PhiloMadrid - Pub Philosophy Meetings in Madrid

Thursday, April 16, 2015

from Lawrence, SUNDAY PhiloMadrid meeting at 6:30pm: Is Artificial Intelligence a threat? + News,

Dear friends,

This Sunday we are discussing: Is Artificial Intelligence a threat?

So by the time you receive this email, it would have gone through and
processed by a myriad of Artificial Intelligence systems. It is the
marvel of the human brain that we can exchange a few million electrical
impulses between us and make things happen in the real world, for
example turning up on Sunday for the meeting. But deep down what we are
afraid of is that our washing machine might one day decide to murder us
in the middle of the night. This won't happen but it does not mean that
the washing machine won't kill us. In my short essay I try to argue why
our washing machine is not a bosom friend and why shouldn't think it is,
useful as it maybe.

In the meantime Ruel has sent us the link to his essay followed by news
from Miguel about Maths tertulias and from David about visits to the
British Cemetery in Madrid.

----Ruel's essay
Hello Lawrence,
Here is the link to the essay I wrote:
See you on Sunday.

---------------From Miguel
Estimado tertuliano,
Por si fuera de interés te anunciamos la conferencia adjunta.
Saludos cordiales,
Tertulia de Matemáticas

------ British Cemetery in Madrid visits
Redacto el presente mensaje tanto en español como en inglés con el
objeto de comunicarles el programa de visitas guiadas al Cementerio
Británico, los sábados por la mañana a las 11.00 horas - el punto de
encuentro es la entrada del Cementerio

sábado, día 25 de abril, con las explicaciones en español
sábado, día 9 de mayo, con las explicaciones en inglés
sábado, día 30 de mayo, con las explicaciones en español

Si prefiere hacer la visita en una fecha no programada y siempre que
formen un grupo de un mínimo de 8 personas, avíseme a <please send me a
private message and I'll pass it on to David>

se pone la dirección.

I am writing this note in both Spanish and in English to provide the
programme of Saturday morning guided visits to the British Cemetery, all
of which take place at 11.00 a.m - we meet at the Cemetery entrance

Saturday, 25th April : the visit will be in Spanish
Saturday, 9th May : the visit will be in English
Saturday, 30th May : the visit will be in Spanish

If you would like a visit on a different date and you can form a group
of 8 persons or more, let me know at <>
for details of location.
David Butler

-------------Essay Lawrence

The problem with artificial intelligence is that it is tainted with an original sin. It is a child begotten from human pride in believing that we can create an intelligent system that is perfect and certainly more intelligent than us. But although we cannot create a system with perfect intelligence we can create a system that is more efficient than us; and certainly more loyal than human beings.

There are a number of limitations of artificial intelligence that confirm that AI cannot be a perfect intelligent system. The first of these limitations is precisely the limitations that we human beings have: our knowledge about the world is based on probability because we are limited to the induction method of thinking. It is not that we cannot say anything with certainty about the physical world but that we cannot say a priori when we have reached an infallible level of certainty. In other words, we cannot say whether a hypothesis is certain about the world in advance before trying to prove it or refute it, and how many empirical examples does to take to refute or confirm a hypothesis before it becomes certain knowledge? Thus, if AI systems rely on inductive reasoning to “learn” then they have the same learning weaknesses as us.

As a methodology AI has the same empirical limitations as us. Of course we have to distinguish here between AI methodology and AI application in say machines (Machine intelligence). If 2+3=5 as an example of AI methodology then 2€ in Bank account A and 3€ in Bank account = 5€ in my bank is an example (although not very imaginative) of applied AI. Thus, although two plus three will always be five there is no reason to suppose a priori that the amount of money we have in our bank account is always accurate. This is the empirical curse of induction, we are condemned to always having to verify the future.

The other important limitation of AI is that without exception until now, AI is always applied to solve human problems especially our interaction with the world we live in. And even then, the nature of the problem is in doing things better, quicker, accurately, longer, repetitively than us and in awkward situations that are uncomfortable for us. For example, accurately filtering out digital and white noise when we are having a conversation with someone else on our mobile phone.

But before we continue, although very relevant to the previous paragraph, it is important to distinguish the difference between AI as a methodology and AI as an application. As I have already indicated we encounter AI in machines, thus our experience of AI and our myths are based on how machines interact with us. So referring to our question the “threat” part is of course a threat to us. We are afraid that some machine might become independent of human control and start doing things to harm us. For example, we don’t worry too much, if at all, that one day a machine that has an AI operating system, for example a machine in coal fired power station, will decide to stop power production because of climate warming. We don’t make AI to be a moral agent; at least not yet. But we are afraid that “somehow” a machine in a power station might “decide” to send out a power surge through the grid to fry the electronic chips in our pc’s and electro domestic hardware.

So one of our first tasks is to get rid of the Hollywood impression and myths on how AI can be a threat to us. The threat is more likely to be the problem of induction if a system is based on data gathering for its operation or a fixed database. And in particular, can an AI system handle or deal with such events as a “black swan” (see Nassim Taleb) or a “dragon-king” (see Didier Sornette)?

Basically, and I mean very basic, a black-swan is an event that happens so rare that we do not even consider it in our calculations; this is akin to the problem of induction but with some new important twists. And dragon-kings (are an even more complex idea) are events (mainly but not exclusively negative events) we can predict a priori by the very nature of a given system. This sounds very similar to determinism but here the problem is that the event is due to the very nature of the system and it is predictable; for example by using pneumatic tyres we can predict a priori that tyres will have punctures. Sornette has successfully predicted certain events, for example in the stock markets using his methodology to predict not only what will happen but also when it will happen.

Indeed, the first limitations of AI systems are the limitations we impose on the system, never mind philosophical limitations. What we choose to include in the AI system and what system we use to solve a given problem will itself determine the kind of failures of the system. Added to this is the very likelihood of human errors, carelessness and unfortunate random event.

For example, the auto pilot of the crashed plane in France was not build to recognise a malevolent procedure from a benevolent procedure by a pilot. Indeed the plane is designed to recognise and maybe prevent illegal moves by the pilot (i.e. protect the plane from the pilot concept of plane design) but clearly if we accept the official version (and this is key) of events the auto pilot was not built to recognise intentional legal manoeuvres by the pilot with illegal consequences. And yet all the relevant information was available to an AI auto pilot to distinguish malevolent from legal actions: why turn off the auto pilot when there is no emergency, the weather is good, and this part of the journey is usually flown by auto pilot, and given the speed, height and location of the plane the new instructions won’t lead to an airfield, why all these changes when there is only one crew member at the controls (there are ways to detect this) etc etc?

The question is not whether the AI system, in this case, the Auto Pilot can identify a legitimate move by the pilot but whether the AI system can identify a morally sound legitimate move by the pilot. The captain would immediately have recognised that the new change in the course of the plane was an illegal and immoral manoeuvre and would have acted to prevent the outcome. In the official version of events the pilot entered legitimate new instructions but was not designed to question the morality of those new instructions; it would however, have questioned the legitimacy of say increasing the speed of the plane beyond the capacity of the engines.

And this goes back to the original sin I started with; we believe we can build a perfect system when in reality the system is built in our imperfect image. And one of those imperfections is that we tend to give more value to behaviour rather than intentions. Indeed the legal profession do recognises this weakness in human beings and therefore actively emphasise the importance of “intention” as a necessary condition in the type of outcome in a legal case.  AI systems mimic behavioural patterns of human beings and not intentional acts of human beings. The auto pilot is very good at maintaining height and speed but not very good at determining whether the new instructions are morally or legally legitimate, a mistake by the pilot or a clever way to bypass the safety features of the system and thus intentional instructions to crush the plane. Hence, AI systems are basically systems that “ask” what is being done and how can it be reproduces? Whereas we humans most times also ask, or should ask, “why is it being done?” and “is it good for me?” And as reasonable (in the legal sense) and rational (in the philosophical) sense we can answer these questions independent of whether we are asked or not.

Before our washing machine can try and murder us in the middle of the night there are other issues that AI systems can become a threat or a risk to us. The singular most important feature of AI systems is that these systems require physical inputs and outputs to function; no private language problems here. Hence, the quality of the output (e.g. keep the plane flying straight) depends very much on the quality of its inputs. Thus, if the system does not have a sensor to input the relevant information for example an infra red sensor and a camera to check whether there is a live second pilot in the cockpit the system cannot decide whether the new manoeuvre is suspicious in the first place. Once again AI systems are limited not only by our choice of what we want the system to do but by our foresight of what our system ought to do.

The question Turing asked was whether a machine is intelligent? In other words can a machine use intelligence to solve human problem?

From our perspective we mustn’t confuse the term Artificial Intelligence with any notion of human intelligence; what we are talking about is basically computer software that is incorporated in machines to interact with the environment and “...takes actions which maximize its chances of success” (Wikipedia: Intelligence/Artificial Intelligence). Sure, this is like all roads lead to Rome but it does not follow that all these roads are the same.

I have also been using the term AI systems to include the notion of input/output algorithms within a machine designed to achieve something for our purpose. Thus, if Gilbert Ryle put to rest the ghost in the machine argument, i.e. the idea that the mind is not some entity working in parallel with the body, we need to put to rest the idea that there is some “genius in the machine” when we talk about AI. What there is, are a group of algorithms and routines, manifested as electrical patterns, that manage the digital computer in the machine. No doubt these routines and digital interactions can achieve amazing things and do require the hard work of some of the human geniuses we can ever meet, but there is no genius in the machine. The genius is some engineer that is underpaid in some impersonal office trying to make a decent living.

The problem with the Turing test is that the interaction with the machine is just a language behaviour interaction. But even if somehow we can build a machine that looks like our closest friend it will still take more than just behavioural performance similar to my friend to establish an intelligent machine. There are, for example, common experiences and shared emotional experiences that it would be unlikely that a machine can use and incorporate into a dialogue successfully. To begin with language exchanges are also emotional exchanges and emotional exchanges need not always manifest themselves into language acts but that they may manifest into physical acts; eg a hug, a pat on the back, a smile etc. An AI system will probably won’t be able to interact as if it was a live person because it takes more than just a code to establish a moral emotional system. If this was possible personal computers would be more user friendly.

Thus the threat here is that AI systems probably cannot be designed to interact as human being by virtue of the fact that human beings can over ride any regulating code or system based purely on emotional impulse rather than a logic circuit or algorithms. Indeed Godel’s first incompleteness theorem reinforces this argument with the claim by Godel that it is always possible to create formal statements that cannot be proved by the system but nevertheless, a human being can still make sense of the “Godel Statement.” The threat here is that an AI system can mimic a behaviour but it is unlikely to emotionally react and give you a kiss on the cheeks or solve a problem just because we are getting emotional. As I said there is no danger that our washing machine might want to murder us but nor will it surprise us with a kiss. And nor will it stop colours from running just because we get very angry.

The problem for us is that we have this bad habit of anthropomorphising inanimate objects just because it makes things easy for us to relate to these objects. AI based machines won’t be having intentional actions independent of us unless we program them to function in a certain way when they detect a certain empirical input. And we can understand things much better if we describe a machine that has broken down or a machine not fit for purpose to be evil or bad rather just a machine malfunctioning. A malfunctioning machine takes away the emotional outrage we are so addicted to.

The other big problem is that we tend to play loose and dirty with language. Artificial Intelligence is just a term some scientists gave a certain engineering activity and problem a few decades ago; AI is just a name and there is nothing else to be implied from these two words put together. This is a quirk of English that we can build these elaborate noun groups. There is nothing intelligent about machines we have designed and built to try and solve our problems and they are artificial simply because these machines don’t grow on trees.

To conclude, AI systems are not a threat, what is a threat is the human component part of the system. Speaking for myself I am not afraid that an AI machine might want to kill me, but I am afraid that I might be killed by such a machine.

Best Lawrence
(typos corrected  19/04/2015)

tel: 606081813 <>
PhiloMadrid Meeting
Meet 6:30pm
Centro Segoviano
Alburquerque, 14
28010 Madrid
Metro: Bilbao
Open Tertulia in English every
From: January 15 at Triskel in c/San Vicente Ferrer 3.
Time: from 19:30 to 21h

from Lawrence, SUNDAY PhiloMadrid meeting at 6:30pm: Is Artificial
Intelligence a threat? + News

No comments:


© of the respective authors,
™ of the respective owners,
® of the respective registered owners.

Philosophy, Social Issues, Classical Philosophy, Citizen Philosophy, Applied Philosophy, Non-Political Meeting, Non-Religious Meeting,