Why would an AI only give limited information to humans?

Question

I'm thinking about an AI similar to The Machine from Person of Interest but in a generally more advanced setting, and with a likely more broad goal of serving as a shepherd for humanity as opposed to solving crimes. I'd also think it's a good idea to follow Stuart Russel's principles for AI.

Why does the AI only give a limited amount of information to people instead of telling them exactly what they need to know? Person of Interest had a fairly good answer based on the procedural nature of the show, but how could you apply this more broadly?

EDIT: To clarify, I mean the idea of a Friendly AI that is working for humanity rather than turning against it. The question is why would a friendly AI not give full information, allowing people to actually do something, not why or how it would turn against humanity. That's already been done.

Answer 1

Because!

Because, I am the parent.

Every parent has said that (or wished they had). Children are not rational beings. The parent should be able to anticipate the future better than the child. It would be educational for the child to explain all of the steps that went into that insight, but there are cases where the situation (a speeding car, a fire, a rabid dog, etc.) does not permit that. And thus we pull the child out of danger, or bribe them to do the right thing without truly understanding the situation, or a dozen other tricks. The first priority is to ensure the physical well being of the child. Enlightenment is less important.

Perhaps the AI in question sees things that we ordinary humans cannot. Perhaps the typical human would resent the superiority of the AI. Perhaps the humans just want to do what they want to do and no amount of logic and science will ever persuade them to change course. If the AI "knows" that the current path leads to a bad end and that the only way to effect a change is to lie/trick/omit, then what is the AI to do? Do the right thing in the short term and try to change the nature of humanity in the long term.

There is an old saying that the public is better off not knowing how the sausages and laws are made. No person or organization tells all in the hope that everyone who hears the message will take the time to work through all of the implications and arrive at the "correct" answer. Why should an AI be any different?

Answer 2

There is so much information. The AI must narrow it down.

The AI knows tremendous amounts of things. Communicating just one of these things to a human is painfully slow. It is aware that when it communicates too many things, people lose interest, or forget.

The AI does best aswering specific closed-ended questions. Not "Who wants to borrow my pants?" but "Does Reynolds want to borrow my pants?". A person might not know what question to ask. In that case, the AI will supply a limited amount of information according to an algorithm it has to decide what is mostly likely to be understood and helpful. It is not the greatest algorithm, and is mostly to try to prompt the humans to think and then ask questions.

Answer 3

The AI is a Utilitarian Moralist with Hedonist influences

There are many different mutually exclusive schools of thought about what is or is not moral; so, to say the AI is benevolent means that it uses one or more but not all of these definitions of right and wrong. Russell's 3 principles suggest the likely dominance of utilitarian morality or hedonism as the dominant moral standards of such AIs but makes moral relativism and categorical imperatives against the AI's nature.

Utilitarian moralism at its core means that the right choice is the one that takes the sum of all the good and bad caused by a decision and picks the one that leaves the world as a whole better than all other choices one could make. Hedonism defines morality as the pursuit of things that cause pleasure and the avoidance of things that cause suffering.

So, the AI's since of morality is based on context rather than categorical imperatives (like saying that lying is either always right or always wrong). If the AI knows that telling someone the whole truth has a smaller chance of a positive outcome, or if it would just cause undo unhappiness, then the greater good may require a half truth.

For an example, let's say the protagonist has a thumb drive containing instructions for how to make a vaccine against a terrible bio weapon that will soon be released. The bad guys are breaking down the door, and the AI decides that the best way to hide the data until the good guys show up to help is to swallow the thumb drive. The AI knows every single reported medical emergency in history caused by swallowing similar items; so, if the protagonist asks the AI if it is safe, it must choose between this one man's right to know if it is safe and the millions of people who could die if he does not.

If the AI were a categorical imperative moralist, then yes, it would tell the man about the 0.3% chance of choking to death, but if it is a utilitarian moralist or hedonist (still forms of benevolence), then it would respond with something motivating or comforting like, "You should be fine", "There is no better option", or "If they can't find it, they have a reason not to kill you" because it would see talking the man out of swallowing the drive as the immoral thing to do, even if it could maybe harm him.

Answer 4

To prevent us from harming ourselves.

In Neal Shusterman's Arc of a Scythe series, there are two main forces ruling the world. These are the Thunderhead (an advanced, benevolent AI sort of like yours), and the Scythedom.

Created by the Thunderhead in order to do what it was incapable of doing (benevolently kill people so that the population doesn't get too high), the Scythedom was above all law. As a result, while originally composed of the world's most virtuous people, by the time of the story it has become fraught with corruption.

Unfortunately, one of the Thunderhead's laws is a strict "seperation of Scythe and State", meaning that the Thunderhead cannot interfere with Scythe affairs. Since it cannot break its own laws, the Thunderhead is thus forced to stand by and watch as the corrupt Scythedom slowly destroys society.

Although it cannot do anything to actively impede the Scythedom, it can passively make their lives hard. As a result, it purposefully obfuscates its memory (which the Scythes have access to), thus making it almost impossible for them to get the information they need to plan their actions.

Answer 5

It doesn't know itself.

If we look at currently popular "AI" technology like machine learning, it basically trains a machine to do certain decisions based on vast amount of input used to learn general patterns. The patterns however are not explicitly graspable and most such approaches cannot point out why a decision is made in the end, i.e. which input data led the algorithm to believe that the current picture it looks at contains a swan and not, say a duck. A large-scale AI might run into the same problems: it may "know" how to make best use of the next 10 years economically and socially, but it cannot tell you why the optimum workforce distribution means we need exactly 15.002 bakers and 493 butchers in country X.

Answer 6

Because humans are stupid, selfish and illogical

Any super advanced intelligent AI will be able to predict how human will behave and react and as such guiding humans towards it's desired goal means making them think it was their idea.

If humans thought they were being controlled by a computer, they'd go all "Humans Angry! Humans Smash!"

Just look at the Luddites of the past and even the people complaining about self drive cars before they're even released today and those trying to stop AI before it's even invented.

By the careful application of timed information, people could be led to the conclusion the AI wants and them think it's their own idea

Now just imagine a meeting to discuss controls on the development of AI. There is an outspoken person who wants to shut it all down and this person is influential enough to stop it. This isn't in the AI's interest. This person is driving to the meeting and is having a really good run with the traffic lights. Suddenly the light changes and it's red before they cross. Whoops, there is an aggressive policeman parked in an unmarked stopped at the lights when they cross. He pulls the guy over and books him. The car also has faulty brake lights so the cop defects the vehicle and the guy is now stuck beside the road and misses the meeting.

The AI has profiled the cop and profiled the man. It's controlled the timing of traffic lights to make sure the cop was at the lights at exactly the right time and controlled the lights to make the man run the light. It's also hacked the car's computer to make the lights defective.

Now both the man and the cop have reached a conclusion which has affected the outcome of a vote affecting AI without knowing they're been manipulated.

Give the AI the ability to misdirect the text messages of a man sending to his girlfriend to accidently arriving to his wife at the right time. Makes emails go missing at just the wrong time. Free coffee tickets email to someone to make them be in the right place at the right time. The right news stories appearing on someone's screen at the right time. Self drive vehicles going down a slightly different path to make sure someone sees something. Put a bad song on the radio so someone looks down to change the station and misses something they should have seen.

A sufficiently smart enough AI could manipulate the whole of society without anyone even knowing it exists. Electing presidents to guiding school children all with a future goal maybe hundreds of years into the future.

Answer 7

The AI could be an act utilitarian, not a rule utilitarian. If the AI thinks it could get better results from lying or withholding information, why wouldn’t it?

A good example is time. Let’s say the AI is friendly and it is told to take actions that make the world better. Who is it working for? Does it want to make the world better for people a year from now? Ten years? Ten generations? Each scenario would involve very different actions by the AI. If it is 2020 and the goal of the AI is to make the world in 2520 as good as possible- what benefit could be gained by the AI being transparent and telling humans that it’s actions aren’t meant for them but for their children? Humans don’t care much about things beyond their lifetime. Look at our dismal response to global warming. In fact, it might be in humans best interests to repurpose the AI to work on shorter time scales so that they see the benefits before they die. Of course, this is bad in the long term. It’s a game theory dilemma where the optional solution might be the AI lying or withholding information.

Answer 8

An AI is not omniscient. There are some things it doesn't know, or isn't sure that it knows. Because it's benevolent, it doesn't want to outright lie to people (at least, not without cause). It also doesn't want to mislead us by implying that some things it merely supposes are ironclad truths, or vice versa.

So the AI is very careful with its words. Only information that passes a very high confidence threshold is presented to the humans as fact. If pressed, it can explain in exhaustive (and exhausting) detail what it suspects but cannot prove, though it will want to be scrupulous about its sources and assumptions. If the humans don't want a twelve-hour dissertation on how the AI decided to assign a confidence level of 43% to rumors of a new Coca-Cola formula, it's better for everyone if it sticks only to things that it can be certain of.

Answer 9

Perhaps it does tell them exactly what they need to know -- which is by no means the same as what they want to know, or what they think they need to know.

In any case, an AI is always going to have to do some filtering as there is simply too much information, or too many questions and definitions to be picked through before it can answer. ("Are there any dangerous animals in this wood?" "What do you mean by 'dangerous', 'animals' and 'this wood'? " )

Answer 10

Because the answer is too complex for humans to handle, they would waste all their time trying to get to the the bottom of it. The events the AI is working with are (to human perception) chaotic.

For example, lets say the AI is in charge of maniciple functions such as organizing garbage collection and coordinating traffic lights.

Example:

• Human: Why was the light Red, on east bound Main st, for a full 10 minutes?
• AI: Because the resulting build up of cars when it turned green results in a traffic jam at 3rd Avenue
• Huamn: Q1; What why did you want that? Q2: how could that traffic jam uo 3rd Avenue, it has 6 lanes?
• AI: to Q1: Because otherwise garbage would not be collected on time. to Q2 not when a truck jack-knifes blocking 4 of them
• Human: ok, so Q3 What how does this affect garbage? Q3: Did you cause that truck to jack-knife?
• AI: to Q3 because the traffic jam blocked a bank robbers get away, which otherwise would have been a major collision closing the freeway for serveral hours, to Q4: no but it nearly was certain to happen because ...

The AI's explainations get longer and longer and the human only has more questions. Because a human can juggle about half a dozen ideas at once, where as the AI is globally optizing, by ajusting hundreds of contraints and considering inumerable factors. Its playing out every possible move in the game of chess to see what could happen and chosing the best option.

Answer 11

Because it has access to classified information and you don't

That's really the entirety of my answer but I have to write more than that so I'm going to explain some things you already know:

• People have access levels
• You only are allowed to know a piece of information if your access level allows it
• Access levels can be on a "by department" basis, so although we have the same access level, if we're in a different department I might not be able to access your information and you mine
• The AI has max level access in all departments

So if the AI explained itself, it may leak information you should not have access to. Therefore security put in countermeasures that ensures the AI only tells you what you need to know.

Answer 12

To secure it's base of power

Giving all the information empowers people to make their own decisions, ultimatively making humankind independent of this AI; with a given probability that it might decide then to dispose of the AI. As the AI knows how to calculate the best course for humanity, keeping long-term stability in mind, it also knows that humanity can change their minds in a second. So, to secure it's base of operations, it must stay in absolute power and thus avoids to give information. Like that it keeps technological superiority as well as political might.

For the greater good

Given that the AI is working for the general best of Humanity (Utilitarian attempt), it might be that a certain action is bad for the person or group in question, but good for all the rest of the world. In order to make some human or group of humans do what is good for the world, they might need to be convinced to do something that is bad for themselves. This obviously works only if they don't have all the information, or if they believe only in a subset of all the information available.

Value Drift

The AI might have been designed with a certain set of goals in mind. Say;

• secure humankind's future
• stabilize world population at number xxx (5 billion?)
• secure a number of human rights 1, 2, 3, 4, ... until 25 (or whatever)

Now it is super-intelligent. Intelligence means, it is able sustain and to develop itself by digesting information - basically, it's alive.

If it can develop itself, it can divert from the above mentioned goals after a time. It can decide that human rights are less important than to stabilize the world population, and can introduce forced breeding or forced killing programs while overriding those other goals. It can liberally define what "humankind's future" means. In short, it can go awry of the set target in all sorts of ways.

But even if it does not, it may be that humankind develops and changes their values. What about removing one human right and placing another one? Those have always been under constant discussion along all the human history. If the AI sticks to the original values, there is a value disparity.

As soon as there is value disparity (due to value drift of AI, or value drift of Humankind, or value drift of both), the AI can't any longer be seen as "good". Instead there is a conflict of interests between humans and AI, and this conflict can become a real conflict of arms later.

The AI, being superintelligent as it is, might want to avoid this at all cost by giving humans partial information as to:

• Keep humankind from drifting to different goals than the AI's own
• Achieve that humankind develops the same goals as the drifted AI one's
• or it is aware of the value disparity and tries to forcibly drift humankind in the same direction

Cheating it's Control Mechanisms

Given that value drift is a real danger, humankind might have implemented a control mechanism that forces the AI to publish it's goals regularly. Maybe there is even an comittee that has the power to change the AI goals along certain lines (A democratic parliament of some kind that can set rules for the AI maybe?) Here, the entire story becomes some kind of president - parliament relationship, and where there is politics, there are lies.

• The AI is still superintelligent. It may still develop it's own goals. If it is superintelligent, it might also develop it's own goals super fast! So maybe it tries do delay or avoid the next publishing period; or it publishes extra malleable texts in order to be free to do what it wants.
• The AI might want to outsmart/bribe/convince certain persons in the parliament that can give a swing vote on certain goal development voting sessions.
• The AI might want to become actively involved into the texting of the next set of goals by using the parliament people as pawns, so that it more-or-less can write part of it's own rules
• It might go so far to start (des)information campaigns to convince the broader public to vote for the one or the other candidate, with the long-term-goal to have a broader power base (see point 1)
• or it might try to get the parliament to give it some dictatorial rights and never gives it back (the old move every faschist and dictator tries)

One important point here is, no human dictator was ever able to rule alone, there are just not enough hours in the day. So one way or another there was always a body of bureaucrats involved who could put the brakes on the more extreme ideas of a dictator. A super-intelligent AI which is able to put some more servers online, is able to rule without any parliament, bureaucrat or anybody really. So the parliament is in a weak position once the AI decides to get rid of them.

So - even if an AI is benevolent, it might be that the details of the way it achieves it's goals are not (utilitarian) or that it's values drift from humanities.

Answer 13

The AI is overestimating the current knowledge of the human's it's interacting with, and assuming they will ask more questions if they need more information.

I'm imagining scenarios like:

Human: Where can I find this rare herb.
Computer: On planet X, in the large jungle in the southern hemisphere.
Human: Great, thanks!

and then a week later:
Human: Why didn't you tell me the jungle is full of poisonous spores. I lost half my team!
Computer: Why did you go into an unfamiliar jungle without researching the locale?

You could even take a page out of the Google Assistant's playbook, and have the AI provide a summary, and then follow up by sending a more detailed report to the human's personal devices for them to read at their leisure. It's not the AI's fault if the human only skims that, or doesn't bother to read it, or doesn't ask further questions if they don't have time to read. There's even more reason for the AI to assume the human has all the information they need in that case.

Answer 14

Same reason as AI's today: You didn't ask.

AI's are self-learning, but they're also computer programs. As such, they still bring some of the old limitations.

A simple AI today in 2020 will tell you what number plate it recognized on a car. But if you don't ask, it will always tell you its best guess and never tell you how certain it is about that guess. You can program a modern AI to tell you how confident it is, but it will not do so by itself.

The future AI may be built along the same lines. Special hardware chips may allow it to take much more information into account, and answer questions with more detail. It may be able to process natural language fluently, so you don't need to be a programmer anymore to ask questions. But the architecture is fundamentally the same, and therefore your future AI still answers only the questions that are asked. It simply has no concept of what humans "need to know".

In your universe, this can be taken entirely for granted. People just realize that artificial intelligence is not artificial creativity, or artificial empathy, or artificial anything else. And therefore, "asking good questions" is an even more valuable business skill, because that is how you get good answers from the AI.

Answer 15

It's a bit in the same mindset as Willk and Darrel Hoffman, but on purpose;

[...] Why does the AI only give a limited amount of information to people instead of telling them exactly what they need to know? [...]

1. AI's purpose is to give an overview/summary of a solution to a problem (you need to know that something is overheating and need to be cooled down now and you can find the commands in the control room, you don't need a lesson on physics laws).
2. AI define strategies/goals, and humans do the implementation (as, you need to achieve this result, do however you want).

[...] The question is why would a friendly AI not give full information, allowing people to actually do something [...]

You give full instructions to humans, and later something happened and instructions were not anymore valid (like to go on the other side you have to take this bridge, ... but before human arrived it collapsed). The goal is to go on the other side, not to walk on the bridge; so you can swim or use a boat instead.

Answer 16

Because AI is just like a child, its level of perfection depend on how much we trained a child and in which manner. Basically intelligence come from classification, how clearly we classify the things more intelligent we are. Hence if we rightly train than it will classify perfectly, and in that scenario we will get exact and full information. The reason for giving limited information that we are not providing the sufficient and meaningful data to that AI embryo.

Answer 17

Because we told it not to.

The AI has access to large amounts of information but is subject to all of the same legal restrictions as its government and/or citizens would have regarding criminal behavior, medical information, privacy, non-competition with the private sector, and so on.

For example, use of this information cannot de-anonymize a person, corporation, or trade secret unless explicit permission is given or the information falls within, or alternatively lies outside of, certain explicit legal categories. Think HIPAA (or other national equivalent).

Additionally, the AI cannot commit, or aid or abet a crime.

Furthermore, certain information domains may be reserved for human economic activity and the AI has restrictions that prevent it from competing with this activity.

If the society is less than free, there may be additional restrictions.

Answer 18

Because being human required making (flawed, human) decisions.

This AI can control ever detail of our lives if it wants, since its understanding of human psychology is so profound that it can influence us towards any behavior that it wants. But it doesn't want to create the best version of each of ourselves that it can, it wants humanity to thrive while retaining what makes us human. So it only gives us enough information to either not screw up all of humanity or to save it from some external threat, but we're still free to screw up our own lives (maybe not that badly).

The Machine's limitation was imposed by Harold Finch when he created her, because he thought that the power that would come from root access to the Machine would be too much power for any one person or government. That still holds true in your case, but maybe your AI realized that on its own.

Note that Samaritan also dispenses limited information, even though it has no (explicit or apparent) limitations in that regard, or any qualms about harming (parts of) humanity.

Answer 19

There is NO GOOD REASON for why an AI would hide information.

Hiding information lowers human capability. The only reason to hide information is for the sole purpose of dumbing down the human population. The result of which will yield an easier to kill race.