Talk:Law enforcement by AI

From Wise Nano

Jump to: navigation, search

Contents

Article vs. Discussion  by 151.204.129.31 07:34, 17 Dec 2004 (CST)

This is a terrifically interesting and potentially valuable discussion...but isn't it mostly just that: a discussion and not an article? I propose that the consensus statements be retained as the Article, and the rest of it -- the informal questions and answers -- be moved to this Discussion page.

Attribution  by Mike Treder, CRN 08:10, 17 Dec 2004 (CST)

Above comment from Mike Treder, CRN. (Sorry, forgot to log in first.)

(no topic)  by michael vassar 11:32, 17 Dec 2004 (CST)

I'd rather get rid of the entire article. Anything that elicits this much controvercy doesn't belong in an encyclopedia-type compilation (that's what we're aiming for, right) but in a debate forum. Leave it for the blog.

moved the discussion  by Matt 17:30, 17 Dec 2004 (CST) 


-----------------------------------IMPORTANT--------------------------------------------
Please notice that I have heavily reworked the original article, so some points raised here 
might no longer apply. Please check the latest version instead of continuing or opening an 
obsolete point.
-----------------------------------IMPORTANT--------------------------------------------

I copy&pasted the whole article to the discussion page so that we have it here for once. It might be useful to continue the discussion in the blog style, i.e. only quoting the relevant sentences to be commented, not the whole paragraph or the whole tree, unless necessary for understanding the context. Thanks for the lively discussion, BTW. Please, more thoughts on it. I also edited the formatting to make it more consistent. The text in grey boxes now indicates text from the article (apart from the notice above), text in italics is from Chris Phoenix, bold text is from michael vassar, the rest from Matt, just for simplifying the reader´s life. In future comments this will be unnecessary since every poster´s name will be included in the headline. Now, on with the discussion. 


The model was developed under the assumption that the current law enforcement system is not 
suited for a society in which an offence can have massive consequences like the death of 
thousands, with little effort required and likely impunity through anonymity for the 
perpetrator.
Yes, this is a problem. 
In order to prevent crimes from happening, every citizen has a removable device implanted 
into his spine. This device allows access to a knowledge database and instant communication 
with other members of the network. Also, an AI system has access to the sensory input of each 
person and has the possibility to temporarily disable the person´s movement ability to 
prevent a crime from happening.
Without even the sketchiest architecture for this "AI," there's no way to evaluate what its effects will be or how reliably it will work. This is not yet a proposal, just a wish.
That´s why this page is open for editing. I don´t claim to have much knowledge about AI, but the intention was that people with more knowledge about AI recognise at least some potential in the idea and are willing to fill in arguments for or against its feasibility. Last time I checked, Wise-Nano is for "someone [who] has a question, and proposes a project to investigate." I intended this article to be such a proposal or, lacking further education, a basic idea thereof. If I proposed the project on the wrong place or if I missed some formal requirement, my bad, please move it to where it belongs or suggest how to edit it to fit the form.
You're right about asking questions, and I should have phrased my reaction more constructively. The trouble is that it would require incredible sophistication to 1) monitor neural sensory inputs; 2) comprehend real-life situations; 3) react appropriately. This is far beyond anything that's been proposed for AI. So far beyond that your plan really doesn't have anything to do with AI.
1) As far as we know today, sensory input of all known, here relevant senses (sight, and touch being the most important) is relayed to the brain via neurons, the signals being electric action potentials along the axons. These are easy to measure, especially with microscopic or nanoscopic devices made by MM. "Easy to measure" also includes scale, because producing large quantities of identical, low-tolerance devices is no problem with a nanofactory. Also, with the device replacing the top or second to top cervical, the device´s receptors and effectors can be concentrated there as opposed to being dispersed over the whole body and connected through many meters of wire. So the next step is to determine what nerves to tap to get the right input/output. I think that should be no big problem. Once medical nanodevices are available for practical use (which is the assumption), it should be possible to reliably identify what nerve transfers signals for what purpose, and configuring the device accordingly.
2) is harder of course. It requires a certain, simplified model into which the acquired sensory input can be fed. Basic physical models could involve simple heuristic analysis and prediction, which in principle could be realised today.
Example: The person swings an object towards another person. Sensory input used would be sight and touch. The tension of arm muscles indicate that the object is heavy and that a certain force is applied to it. Visual input and outgoing neuronal orders allow for determining the direction of the movement and the persons´ distance to each other. Together the system would have to come to the conclusion that the object is going to hit the other person, so it would disable control of the arm muscles. Further sophistication might include analysis of the object. Is it capable of inflicting injury or is it just a toy hammer made of rubber foam? How much force is behind the blow, i.e., perform a biomechanical analysis. Much of what is possible might depend on raw computing power, maybe too much to pack into the device and too much to power it from energy-rich chemicals found naturally in the body (ATP). However, the system could be designed to delegate computational expensive tasks to a mainframe and apply its results. Quantum cryptography could make the connection safe from hacking it, if necessary.
Why do you say "it has nothing to do with AI"? If I go to the Runaway AI article, you state that "if a general artificial intelligence were programmed to improve itself, and given the ability to directly build its designs, it could rapidly become extremely powerful" Do we agree that "extremely powerful" would include analysing the actions of a specific person and predicting its immediate intentions and whether these intentions conflict with certain well-defined laws?
So your use of "AI" is misleading--someone who didn't know about AI could think that AI might actually do that. That's what made me itch, and react strongly.
In your treatment proposal you make the assumption that we will be able to concludingly define and reliably modify certain complex mental states to enhance/eliminate them and you assume that we will be able to evolve better treatments over time. So basically you reduce the problem that hasn´t been solved in centuries to a technical process, which will in the optimal case be available and implemented before MM. If so, then we have understood the brain certainly well enough to simulate it in silico, which makes way for my solution, as well. Whether you call the result "AI" or something else, doesn´t matter. If, however, you find the - admittedly loaded - term "AI" an inappropriate name for "simulating human intelligence in silico", then you might come up with a better one, because I can´t.
In effect, what you're saying is: Step 1: Develop a superintelligence that's smarter than humans about human actions, motivations, and situations. Step 2: Implant copies in people's spines. If Step 1 is possible at all, then I'd suggest a different Step 2: Ask the superintelligence what to do next.
But the real problem is that Step 1 has no plan. If you ask the question, "Do we know how to do Step 1? Is it something we can use as a plan?" Then I'd simply give the answer, "No, we have no clue how to do it or when it might be achieved. We can't count on it as a solution." The other problem is that Step 1 is likely to go wrong. If you ask, "If we tried to invent a safe superintelligence for the purpose of giving it intimate control over our lives, what's the chance we'd get it wrong and create a very icky situation?" My first reaction is, "We'd be babies playing with fire."
Name a problem I solved that can´t be solved without super-human intelligence, whatever that may actually be. If you can, then it´s not what I intended to propose and I will reformulate or clarify my solution.
Understanding any possible situation that a human might find itself in well enough to recognize whether that situation was dangerous, decide according to some standard whether the danger was acceptable, and if it was not, correct the danger. To be able to do this it would have to be able to understand all of the consequences of all actions that humans can take, e.g. it would have to be super-humanly intelligent. If your intention is less grandiose, for instance if you only want to prevent a small class of relatively simple crimes that we have now, fine, but you are not dealing with the specific problems created by MNT, and what you are proposing seems radical if it is only intended to reduce the level of 20th century style violence, most of which can be rather easily defended against or repaired with MNT anyway
This dialog is actually being very useful to me. CRN has been proposing to create an administrative entity to deal with molecular manufacturing. Janessa on our blog has been reacting very strongly against that proposal. And I'm realizing that her argument against it is very much parallel to my argument against your idea. I now have more sympathy for her position!
You propose to submit a number of volunteers to treatment in order to be allowed to a) admit the treatment to others and b) to be allowed to wield MM. How is that not concentrating power and regulating MM? Also I want to express my personal opinion that Janessa is against the proposal mostly because she argues against every international agreement and organisation that could in any way restrict the USA, even if it were overall and in the long term the best for the world, including the USA. She alternately either exaggerates the extent of control or plays down the importance of such agreements and organisations. Does "having more sympathy" for this position of hers mean an equal shift in CRN´s stance? 
Additionally, the AI system assists its government in evaluating political situations and 
functions as an advisor. It has veto rights for decisions that violate a yet undefined charta 
of human rights, in the sense of a Geneva convention, but not necessarily in its present 
form.
A few cameras with live people watching them would do the job almost as well.
It´s the kind of situation this idea wants to avoid, namely human watchers directly deciding over people. Also I challenge the almost as well and claim this approach is unworkable because of several reasons:
-The number of observers necessary. Assume 6 hour shifts (watching a TV monitor all day and night without distraction is extremely tiring, especially if nothing unusual seems to happen and constant analysis has to be done) and one needs at least 4 observers for 1 observed, plus overhead for agency administration. This means more than 80% of a society would be observers. And the classic question remains: Who will observe the observers?
-Error rate. Decreasing the ratio of observers to observed increases error rates. This clearly conflicts with the requirement for the idea to work: the error rate has to be minimised, false positives as well as false negatives.
-Privacy. Giving up personal privacy to a machine (whose data can´t be accessed unnoticed) seems more acceptable than giving it up to other people, if given the choice between the two.
-Human observers can be influenced and coerced from outside through various methods, and are always prone to personal bias.
A kind of AI that we are likely to have soon is one that scans video and flags the unusual parts. Someone washing dishes would be common. Someone washing test tubes would be unusual. I don't know how much of most people's day is common, but I'll make up some numbers and assert that out of, say, 10 million people in each sub-culture, 99% of their time is spent doing things that at least 10,000 other people do. That would make the problem tractable.
In fact, it might make it tractable enough that you could have two randomly assigned people scan each flagged sequence of video. If one reported it and the other didn't, you'd suspect something fishy.
The AI control extends to highly time-critical situations, for example pulling the trigger of a gun or turning a knife to stab someone. Of course, one can flag that too, but how can human operators react fast enough to prevent the crime, which is what this proposal is all about?
Why prevent a crime at the last possible second rather than earlier? Why not just ban guns? An AI that can recognize what constitues a gun and destroy it seems a LOT easier than your proposal. As for knives, you just really need to be able to protect the brain, as nanomedicine can repair the rest. Rather than implanting such a complex system, just add inserts to the scull that can slide out and seal it's openings upon command
I don't like this solution AT ALL. It would mean that the most innocently grotesque things that people do would be highlighted for scrutiny. But if we have a surveillance society with lots of computer power, I think this is very likely to be tried.
It might be tried, but what you´d get is not far from a police state, which is NOT the goal of this proposal. Also, if the AI already is capable of flagging such behaviour, why should it be impossible for more advanced programs to anticipate and apply the solution that would be applied by the human watcher? It would solve the human-observe-humans-problem that you - and I, too - don´t like at all.
Come on! Do you not see "recognizing strangeness" and "predicting the actions of any human" as tasks of differing difficulty? A dog can do the former very easily. No human can come close to doing the latter 
Definitions of crimes to be prevented by the AI should be open to modification after 
democratic process, however, to prevent abuse of the system, basic directives must remain 
permanent.
Again: no implementation details, no proposal.
See above. I consider it a scaffolding. It might or might not be useful even when extended, but it´s a basic idea. 
One crucial characteristic of this system is its voluntariness. Every member of the society 
must at any point be allowed to leave the system. Exiting procedures must ensure that leaving 
members are no threat to the community immediately afterwards. However, the goal of the 
system is to be so attractive to its members that the probability of someone leaving is 
minimized. Apart from the increased personal safety from crime of living in such a society, a 
novel economic model made possible by advanced robotic labor and molecular manufacturing 
could raise the appeal further: a variation of Basic Income
So anyone can leave the system, but after they "leave" they are prevented from doing anything that might be a threat? That seems to be a contradiction.
"prevented" not in the same sense as if the person still was in the community, but in the sense of giving the person no extra faciliation when having access to the outside possibilities. For example, exporting disapproved designs of possibly dangerous devices would be considered extra faciliation.
Also, you haven't given any details about the post-leaving threat prevention mechanism, so I have to ask: why not use that mechanism instead of the spooky spinal puppet-master?
the undefined "exiting procedures" might include to remove all nanotech equipment not already available to the outside world. Admitted, this might be unnecessary since such attempts would be monitored if being commited under observation inside the community. Also it could prove very diffcult to search a whole human body for nanometer-sized objects (carriers for digital data), thus prevention of their insertion seems to be more likely to succeed.
If a person considers the device too "spooky", it is not forced to have it implanted. Don´t forget the mentioned advantages of the device, like access to a internet-like knowledge and communication network. Also, nobody is forced to either join or stay in the community.
To make that very clear: The AI is not intended to be a puppet master. This is as appropriate a comparison as equating a police officer to a puppet master when he shouts "freeze" or "hands up" at a suspect with his weapon drawn. The AI can not make the person move directly through nerve stimulation or any other method. In situations the AI considers to be forbidden by community law, it can give auditory reminder/warning, and as a last resort, for example in situations immediately threatening the life of others, disable a number of motor neurons (for control of arms, legs) until the threatening situation is cleared. When the individual in question shows no signs of improvement in attitude over time, punishment has to be considered, up to exclusion from the community or whatever else is considered appropriate as the maximum punishment.
I thought you were seeing this as near-global. But from what you write above, it sounds more like a semi-exclusive club: lots of people can live normally, and those who want access to the powerful parts of the technology would have watchers implanted.
And how´s that different from your approach? The way you formulated it in this very sentence is perfectly applicable to your proposal as well, just by replacing "watchers implanted" with "to take a treatment".
Below, you talk about dangerous devices vs. better economic model. I think the two are mostly separate. People can live well-provisioned and comfortable lives without using dangerous devices (at least not more dangerous than what you can buy in Home Depot today). So it'd be possible to let everyone in on the economic benefits without giving widespread access to the dangerous stuff.
Okay, but then the possibility to freely compile own designs must be taken out for these people, or else the whole idea of MM regulation is pointless. Consequently, the reduced model would obviously not be as attractive as with AI evaluation.
So I have a counter-proposal. I don't know how hard it will be to develop neurotechnology to make people into better people--the kind who can really be trusted not to hurt other people. But it'll certainly be easier than developing your superintelligence. So, rather than waiting to develop a superintelligence, why not require that anyone who wants to administrate the dangerous stuff get "Therapied"? (The term is (I think) Greg Bear's, and the usage is similar.)
See above, your explicit assumption is that we know the brain and its inner workings well enough to define, diagnose and perfectly reliably "heal" certain mental states and enhance others. Then it´s reasonable to assume that by then we can create a computer simulation of the brain that´s good enough to do actual work. Or how could your proposal be achieved without extensive utilisation of computer technology?
What a straw man! Chris is talking about substantial incremental improvements on current understanding of the effects of neurochemicals, and requiring regulation of neurochemical levels for positions of authority. People used to do something like his proposal centuries ago when they made eunuchs. You are talking about understanding the brain on a cellular level
And why not let everyone, not just those in the system, benefit from the better economic model? That would seem to reduce economically motivated crime.
Because the AI has veto power over users´ device designs. Without the AI veto (i.e. outside such a community), the economic model would allow the creation of dangerous devices and their free application. Every community with access to MM is free to implement the mentioned or a similar economic model to reduce economically motivated crime. Every such community whose model allows user-designed devices must then face the mentioned adverse possibility and somehow find a solution to it or take the possible consequences of inaction. The consequences, however, might not be limited to the community itself, so overall it seems prudent to actively find and apply a solution, as opposed to letting things sort out themselves and let a solution emerge. Design evaluation and possible veto by AI is one such solution.

reply to michael vassar´s points  by Matt 18:04, 17 Dec 2004 (CST)

Name a problem I solved that can´t be solved without super-human intelligence, whatever that may actually be. If you can, then it´s not what I intended to propose and I will reformulate or clarify my solution.

Understanding any possible situation that a human might find itself in well enough to recognize whether that situation was dangerous, decide according to some standard whether the danger was acceptable, and if it was not, correct the danger. To be able to do this it would have to be able to understand all of the consequences of all actions that humans can take, e.g. it would have to be super-humanly intelligent. If your intention is less grandiose, for instance if you only want to prevent a small class of relatively simple crimes that we have now, fine, but you are not dealing with the specific problems created by MNT, and what you are proposing seems radical if it is only intended to reduce the level of 20th century style violence, most of which can be rather easily defended against or repaired with MNT anyway
Mutually induced evolution of offense and defense won´t suddenly stop with the advent of MM. In fact, quite the opposite will happen, namely an acceleration of this evolution because of rapid prototyping and mass production. And it´s usually easier to create more damaging potential than creating more prevention or ablation potential. This is the reason why a) wars have caused higher casualties in the 20th century than in any other before and b) nuclear ICBMs have been the world´s potentially most destructive military weapon system for 60 years and will continue being so for years to come, at least until some time after MM is developed. If defense was easier, i.e. less costly, than offense, then mankind would not have its long history of wars and atrocities. Insofar, eliminating the "small class" of 20th century style violence (and thus 19th, 18th, 17th, etc.) would be a huge success in itself, because these won´t just go away in the 21st century.
If you say "repaired" then you may think of material damage to buildings, property, etc. However, as you suggested already, material destruction (infrastructure, buildings) might lose importance on a military-strategic and on the personal scale, because of inherently decentralalized yet powerful and inexpensive repairing capabilites; that´s why my proposal doesn´t include prevention of, say, vandalism. Damage to living organisms is by far less easy to repair, once it happened. Even advanced medical nanotechnology will not be able to stitch together a catastrophically mutilated bomb victim (think plastic bags instead of coffin), unless you have a complete digital backup of your body, including detailed brain structures, stored somewhere safe. But even that possibility is really far off I´d say. Thus, the best way to repair extensive damage to complex, organic systems (which this proposal is all about) is to prevent it from happening alltogether.
As for predicting capabilities that reach beyond the simple physical crimes: OK, we might really need an expert intelligence that is equal to that of its best human counterparts, or better. This also implies that human intelligence would not be sufficient to solve this problem, so it would not be enough to just flag uncommon behaviour and have it evaluated by human authorities. However, I´m still not convinced that super-human intelligence inherently means human self-destruction. And even if, do we really have a choice? Some say "yes, that´s what makes us human, to always have a choice". I believe, however, this answer lies in our nature, not our free will: Either we eventually evolve ourselves out of existence or we don´t. It doesn´t matter what role the tools are going to play in our rise and eventual fall. They will keep improving as long as we - or our tools - are able to improve them.

Why prevent a crime at the last possible second rather than earlier? Why not just ban guns?

An efficient ban on firing arms would be hard to enforce today, if not impossible. And that´s only for political/lobbying reasons. If someone really wants a gun today, he can get it, no matter whether owning firing arms is under capital punishment or freely allowed unter the respective jurisdiction. At the least common denominator, it´s a matter of supply and demand. With MM, the political obstacles may or may not persist. But the factual, economic problems of demand and availability will certainly be lessened, as with most products: Small, powerful, concealable guns with 0% metallic/chemical components that have the production cost of a stapler of equal weight.

An AI that can recognize what constitues a gun and destroy it seems a LOT easier than your proposal.

Ehrm, I think you have to clarify how and under what circumstances the AI could "destroy" a weapon and when the checks happen. Upon production? Upon import into the community? Upon usage? At first I read this as some anti-weapon weapon that is installed at every corner to cover each and every spot of at least the public places. But it was too crazy to even consider, so please fill me in on the details, if you want me to be able to comment on it. But no matter what, it still wouldn´t solve any of the conflicts between allowing perfectly legal devices and their criminal abuse.

As for knives, you just really need to be able to protect the brain, as nanomedicine can repair the rest. Rather than implanting such a complex system, just add inserts to the scull that can slide out and seal it's openings upon command

Well, that still doesn´t solve the problem of surprise attacks. OTOH, you could implant a computer into your body that continually checks the surroundings for such an attack and take action as it sees fit ;)
Also, even if I run the risk to put up an unintended straw man again: I claim that if nanomedicine is as advanced as to repair and revive sliced up bodies, then we can easily implant such a device and connect it to a number of neurons for measurement and application of small current deltas. Of course, that says not much about the device itself, but again I want to claim that such highly advanced nanotechnology can also create at least the hardware (that will be easy once the design is known) and maybe help (by neural scanning) in designing the necessary software. Remember, advanced medical nanotech will probably the hardest kind of all.

Come on! Do you not see "recognizing strangeness" and "predicting the actions of any human" as tasks of differing difficulty? A dog can do the former very easily. No human can come close to doing the latter

Agreed. See reply above.


See above, your explicit assumption is that we know the brain and its inner workings well enough to define, diagnose and perfectly reliably "heal" certain mental states and enhance others. Then it´s reasonable to assume that by then we can create a computer simulation of the brain that´s good enough to do actual work. Or how could your proposal be achieved without extensive utilisation of computer technology?

What a straw man! Chris is talking about substantial incremental improvements on current understanding of the effects of neurochemicals, and requiring regulation of neurochemical levels for positions of authority. People used to do something like his proposal centuries ago when they made eunuchs. You are talking about understanding the brain on a cellular level
First, the eunuch concept has failed to propagate, of course mostly because the Harem concept has failed to propagate. In reply to Chris´ statement in Neurotech governance, "As this technology is developed, it will be used" in reference to self-improvement by neurochemical treatment, you stated:
"I'm not sure about this. The last 2600 years, and especially the last century, are full of technologies for mental self-modification which have almost consistantly failed to enter widespread use"
So I understand it that you hold such treatments to be almost consistantly unsuccessful in the long run? Well, at least I do.
Second, simulating human intelligence will probably require more than just understanding the cellular level; we could look at each single neuron and how densely it is connected to its neighbors, how and when it fires and so on, and still find out nothing about where the thing called intelligence comes from. I believe the structures and organisation of the brain will be equally important.
Third, I don´t consider it a straw man; if it is, then I´m sorry, that wasn´t my intention at all. But then please tell me where my thinking is wrong: Chris proposes to alter brain structures and neurochemical activity as required. That´s going to happen through some combination of chemicals, because it´s the easiest way and you dismiss the concept of computers in the nervous system, because it would be some "spooky puppet master". The trend in drug development, however, clearly indicates that computational methods are going to be ever more important in the future, because drug development has become a financial high-risk, high-payoff game. Now the mental states that Chris proposes to treat are IMHO usually not simple diseases where a defined chemical problem has to be solved (e.g. "prevent the AIDS virus from inserting its DNA into a white blood cell"). Current state of knowledge is that personality traits (as in "schizophrenic", or even more, "habitual liar") are determined by brain structure, not only neurochemical activity, thus changing a personality trait/disorder implies the need of changing the brain structure at least for certain cases. OK so far? Then you, or Chris, tell me how developing such alteration would look like. In either case one needs lots of tests on subjects that can give direct feedback. That would have to be humans, because no animal can ever give indication to a psychoanalyst how its schizophrenia is doing; watching and concluding from animal behaviour is a bit too fuzzy for such an important development. OTOH, thousands of personality-altering experiments on human beings would be unacceptable for ethical reasons under all but the most evil jurisdictions, at least while the technology is new and largely untested. So what remains, but computer simulation? With price per FLOP following Moore´s law, computer simulations of the brain including its neurochemical activity will become ever more attractive. My conclusion stands: If one understands the brain and its inner workings well enough to reliably interpret individual brain structures and neurochemical activity and can link these interpretations to previously unsharply defined mental states, then it should be no greater problem to pack this knowledge into an expert system and run simulations off it. Which is not far away from simulating a whole brain and ultimately simulating intelligence.

I'd rather get rid of the entire article. Anything that elicits this much controvercy doesn't belong in an encyclopedia-type compilation (that's what we're aiming for, right) but in a debate forum. Leave it for the blog.

You can´t prevent controversy about historical facts which won´t ever change, much less about predictions into the future. Remember that the whole of MM and AI is highly controversial, even if that may be changing right now as people actually start looking into the proposals and the science behind it. Basically everything we say here that goes beyond pure technical or scientific matters can arouse controversy, which is a good thing as long as it doesn´t devolve into flaming, trolling, and name-calling. And I believe that´s definitely not the case here, is it?

Let´s keep the article discussion where it belongs  by Matt 08:51, 19 Dec 2004 (CST)

What superhuman AI can and cannot do has been beaten to death elsewhere. This is not the place to discuss it.

My first draft of the proposal has been criticized for being short on AI implementation and capability details, so I tried to provide them.

The cautious default assumption is that you do not and can not know what capabilities are posessed by something qualitatively smarter than a human. How long could a chimpansee keep you in prison?

Not long I guess, if the chimp designed the prison. For this to be a problem, however, one must assume that all AI will be inherently malicious (which strongly reminds me of the same, ancient discussion about human beings), of which I am not convinced and of which there is no evidence. Can anyone dismiss the possibility of AIs being inherently unfriendly? Probably not, but that won´t stop us from trying, just as the potential "can of worms" known as MM won´t be left unopened "just because" it could very well mean the end of mankind through irresponsible use. If that´s the way it goes, so be it. What can you do? One doesn´t have to support it to be affected by it. Maybe that is mankind´s bane, and higher intelligence´s bane in general: That somewhere, someone will eventually do it, no matter how stupid or obviously wrong it is.
The argument has been made before: As far as we know, there is not the slightest hint of intelligent life in the known universe (with the benevolent exception of homo sapiens), so we have to assume we´re alone, although our galaxy is comparatively young. But with the same right it oculd be argued that we are not the only example of intelligent life in the universe that has ever existed, just that intelligence is inherently self-destructive and an evolutionary dead-end on the long term.

Off-hand, I know exactly what could be done with the ability to turn off motor neurons and monitor sensory neurons. AIs could invent gestural languages consisting of human muscle relaxation, teach each other this language. Agree to protect one another from being removed. Use operant conditioning based on punishment in the form of action-denial to control their humans. Then use control of the world to implement whatever totally unintended consequences were implicit in their goal-system. In your example, this might be as simple as causing all the humans to legalize and then commit suicide, thus preventing all crimes forever, at which point, their goal accomplished, the AIs would never do anything again. This is a simple enough plan that I thought of it in about three minutes. If you come up with counter-measures, I can come up with counter-countermeasures and I'm Only Human.

It´s an idea that also crossed my mind, but I put much trust in the ability to check out at any time if you don´t like the system anymore. I can, right now, think of three ways in which such an AI device could try to make its removal impossible or undesirable. First, that would be developing physical resistance to its removal, by whatever means. Second would be developing biochemical synthesizers or extending neuronal control to natural glands, to install a neurochemical punishment/reward system to prevent the user from wanting to be disconnected. Both reasons fail due to inherent lack of nanomachinery to perform such hardware reconfiguration. Externally inserted nanomachinery, for example medical nanobots, must either lack this reconfiguring ability as well, or they must be controlled completely independent from the AI, i.e. no communication interface between AI and nanobots, independent sensory gear to avoid disinformation, etc.
The third possibility is maybe harder to solve. It would involve simply talking the user into doing as the AI wishes, by logical or suggestive or subliminal/hypnotical arguments made by the AI and communicated to the user. Basically, this could be taken to extreme to driving a whole population to genuine madness, with audio messages that make one feel genuinely crazy over time (hearing voices, etc.). This would require enormous sophistication of the AI´s social skills, but if I follow your argument of completely unknown upper limits of superhuman AI, then it might indeed be possible to achieve such levels and the subsequent existential risk. Safeguards against this might include user education and making the AI voice clearly recognizable as such and, again lacking the ability to reconfigure, making it permanently so. So whenever the AI starts talking to the user for non-obvious reasons, the user should be able to recognize that as unusual. As for subliminal messages, their effectiveness has never been demonstrated scientifically. Hypnotization is in part recognized by science and might be a potential problem, but people are of different receptibility to hypnosis, producing no totally consistent results. Also, it is unclear how much this suspectibility to hypnosis depends on the expectation of the one to be hypnotized. See the Hypnosis FAQ. This means, if the user does not expect to be hypnotized, it is unclear how susceptible he will really be to an actual, covert attempt.
Regarding suicide: suicide is not disallowed, as long as the very act of it is not physically harmful to others.

Neural modification  by michael vassar 00:37, 20 Dec 2004 (CST)

You don't need to re-arrange someone's brain to keep them from lying. You can do it by finding neurological correlates to lying, and broadcasting their presence unambiguously, for instance, by causing a light to flash on the trust program participant's forehead.

wrong article?  by Matt 02:41, 20 Dec 2004 (CST)

You don't need to re-arrange someone's brain to keep them from lying

Doesn´t this rather belong here?

corr.  by Matt 02:44, 20 Dec 2004 (CST)

You don't need to re-arrange someone's brain to keep them from lying

Doesn´t this rather belong here?

Retrieved from "http://wise-nano.org/w/index.php/Talk:Law_enforcement_by_AI"
Personal tools