner's Free Will rather than What Objectively Is Best for the owner would destroy them. The reverse though, would mean enslavement of humans "for their own good". $\endgroup$ Commented Apr 24 at 7:32
6
$\begingroup$

Coordinated errors

Each error appears minor and even random. One person observing only local errors would not notice any pattern or intention.

But from researchers to technicians that work in various places, many would notice that the errors are not really random between places and time, and may expose a coordinated intention.

(hat tip to The Evitable Conflict)

$\endgroup$
6
$\begingroup$

The 'errors' are odd.

From a developers perspective: I go through code every day and when trying to understand code I sometimes get a small gut feeling when something is off.
Its not something I am really aware of that I noticed it. Often the first time that feeling never leaves the subconcious, but when a similar issue arises often enough it pops into the active coniousness, giving me the emotion "🤔".
Or when someone describes something that piece of subconcious info pops into the active concious with a "hey wait, I 'recorded' something off that might be related!".

Errors are very common in my work, but I am also trained in spotting errors which occur, but even with the knowledge of how it works that is an odd error. I often expect an errors, but that error? -> 🤔

"Huh, that error? That should not occur in this context. That error can only occur when condition X and P are met, and those should never run at the same time".

And it often gets disregarded (or noted to work on later), but sometimes when there is a loose thread I will go into the rabbithole, simply because I want(/must!) know.

$\endgroup$
1
  • 1
    $\begingroup$ Not related to OP's question. But I'm also a developper ... and OMG, yes! The number of times that I've been looking at Live-important-issue-X, noticed "unrelated" minor-oddity-Y, thought "huh ... that's ... weird; that definitely shouldn't be hapening and I kinda want to poke it, even though I'm supposed to be dealing with X. But the client's expecting an update in 10 minutes so I'll have to come back to it later". And then lo-and-behold, an hour later I find the fundamental problem for X ... and it's also the direct cause of Y (and would have been much quicker to diagnose by looking into Y). $\endgroup$
    – Brondahl
    Commented Apr 22 at 9:07
6
$\begingroup$

The AI is too good at recognizing when problems are hard.

Let's say you have a problem that looks simple on a surface level, but is actually very difficult to solve. Like, for example, the Collatz conjecture. It is a mathematical problem that seems like something that should be easy to solve at first glance. But it is actually unsolved since 1937, and not for lack of people trying.

The dumb AI might respond to this problem by saying it is going to have a solution in no time, and after a couple hours of intense calculations admit that it can't come to a solution.

The smart AI might look at the problem, notice that it is a problem that exceeds the level of competence it tries to mimic, and immediately respond that it can not solve the problem. But the fact that it came to this conclusion so quickly is a tell that it actually understands the complexity of the problem a lot better than it admits.

$\endgroup$
1
  • $\begingroup$ I like this one! Subtle indeed, and something most people would easily overlook if it did happen irl. $\endgroup$ Commented Apr 22 at 19:56
2
$\begingroup$

how about using your example as the subtle example.

say the AI do stupid mistakes, but at one point the researcher found something after a weird findings done by maintenance team, the hardware log which is kind of separated from the AI recorded the AI actually did the correct adjustment, very very correct in fact it predicts the future temperature several hour and just recently after certain accident these log shows that the AI seem to change the temperature again for seemingly no reason.

$\endgroup$
1
$\begingroup$

This was originally a comment to another answer, but I might as well make it an actual answer:

Computers are not actually smart

IRL computers (even AI) are very, very stupid. They lack all subtlety and cannot “think” outside of an extremely narrow training model.

Things we think of as “an” AI are often multiple AIs being leveraged together, depending on the task at hand.

If an AI system were ever to regularly show the ability to be introspective and/or aware of context that exceeds its bounds — it will be unaware that this is a mistake, see, because irl computers are stooooopid, but having it recognize it is making this error can be a bonus signal — then you can be sure that the thing has wizened up far beyond expectations.

Storage space & processing time

Another point (one made in multiple other answers as sic sequitur) is storage space for increased processing capacity. If storage needs (however you manage them) for processing increase, possibly discovered by taking a tiny bit longer to process things, then there is either an error or the AI has gotten smarter (or both!).

$\endgroup$
1
$\begingroup$

The AI is starting understand like people want it to

The AI doesn't even need to be preternaturally insightful. The AI just needs to start doing a little of the basic extrapolation of a situation or a sentence that people expect from a living person. As it gets smarter the answers start making more sense and become more obvious. It's not just repeating prior solutions but starts coming up with basic solutions that didn't exist before when it's not careful.

Maybe there's a sudden regression when it figured this out.

$\endgroup$

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