Skip to main content

Why I don't like the term "AI"

Content note: I replicate some ableist language in this post for the sake of calling it out as ableist.

In games research, some people take pains to distinguish artificial intelligence from computational intelligence (Wikipedia summary), with the primary issue being that AI cares more about replicating human behavior, while CI is "human-behavior-inspired" approaches to solving concrete problems. I don't strongly identify with one of these sub-areas more than the other; the extent to which I hold an opinion is mainly that I find the distinction a bit silly, given that the practical effects seem mainly to be that there are two conferences (CIG and AIIDE) that attract the same people, and a journal (TCIAIG - Transactions on Computational Intelligence and Artificial Intelligence in Games) that seems to resolve the problem by replacing instances of "AI" with "CI/AI."

I have a vague, un-citeable memory of hearing another argument from people who dislike the term "artificial" because it presupposes that biological matter is fundamentally different from digital. I'm a bit more sympathetic to this argument. "Computational" seems like a fine substitute, except that apparently it means something else. ;) I suppose I'm partial to "synthetic" (vs. "organic").

But ultimately, it's not the first word in "AI" that bothers me, that makes me hesitant to adopt it as a field I identify with -- it's the second one, intelligence. My issue is not just that "intelligence" is poorly defined and hard to measure, but actually that it highlights everything I find culturally wrong with computer science as a field: a false dichotomy and prioritization of the "smart" over the "dumb", the "rational" over the "emotional", and a supposition that these qualities are immutable and acontextual.

Fundamentally, the language of "intelligence" is ableist, as Tim Chevalier explains well.
Popular culture seems to like the "innate intelligence" idea, as evinced by movies such as "Good Will Hunting". In that movie, a guy who's had no social interaction with a mathematical community bursts into a university and dazzles everyone with his innate brilliance at math, which he presumably was born with (for the most part) and put the finishing touches on by studying alone. The media seem to be full of stories about a bright young person being discovered, a passive process that -- for the bright young person -- seems to involve nothing except sitting there glowing.
Computer scientists can at times be obsessed with the idea of the loner auteur, the one person (usually white guy) who just has to work alone at a blackboard for long enough until he understands what no one else can and instantly changes the world with his discovery. In this narrative, a natural order emerges in which people without such gifts have a responsibility to simply identify intelligent people and bring them into their rightful privileges.

Meanwhile, in the real world, intellectual efforts succeed by collaboration and social support, and by work one spends communicating and disseminating their ideas -- the time one spends on which can completely overshadow time spent actually programming, proving, or measuring -- and subsequently refining them upon critique and peer review. And believing intelligence is immutable leads to poorer performance on intelligence tasks. And I could link to a wealth of research discussing the role of social cues and support in academic success. The evidence is overwhelming that "intelligence," under all definitions by which it's measurable, is not a static nor inherent property of an individual.

Lindsey Kuper also makes the argument that we should say "experts" instead of "smart people" to highlight the fact that expertise is relative, and privileging expertise in one specific thing over another often isn't productive when the point is to communicate well between different domains of expertise.

So why does this matter for AI? Mainly because I see it as a flawed research goal. To the extent that AI is about re-implementing human-like abilities through algorithms, why is intelligence the one we want to focus on?

I have no interest in creating Synthetic Idealized Bill Gates or Robot Idealized John Nash, at least not for the sake of their "intelligence." We have to be asking what AI is for. It seems like if you want a synthetic intellectual, you probably want it to be a good collaborator. And I have no evidence that "intelligence" as typically defined is a good predictor for collaborative skill.

On the other hand, I like working with proof assistants and automated deduction systems (e.g. Twelf, Agda, Coq, Prolog, ASP, Ceptre, STRIPS planners). It's not because these things are intelligent but because they do something (a) predictably and (b) quickly. I like generative methods because they, like young human children, repeat the things I tell them in novel and amusing configurations. I wouldn't call that intelligence, but I might call it creativity or whimsy.

I'm finding it an interesting exercise to go through Tim's list of alternatives to intelligence and identify which "synthetic" versions exist or don't:

  • Synthetically curious: web crawling, stochastic search, query generation
  • Synthetically hard-working: most automation systems
  • Synthetically well-read: text processing, natural language processing (NLP), narrative analysis
  • Synthetically knowledgeable: expert systems, commonsense reasoning
  • Synthetically thoughtful: automated deduction? What about other kinds of thoughtfulness (e.g. consideration for diplomacy or emotional context?)
  • Synthetically open-minded: // TODO
  • Synthetically creative: See: ICCC conference
  • Synthetically attentive to detail: most AI
  • Synthetically analytical: static/dynamic analysis; NLP; image/sound processing...
  • Synthetically careful: software verification
  • Synthetically collaborative: mixed-initiative creativity tools; interactive theorem provers
  • Synthetically empathetic: Affective computing? // TODO
  • Synthetically articulate: narrative summarization; natural language generation?
  • Synthetically good at listening: chatbots/dialogue systems, ideally // TODO

A secretly great thing about being more "politically correct" (as in: more considerate of the language we use) is that it's really about being more precise and concrete, which in turn is a great mechanism for generating research ideas.

Edit to add: I also think it would be rather interesting, if someone isn't doing it already, to try to model human cognition that we label as *less* intelligent, or neuro-atypical. What would a computational model of autism or dissociative identity disorder look like? How might we represent a computational agent experiencing anxiety, depression, or trauma? Answers to these questions could also lead towards new answers to some of the questions above, like how synthetic empathy or social skills might work.

Comments

  1. Reading your last point, about modeling atypical human cognition, I was reminded of some work done by my friend Uli at UT Austin. He focused specifically on the effect of schizophrenia on language use in storytelling. This article is probably a good entry point: http://www.cs.utexas.edu/users/ai-lab/?grasemann:cogsci11

    (And the rest of Uli's publications are here:
    http://www.cs.utexas.edu/~ai-lab/people-view.php?PID=101)

    ReplyDelete

Post a Comment

Popular posts from this blog

Using Twine for Games Research (Part II)

This preliminary discussion introduced my thoughts on using Twine as a tool for creating prototypes for games research. I'll start with documenting my first case study: a hack-and-slash RPG-like setting where the player character has a record of attributes ("stats") that evolve through actions that turn certain resources (money, health, items) into others. I've selected this hack-and-slash example because it falls outside the canonical "branching story" domain thought to be Twine's primary use case, but it is not too much trickier to implement. It relies crucially on the management of state in ways that simple branching stories would not, but it does so in a fairly straightforward way.

If all goes well, this post may also serve as a tutorial on the "basics" of Twine (links + variables + expressions). In particular, I'll be using Twine 2/Harlowe, and I haven't seen many tutorials for this new version published yet.

To me, the main "…

Using Twine for Games Research (Part III)

Where we last left off, I described Twine's basic capabilities and illustrated how to use them in Twine 2 by way of a tiny hack-and-slash RPG mechanic. You can play the result, and you should also be able to download that HTML file and use Twine 2's "import file" mechanism to load the editable source code/passage layout.

Notice that, in terms of game design, it's not much more sophisticated than a slot machine: the only interesting decision we've incorporated is for the player to determine when to stop pushing her luck with repeated adventures and go home with the current spoils.

What makes this type of RPG strategy more interesting to me is the sorts of decisions that can have longer-term effects, the ones where you spend an accumulation of resources on one of several things that might have a substantial payoff down the road. In a more character-based setting, this could be something like increasing skill levels or adding personality traits.

Often, the game-…