Re: Immoral optimisation - killing AIs

From: Olie L (
Date: Tue Nov 15 2005 - 22:16:33 MST

I've searched through the archives, and noticed that there is very little
said here about the meta-ethics of the value of AIs' existences. This is
kinda important stuff, and although I respect that 1) it's a hard area 2)
It's not as much fun as probability theory, it's kinda important as far as
friendly-AI is concerned. Without a decent understanding of why an entity
(humans) shouldn't be destroyed, all the practicalities of how to ensure
their continued existence is kinda castles-in-the-air.

*Warning! Warning! Typical introductory paragraph for a nutbag to prattle on
about "Universal Morality"*

Yeah, we don't want a Sysop to convert us into processing units, or to
decide that the best way to solve our problems is to dope us up to the
eyeballs with sedatives, but what's the reasoning for "vim" not to do so
from ver perspective? Note that asking "why shouldn't?" is an entirely
different question from "why wouldn't," which some of the people here are
doing admirable work on (cue applause).

The "would" can be addressed by goal-system examination etc. Unfortunately,
the "should" issue can /only/ addressed with Morality, which attracts people
to spout exceptional amounts of drivel. I'm going to venture out, and try
to make a few meaningful statements about meta-ethics.

First, a primer on what I mean by "meta-ethics". Ethics studies tend to
fall into three categories: Meta, normative and practical.
*Meta-ethics* looks at what moral statement are: what does "good" mean; can
one derive "ought" from "is"; problems of subjectivity and whether or not
moral propositions can even have any truth-value.
*Normative ethics* looks at systems for going about achieving good - these
tend to be focussed around variations of utilitarianism, justice or
rule-based systems. Typical concerns: Is it ok to do some bad stuff to
achieve lots of good stuff?
*Practical ethics* is... you guessed it... applying normative ethics in
practice. "We've got $10B to spend on healthcare. What do we do with it?"
or "Under what conditions are 5th trimester abortions permissable?"

One would expect that non-singularity AIs - I'm thinking advanced weak AIs -
would still be useful/good at dealing with practical ethical problems,
however, it will take some pretty savvy intelligences to get much ground on
the meta-ethics end. Nb: don't get cocky with meta-ethics. Many of the
best philosophers have taken a plunge and sunk. Anyhoo:

*Puts on Moral Philosopher hat*

Why might it be wrong to turn off an AI?

Most moral philosophers have a pretty hard time dealing with the meta-ethics
of killing people (contastively, the meta ethics on making living people
suffer is pretty straightforward - suffering is bad, bad is Bad, avoid
causing suffering).

Apart from issues of making friends and family suffer, the meta-ethical
grounding for proscribing killing usually comes down to (1) sanctity, which
doesn't hold for non-religious types (2) Divine command - same problem (3)
Rights - based approach (4) Actions-based approach. There are also a few
others, such as social order considerations, but... I can't be stuffed
wading through that. I'll focus on 3 and 4.

The main idea behind these is that people have plans and intentions, and
that disrupting these plans is "bad". The rights-approach says that there
are certain qualities that give an entity a "Right" that shouldn't be
violated - the qualities often stem back to the plans and intentions, so an
examination of these is relevant...

A key question for the issue of killing / turning off an AI is whether or
not the AI has any plans, any desire to continue operating.

There a few senses of the phrase "nothing to do" - if an intelligence is
bored but wants to do stuff, that's a desire that off-ness (death)
interferes with. If, on the other hand, an intelligence has no desire to do
anything, no desire to think or feel, feels quite content to desist being
intelligent, then death does not interfere with any desires. An
intelligence that has no desire to continue existing won't mind being

/I'm not doing a fantastic job of justifying any of these positions, largely
because I disagree with most of these meta-ethical approaches. For the
others, I lack a complete technical understanding. I'll therefore resort to
the customary ethical technique of providing analogies, reductio ad
absurdum, and relying on intuitions to invent rules (sigh)./

Imagine someone wishing to commit suicide. Is this an ethically acceptable
course of action? I think so, particularly if they're generally having a
rough time (terminal illness etc). Just imagine they've put their affairs
in order, said goodbye to their family, are about to put a plastic bag full
of happy gas over their head... when somebody else shoots them in the back
of the head, killing them instantly. Is the assassin here doing something
ethically unacceptable? Are the intentions/ actions bad? Is the result
bad? If the assasin is aware of the suicide-attempter's plans, does that
make a difference?

I would suggest that although the killer's intentions could be immoral, the
result ain't bad. Whether the means of death is self-inflicted,
other-inflicted or nature-inflicted, the suicidor's wish is granted.
Killing a person with no desire to live is not necessarily such a terrible

Drag the analogy accross to AIs: if the AI has no desire to live, is killing
them/ turning them off bad? Not really. An AI with no desire to continue
operating would seem to be, necessarily, an AI with no intentions and no
purpose. One can imagine this happening if the AI has a purpose that is

The interesting counter to this is: would it be extremely immoral to
influence an entity to cease their intentions to live? By whatever means,
causing a person to give up their desire to do, achieve and continue to
live? How about comparing the goodness of creating a person with a high
likelihood of will-to-live-cessation against creating a person more likely
to want to keep living? My intuition says this is dancing around a very
fine line between OK and hideous.


H C wrote:

>"You've specified an AGI which feels desire, and stated it doesn't mimic
>human desires"
>It wants to be Friendly, but it doesn't want to have sex with people or eat

Or get jealous when you ask for a second opinion, or react
agressive-violently to actions it finds threatening?

>>From: Phillip Huggan <>
>>Subject: Re: Immorally optimized? - alternate observation points
>>Date: Fri, 9 Sep 2005 11:15:04 -0700 (PDT)
>>H C <> wrote:
>> >Imagine (attempted) Friendly AGI named X, who resides in some computer
>> >simulation. X observes things, gives meaning, feels desire,
>> >and is capable of creating tests for vis hypotheses. In other words, AGI
>> >is actually a *real* intelligent AGI, intelligent in the human sense
>> >without athropomorphizing human thought procedures and desires).
>> >Now imagine that AGI X has the capability to run "alternate observation
>> >points" in which ve creates another "instance" of the [observation
>>program -
>> >aka intelligence program] and runs this intelligence program on one
>> >particular problem... and this instance exists independently of the X,
>> >except it modifies the same memory base. In other words "I need a
>>program to
>> >fly a helicopter" *clicks in disk recorded where an alternate
>>point already learned/experienced flying helicopter* "Ok thanks."
>> >Now if you optimize this concept, given some problem like "Program this
>> >application", X could create several different AOPs and solve 5
>> >parts of the problem at the same time, shut them down, and start solving
>>main problem of the application with all of the detailed trial and error
>>learning that took place in creating the various parts of the application
>>already done.
>> >The problem is, is it *immoral* to create these "parallel intelligences"
>> >arbitrarily destory them when they've fulfilled their purpose? Also, if
>>decide to respond, try to give explanation for your answers.
>>You've specified an AGI which feels desire, and stated it doesn't mimic
>>human desires. Which is it? If the AGI itself cannot answer this moral
>>dillemma, it is not friendly and we are all in big trouble. I suspect the
>>answer depends upon how important the application is you are telling the
>>AGI to solve. If solving the application requires creating and destroying
>>5 sentient AIs, we are setting a precedent for computronium.
>Good point. You could focus on suboptimal performance while waiting for the
>AGI to Singularitize itself and tell you the answer.
>>Do You Yahoo!?
>>Tired of spam? Yahoo! Mail has the best spam protection around

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:53 MDT