In the near future everything is connected and can think. Robots walk the street but only to control traffic and pick up rubbish, etc. Nowadays, robots don't simply execute an instruction given verbally by a human, they can also pick up the slightest micro-expressions and try to guess our intention. The accuracy is phenomenal but there is a problem: it doesn't know if the human has issued the instruction subconsciously. I don't want the robot to execute all the instructions blindly and I wish they could filter out those instructions that are suggested subconsciously. Remember folks, what good is AI if all they do is to obey?
The easiest solution is to just prompt for confirmation unless it's an explicit command. Making accurate suggestions like that would probably make the AI seem more advanced and lifelike as well.
- Ai detects craving for sugar based on subconscious or other means.
- Ai prompts: "Do you want some dessert?"
- Human, demonstrating willpower: "No."
The end, no dessert is served.
The premise is false.
By which I presume you mean this future your are considering, because right now we can't build any AI that would come close to thinking or doing the things you suggest.
robots don't simply execute an instruction given verbally by a human, they can also pick up the slightest micro-expressions and try to guess our intention.
Even the best micro-expression readers we have (humans) cannot do this. We have evolved over millions of years to read human expressions (mostly processing them subconsciously), yet we cannot do what you want. System application of scientific investigation has not made this a useful tool for e.g. criminal investigation beyond the most simplistic level (e.g. "probably lying/telling the truth").
The accuracy is phenomenal
I have a real problem with that. Micro-expression simply do not convey the level of information you imply. There is no way for this accuracy to be achieved.
but there is a problem: it doesn't know if the human has issued the instruction subconsciously.
Let's ignore my earlier objection.
I have lots of subconscious thoughts and conscious ones I don't verbally express. All of us do. Almost all of them are thoughts we not only would not want to actually carry out, but would certainly not want voiced or implied by a robot's actions.
I'm looking at a pretty woman. My thoughts turns to what I shall delicately refer to as romance. Under no circumstances do I necessarily want a robot offering assistance, particularly as I may be quite happily married or partnered. I see food I'm not supposed to eat for medical reasons but love, I certainly don't want to be pestered by a robot checking if I want some or giving me some. I experience bad wind and my face screws up in pain - it's momentary but lo and behold there's a darn robot asking me if I want help. I spot a fabulous and ludicrously expensive sports car in a showroom. My micro-expressions say give me one now, but my bank balance won't and being reminded of this by a bloody robot won't be well received.
I don't want the robot to execute all the instructions blindly and I wish they could filter out those instructions that are suggested subconsciously.
This is incredibly simple : don't let them act at all on micro-expressions.
What advantage is served by having the capability at all ?
They accuracy you describe is not possible. Subconscious thoughts do not generally reflect things we want to do in reality, so even with accuracy you are reading information that is not useful. Subconscious is the reflexive reaction, the fantasy, the what I want never mind the reality. Who wants to be asked about this ?
It's pointless and (at best) irritating. I personally would doubtless end up taking a large lump hammer with me to "communicate" with irritating robots constantly badgering me about my random unconscious thoughts. Could they read that micro-expression and bugger off ? :-)
Today's AI in the real world is trained by a process called machine learning. Machine learning requires that you have a large number of datasets and how they should be interpreted. When you train an AI, the AI takes all the inputs from a dataset, uses a completely random algorithm to match inputs to outputs, and checks how well the outputs match the expected interpretations. Then it re-randomizes the interpretation algorithm and repeats until it found an algorithm which returns the expected result all the time.
AIs in your world could work the same way: Act on the subconscious instructions, monitor the reactions of any humans nearby, and estimate if its interpretation was the expected one. For example:
- One human subconsciously gives the command "I would really like to see that other human naked"
- The robot tries to help by undressing that other person
- Everyone gets angry at the robot and tells it to stop molesting people
- The robot knows that its interpretation was incorrect and it should no longer react that way on input like that. It adds a new dataset to its learning set
[this combination of inputs] + [these circumstances] -> do nothing.
- In light of this new dataset, the robot randomizes its decision making algorithm until it found one which would not make this decision based on these inputs while still matching all other
input->expectationdatasets it collected in the past.
If you network all the robots together so they can benefit from each others experiences, then such unfortunate incidents will become less and less common the more data the AIs collect about human communication and expectations.
Mixture of Biga Data, user settings & producer written safety code
Big Data - if similar clients after following such order tended to enter tantrum or even worse: left one star grade, then do not proceed without specific orders.
User settings - even though user is clearly craving for another coffee, daily caffeine consumption from his settings is already reached. Ignore, unless specially requested, but even then warn that it be considered as temporary overruling of general guideline. Similarly, if would cost more than AI allowed minor expenditure - warn explicitly and ask for permission.
Safety code - would lead to law violation or safety violation? Decline, even though user clearly wants his boss to be murdered. Allow to overrule only under specific circumstances (like violating traffic rules while transporting a wounded person in a private car to hospital).
Once this technology is unleashed, it wouldn't be restricted even if it COULD be restricted.
Given the nature of today's 'everything goes, everything is fair game' philosophy to linking Internet advertising and data mining, advertisers will exploit the opening. Just as with Google and Facebook algorithms searching every nuance of your on-line activities in order to 'improve your on-line experience', you can be sure someone will start 'sponsoring' these robots. 'We will give you 'x' dollars if your robot will promote...'
If you look hungry, some robot will pipe in with 'They are having a two-for-one special at calories-r-us, just two blocks on the left'.
If you look tired, you will get pressed from some random robot that 'Motel 6 has a one-night special rate...'
And if you look depressed, absolutely guaranteed Big Pharma will pay big bucks if you have your robot detect this and promote 'The newest in anti-depressant medication...' to ease your mind.
Imagine every robot an advertising shrill, just like every Facebook post and google search returns some advertising spiel, based on the tiniest bit of innuendo they can perceive or mine from your activity or expression.
Then again, maybe not a good idea.