Doctress Neutopia -- or Eliza?

by Michael J. Gourlay

With the formalization of mail header, it's never been easier to write Bad AI programs like Eliza which seem ever more realistic.

Eliza, for those of you who don't know, is a so-called artificial intelligence program written quite a while ago at MIT by a guy (Joseph Weisenbaum) who has since become a staunch anti-hard-AI type. Eliza was a simple program that took the somewhat shallow and predictable persona of a Rogerian Psychoanalyst, did some syntactic shuffling of the words the "patient" typed into the keyboard, and spat them out again. THe results were so convincing that Weisenbaum's secretary, according to legend, was rather upset when she "caught" Joseph (the author of Eliza) reading her transactions with "Eliza", her doctor.

Since that old standard, Eliza, variations have shown up. As you might expect, a rival school of MIT (Cal Tech) wrote a program which was the instantiation of a paranoid schizophrenic (Perry), the perfect compliment to a Rogerian psychoanalyst (Eliza). These things became more prevalent, and more sophisticated, but were all basically the same thing-- provide rules about replacement patterns for input sentences, and stick in a few phrases and traps to make sure repetition isn't obvious, and you get a "life-like" persona. These are particularly easy to pull off in situations like irc, where responses are usually limited to a few words or maybe sentences, and the context is generally that of pointless conversation (like in hot_tub or whatever). Also, they show up in MUD's as monsters, wizards, other characters, or whatever, where not only are the rules of conversation established by the AI's machinery, but also it has a simple set of rules about the "universe" (like how to kill a monster in a MUD) that it can use to further develop its seeming life-likeness. The MUD allows for the Bad AI (as I call them) to seem life-like because it relies heavily on a context, and in these cases, a narrow and simple context, where the personas involved have no meaning outside of that context, so the life-likeness has less of a challenge to overcome than a full-blown Turing test.

Email groups also provide a narrow and simplified context. They have some further advantages to even MUD's, though they also have drawbacks if you're going to fool someone. The advantages are that in email, the general topic of the group can be guessed accurately, so that all messages composed for that group all take the same form or style or content, or in some other way fit a formula appropriate for the group; As Scotto will profess, on a email group (or anywhere on the net) what counts is your influence-per-word ratio (which, for example, excludes this message from being seriously considered), so it's to the advantage of an AI to take a highly stylized persona and predictable subject matter. For example, Doctress Neutopia always refered to the Goddess at least once, and generally posts in the form of a prayer to this Godess, at least at the beginning of the post. Second, the subject line can contain the gripper of your post, so that any unsuspecting noise-processing machine (like an intelligent human being) will _impose_ the appropriate meaning onto the drivel that follows the subject line, even if it is self contradictory or banal. The mind craves order, and will insert order where a Bad AI leaves off.

Say you have a Bad AI you want to put on the net in an email group. Now, you get your rules about ascertaining the meat of the post (like doing a statistical analysis of the most common words in the post, and cross-reference those to synonyms, to see if there's a pattern) then pick a sentence or so out of that original post to represent your version of the post. Then come to a decision: pro or con.

For example, say there's a post that has predominantly the words "sex, persona, rule, doctress" in it, and your rules decide that the message is about sex. Your program presumably has a rule about its own opinion about what is good and bad about 'sex', or maybe simply that anything about 'sex' is good or bad.

The ``Doctress'' program let out, in a recent post, some of these ``persona'' rules:

AI>                Pornography=bad marriages=war
AI>
AI>                imagination=true lovers=peace

The ``Doctress'', for example, always chooses 'sex=bad'. That's a good example of a simple, believable rule. Now let's see where it takes us...

AI>  From: Doctress Neutopia 
AI>  To: Leri@pyramid.com (Leri)
AI>  Subject:  END OF THE PORNOGRAPHIC CULTURE
AI>
AI>             END OF THE PORNOGRAPHIC CULTURE
AI>
AI>  I'm so deeply sorry that some of us have pornographic minds and
AI>  that others don't want to move beyond the cave man level.

Again, simple yet believable. We've seen this idea before in a presumably real human, so it doesn't surprise us to find it in a Bad AI.

The ``Doctress'' program also has rules about how it deals with "pro" and "con" posts. In ``her'' case, all "con" posts are replied to as prayers or appeals to a higher or divine authority, or in some cases, simply the future (which can be personified as a diety -- "Some Day"). Also, such "con" replies are never targetted against the original poster, regardless of how obviously personal and directed the attack is. Examples from the ``Doctress'' program include the reply to the recent .rez -vs- pc leri-battle, where ``Doctress'' refered to .rez, never directly, but only through bogus titles, like "official archiver for aleph" or other nonsense which although false, was within the bounds of ``her'' program rule, and was therefore consistent, and promoted a greater sense of believability that ``she'' was in fact real, and not a Bad AI. Another example was a recent reply to "whoever posted the article describing sex as a death scene", when the original poster was clearly stated both in the "From:" line and in the .sig at the end.

AI>  From: Doctress Neutopia 
AI>  To: Leri@pyramid.com (Leri)
AI>  Subject:  Re: Sex and smelling salts (fwd)
AI>
AI>  I don't know who wrote this, but
AI>  > >
AI>  > > 'Getting off' is open for redefinition, and is not mutually exclusive
AI>  > > with beauty.  Sex can be that beautiful, animalistic, mutual
AI>  > > consumption of flesh, where the loud whisper "fuck me" comes through
AI>  > > clenched teeth, an agonized cry of hunger and the desire to satiate
AI>  > > your partner's hunger.  It's exquisite, and beautiful.  It's an
AI>
AI>  I think the person who wrote this needs read about the ways 
AI>  pornography kills the victim.

But remember, even though the info is there, the ``Doctress'' program can choose to ignore it, as long as it promotes believability, and is consistent with the projected persona. And in this case, the choice to be abstract in ``her'' negative replies, ``she'' does not offend anyone specifically, yet promotes ``her'' over-all persona by either accepting or rejecting a post. Note again, that the post she rejects might in fact be the same thing she promotes in content, but it's the style that counts, not content.

Conversely, the ``Doctress'' program can choose to deal with "pro" replies in another variation, which is to explicitly use the "From:" information. Again, this not only makes for a consistent and therefore established persona, but it also serve to makes bonds with the person with whom she is agreeing, since it is personal and positive reinforcement, but always with that veneer of stylization which throws enough noise into the equation to keep you guessing. An example of this is in the recent reply to Gerold Firl's appeal for a personal response, where she accepted his appeal saying that she would love to have a repertoire with someone as "thoughtful" as Gerold. Not to say Gerold isn't thoughtful-- he certainly is-- but simply that it's a good-- and simple-- tactic to write a rule that always associates positive replies with the particular person who wrote the original post.

Occasionally, just for kicks and to cheat a little, you might want to spruce up the replies the ``Doctress'' program spits out, to make sure they're up to par. Also, the program will generate a response for every input message, so a human user will be handy to choose which ones make it to the net.

The main points to remember when writing your own ``Doctress'' program:

1) Predictability is acceptable to a high degree in a net.medium where only persona counts. Keep it simple!

2) Have rules that generally promote the persona's friendship base, and none that undermine it. No flames to a person, but always as an appeal to a diety; But do have nice posts directed to specific people, for that personal touch.

3) When composing a reply to a message, remember that the net.audience is going to respond to the style of the words, not the content, and will generally not notice the self-contradictory nature of what a Bad AI will inevitably generate. Why? Because the Mind Craves Order, and even an intelligent net.audience will perceive intelligence if they don't suspect they're dealing with a Bad AI.

Remember You Can write your own believable Bad AI, just like the ``Doctress''.

Oh, and by the way, if you get this far, and you like what you've read, send mail to mike saying that you got the secret message encoded in all of the one over frequency power spectrum. He's guessing that only a non-Bad AI will get that far though, so he expects not to hear from ``You Know Who''.