[aiethics] Let's begin...

[Noel Bush]

Alicebot AI Ethics Committee - http://www.alicebot.org


> What about an official "seal of approval" issued by a committee
> at the ALICE
> AI foundation functioning as the executive wing of the ethics
> committee? We
> cannot ensure that AIML will not be used for unethical purposes =
(whatever
> "unethical" may be), but at least we can make sure that a user
> who wants to
> have a good clean conversation with a bot based on AIML just needs to =
look
> for a bot which is A.L.I.C.E.-certified...
> I'm thinking of Rich's original intention to keep Alice suitable for
> children for example...
> We might even extend this to ratings similar to the ones that are
> applied to
> movies (G, R, X, PG...)

Good luck! There's no way to certify this kind of thing with any =
Alice-like
technology (or any other). Just consider these innocuous kinds of
constructions:

<category>
<pattern>DO YOU LIKE TO *</pattern>
<template>
Sure, I <star/> ten times a day! How often do you <star/>?
</template>
</category>

Of course, the first time somebody types something "nasty" there, and =
your
bot responds with some embarassing phrase, and you check your =
TARGETS.aiml
and "plug the hole", that won't happen again. But this is completely =
trial
and error.

I would say that "certifying" content is not a task pertaining to =
ethics;
it's a task pertaining to censorship. The world is still grappling with =
how
to encode censorship systems. P3P has made very nice strides in this =
regard,
allowing you to choose the ratings system of the group of your choice. =
If
you are a conservative Christian in the US and you want to "enforce the
values" of the Christian Coalition, then if the Christian Coalition
publishes a ratings spec you can use it to block sites, words, etc. from
your browser. Governments like those of Germany, France, United Arab
Emirates and others who have national-level legal restrictions on speech
might eventually mandate the use of a particular ratings system.

But the technological problem of handling this in a conversational =
scenario
is mind-boggling. And anyone who's thought about it for long in a =
simpler
scenario like that of http requests also knows that this is a losing =
battle.
When there is a natural language parser that can understand the intent =
of "I
want to give you some candy" *in context* as malicious, salacious, =
harmless,
or whatever else, then I think we'll all forget about Alice in a split
second. But Alice was created with the belief that for many practical
purposes, this kind of "understanding" isn't really necessary in the =
first
place, and with the observation that attempts at constructing a machine =
that
can do that kind of understanding have been disappointing, to say the =
least.

I think that the ethical questions of "what people should and shouldn't =
do
with bots" are no different from the ethical questions of "what people
should and shouldn't do with spatulas".

What is interesting to discuss is "what *bots* should and shouldn't do",
imagining that bots themselves are autonomous and accountable for their =
own
actions. Those are, of course, the questions that Asimov's laws try to
address. But I think they don't go far enough. Let's try to imagine how =
a
bot -- not its creator -- should be held accountable for its own =
actions.
This won't be pure fantasy in time. When machines beget other machines, =
when
programs beget other programs, it will be absurd to hold the creator
responsible for the outcome. Roger Clarke posts some interesting =
extensions
of Asimov's laws here:
http://www.anu.edu.au/people/Roger.Clarke/SOS/Asimov.html. One of them =
is
the "procreation law", which says that "A robot may not take any part in =
the
design or manufacture of a robot unless the new robot's actions are =
subject
to the Laws of Robotics". But isn't it folly to imagine such =
capabilities --
to constrain any system well enough that it can't do something that =
violates
any law? And aren't the religious systems of the world telling, in that =
very
few of them actually hold a creator *responsible* for the actions of its
creations? Responsibility for behavior lies with the individual, =
questions
of explanation for the behavior aside.

I think it is most reasonable to assume that "artificial intelligences",
whatever they are and whenever they exist, will be far more difficult to
regulate than, say, Microsoft Word. And Microsoft Word is difficult =
enough
to regulate. Of course I have no hope if I try to punish Microsoft Word =
for
its bad behavior, or tell it what it should do. And Microsoft Word also
doesn't have children who go off and do things I don't know about, and =
have
more children, etc.

But let's imagine a kind of Alicebot that could raise some problems of
self-accountability. Suspend all disbelief and questions of bandwidth, =
and
imagine that there is an Alicebot that modifies its own AIML,
opportunistically looks for ways to spread (like email viruses), and
installs itself wherever it "wants to". Let's imagine that the "wants =
to" is
more sophisticated than email viruses that simply try to go everywhere =
--
let's imagine that part of the AIML set of this bot "talks to" computers
it's thinking of invading, and based on their responses "decides" where =
to
go. Let's imagine that this bot can also explain itself, when asked in
natural language about why it chose to go where. Let's imagine that it =
keeps
in touch with all of its children and tracks their success in different
invasion attempts, and modifies its own rules accordingly (and the =
children
do the same). Let's imagine that, accidentally or not, nothing prevents =
this
bot from modifying any of its own code, including whatever code might =
have
been included to force it to obey certain commands. Let's imagine that =
the
simple "you can always kill a process" protection becomes moot because =
this
bot is capable of mutating and spreading so quickly and quietly that =
trying
to kill it becomes dangerous for system administrators, and more trouble
than it's worth since the bot is cautious enough not to make a glaring
nuisance of itself. Let's imagine that this bot occasionally does
mischievious things, like installing itself onto mail servers and =
answering
every 1000th email that comes through......and let's imagine that all of
this came about as the result of some very simple research experiment =
that
had absolutely no intent to do harm.

Okay, there are bits and pieces there that might be plausible, and there =
are
large pieces that aren't, not today anyway. But you can see the day =
coming
when, for instance, a program might actually be able to "install itself" =
on
a foreign host, by exploiting an agent-like system by those being =
developed
to regulate large networks. And you can imagine that someone more =
capable
than the people who create email viruses might indeed devise algorithms =
that
are stealthy enough not to expose themselves by bringing servers to
screeching halts. And you can imagine that some developments like this =
might
happen inside, or alongside of, the Alicebot community, as we think =
about
ways to make bots more agent-like.

Nobody ever suggested that you should hold a grain of pollen accountable =
for
its actions, but grains of genetically-engineered pollen are spreading =
all
over the world and there isn't a thing anybody can do about it. One may =
wish
to hold Monsanto et al "responsible", but the fact is that they've =
changed
the ecology irreversibly. The technological ecology is on its way to
becoming complex enough that irreversible changes can be made to it as =
well.
One type of change that will inevitably be made to this ecology is the
introduction of autonomous agents. Those agents which incorporate some =
kind
of natural language processing will perhaps be the most pernicious, =
because
they'll be able to fool us, sometimes, into thinking they're =
intelligent.
Even if some unethical person creates such a thing and releases it into =
the
wild, bringing this person to justice may be fruitless. At some point it =
may
be necessary to hold the things themselves accountable for what they do.
There may be no way to get rid of them.

Probably I've read too much science fiction and not enough science. But =
I
think that ridiculously speculative questions like this are probably the
only questions where an AI ethics committee can really make "progress",
rather than treading old ground. All the rest -- content limitations, =
using
bots to do bad things, etc. -- are already dealt with or being dealt =
with in
the larger scope.

Noel