Delusions about Eugene (A reply to Andreas Schou)

Andreas Schou writes:

+Daniel Estrada finds this unnecessarily reductive and essentialist, and argues for a quacks-like-a-duck definition: if does a task which humans do, and effectively orients itself toward a goal, then it’s “intelligence.” After sitting on the question for a while, I think I agree — for some purposes. If your purpose is to build a philosophical category, “intelligence,” which at some point will entitle nonhuman intelligences to be treated as independent agents and valid objects of moral concern, reductive examination of the precise properties of nonhuman intelligences will yield consistently negative results. Human intelligence is largely illegible and was not, at any point, “built.” A capabilities approach which operates at a higher level of abstraction will flag the properties of a possibly-legitimate moral subject long before a close-to-the-metal approach will. (I do not believe we are near that point, but that’s also beyond the scope of this post.)

But if your purpose is to build artificial intelligences, the reductive details matter in terms of practical ontology, but not necessarily ethics: a capabilities ontology creates a giant, muddy categorical mess which disallows engineers from distinguishing trivial parlor tricks like Eugene Goostman from meaningful accomplishments. The underspecified capabilities approach, without particulars, simply hands the reins over to the part of the human brain which draws faces in the clouds.

Which is a problem. Because we are apparently built to greedily anthropomorphize. Historically, humans have treated states, natural objects, tools, the weather, their own thoughts, and their own unconscious actions as legitimate “persons.” (Seldom all at the same time, but still.) If we assigned the trait “intelligence” to every category which we had historically anthropomorphized, that would leave us treating the United States, Icelandic elf-stones, Watson, Zeus, our internal models of other peoples’ actions, and Ouija boards as being “intelligent.”

Which leads to not being able to express the way in which Eliza, a relatively simple stateless text parser which returns “conversational” results, meaningfully differs from a human. Which makes it difficult to define additional problems. Which makes the definition not necessarily helpful for that particular purpose.

I think the worries about anthropomorphism in this debate are entirely misplaced. But we need to step back and cover some basics.

First of all, let’s clearly reject the “naive capabilities” view. I think this view is a misunderstanding of Turing, and although it is close to the consensus interpretation of his test, I argue (in my dissertation) that the consensus view is a poor interpretation of Turing’s project. The “naive capabilities” view is traditionally presented as an alternative to an essentialist view of intelligence. Essentialists believe that there is some core nature that is constitutive of intelligence, and either you have it or you don’t. Essentialists can be mystics (minds = souls), but they can also be materialists; Searle is an essentialist about minds, which he thinks reduce to biophysical facts about brains. Since digital computers don’t share that causal structure, they don’t have minds, qed.

The naive capabilities view rejects essentialism, and argues that intelligence is a matter of behaving in certain ways. Turing’s test is typically interpreted as saying that if a machine can conduct a convincing (=indistinguishable from a human) conversation, then it is intelligent. At the limit is the so-called “Total Turing Test”, a machine that is behaviorally indistinguishable from humans in all respects. On this view, we might put together a list of capabilities that we identify as “quintessentially human” (playing chess, recognizing faces, engaging in petty office gossip, etc) and judge machines by how well they perform at these tasks. Although Andreas rightly points out that such a list is historically contingent, the view seems philosophically satisfying, especially if we’ve already rejected an essentialist view. Different historical communities judge human intelligence differently, so machines have to meet different conditions to qualify as intelligent within those communities. But the capabilities view lets us operationalize these standards and subject them to rigorous testing, and there are now a huge number of domains in which machines are clearly superior in direct human-machine competition. Operationalizing tasks is the bread and butter of AI, and even the harshest critics of AI agree that it has a long record of truly stunning successes.

The naive capabilities view is better than essentialism, but Andreas suggests it isn’t satisfying, in part, because of the open-ended nature of the list of necessary capacities. I think Andreas is still seduced by essentialism if he thinks a definitive list is required for philosophical satisfaction, but fortunately there is an alternative to both views that has not yet been explored in this discussion. But before explaining the alternative, we should reflect on how this even the naive capabilities view makes the worries about anthropomorphism irrelevant. Andreas worries that our greedy anthropomorphism inclines us to attribute agency to systems that aren’t really agents. You point to Eugene or Clever Hans or Eliza as examples of misidentification. But are these fair examples? None of them qualify as “indistinguishable from human behavior” under any rigorous scrutiny. They might immediately trigger a false positive (after 5 minutes of questioning among 30% of judges, the standard for a Turing Test), but any rigorous testing would reveal countless limitations behavioral quirks that do not constrain their human counterpart. Despite our greedy anthropomorphism, the distinctions loom large. The argument that greedy anthropomorphization has somehow blinded us to these distinctions is ridiculous.

The point here is that the Turing Test was never meant as a rigorous test for intelligence. It sets the bar far too low to be useful for thoroughly testing the capabilities of the machine. Instead, the Turing Test is meant as a test of *our* willingness to attribute intelligence to the machine. Some would say it is meant to test how well the machines can fool us, but this simply assumes the display is a deception. I’d rather say that the test is meant to examine our willingness to extend conversational charity to the machine. How willing are we to give this potential interlocutor the benefit of the doubt, and how easy is it to burn that bridge?

For the first few seconds of talking to someone with Werneke’s aphasia, you might believe that you are holding a meaningful conversation with mutual comprehension on both sides. But after a minute or so you start to realize that the words you hear are gibberish. Does this person pass or fail the Turing Test? Well, that’s ultimately in the eye of the judge, in how willing they are to play along. I can imagine some people (perhaps who have a special relationship with the person, like a granddaughter) entertaining the speech and carrying on a “conversation” for hours and hours, and find it meaningful and rewarding despite the difficulties in comprehension. I can also imagine a judge interested only in whether to check a box marked “intelligent” coming to a satisfying conclusion in only a few minutes of affectless interaction, well within the Turing test timeframe. Holding capacities constant, the results here vary as a function of the relationship between interlocutors.

And emphatically: the granddaughter’s continued conversation with the aphasic grandfather is *not* a mistake or a delusion, and it is not the result of greedy anthropomorphism. It is, much more simply, the extension of social care across cognitive difference. This is something human communities have always done. Our greedy anthropomorphism means we extend this social grace even beyond the bounds of the human, to domesticated pets and fictional characters and, now, the lively machines that inhabit our world. The Turing test assumes from the outset that we’ll extend the benefit of the doubt arbitrarily, even to machines that might not ultimately deserve it (Turing calls this “fair play for machines”). The real question is: which systems talk back well enough to warrant this social grace?

So I think my cat is intelligent, but that’s mostly because it does a lot of intelligent things. I’m a greedy anthropomorphizer so I tend to see more intelligence and agency than is there; I think my cat is crying because it “wants something” when more likely it is just making noise because it is bored and crying is something to do (and something humans seems to do a lot). But I don’t really anthropomorphize the cat radically beyond what it’s behavior warrants as a reasonable interpretation. I don’t expect my cat to keep promises, or to hold down a job, or that it’s flirting with me. These would be delusions: they are interpretations the behavior simply doesn’t support.

You seem to be worried that greedy anthropomorphization makes it so we can’t tell the difference between delusions and interpretations supported by the evidence, but this misunderstands the disposition to athropomorphize. Anthropomorphization is a prejudice, a bias, but one that is tempered by further evidence. I hear a creak behind me and I think someone else is in the room– until I look around as see an empty room, and realize my bias was mistaken. Greedy anthropomorphization isn’t stubborn or blind to correction: we check (and recheck) our results, and we tend to maintain the views our evidence can reasonably support. My prejudice makes me tend to consider some possibilities over others, but that doesn’t make me a dogmatist.

So after 5 minutes of questioning Eugene convinced 30% of judges it was human. It’s not good enough to convince 80% of judges after 5 minutes; and it wouldn’t convince 30% judges after 5 hours of questioning (the typical length of a full psychological assessment). It passed the Turing Test, but indistinguishability from human behavior this isn’t. Eugene is not evidence that human anthropomorphization has induced a mass delusion in attributing agency where it isn’t. Eugene is only evidence that the bar for passable human conversation is incredibly, almost laughably low.

But to a naive capabilities theorist, Eugene’s results are encouraging for the operationalist reasons discussed above. There are probably other dimensions of human behavior that are low-hanging fruit, where we can rack up some victories and close the game between human and machine behavior. For instance, Google and Facebook are apparently “racing to solve Go“. Not because it will provide insight into the mind or will promote the development of new technologies. On the contrary, the goal is simply to plant a flag in an as-yet unclaimed island, and to test the capabilities of technologies that are already widely in use. Eugene’s rather poor performance suggests that either the “race” to build a widely convincing chatterbot either doesn’t have very stiff competition, or it’s still just so damn hard that fucking Go looks like easier pickings.

Perhaps today’s deep learning methods are sufficient for cracking Go, but that has yet to be seen. As of right now, computers still can’t compete in a fair game at the highest levels. If nothing else, the low-hanging fruit of Go-playing is providing a battleground on which today’s software giants can display their AI skills. And for these socioeconomic and historically situated reasons, the game of Go has become an important marker of intelligence in our community. That doesn’t tell us about “intelligence” so much as it tells us something about what practices we find meaningful. 50 years ago chess was hard enough and go was basically impossible; holding AI to the standard of Go-playing was as unreasonable as expecting my cat to floss. Today those standards have changed, and go-playing is less a theoretical challenge and more a reason to grab headlines in Wired. Not only has tech changed, but the standards by which we judge it have changed, dramatically. We’ve consistently raised the bar because our tech has consistently met these challenges.

And here we see the cracks in the “naive capacities” view. We don’t yet know the full capabilities of our existing technologies! If capacities are something you “just have” independent of exercising them, then this is just essentialism by another name. We ought to be wary of attempt to compile lists of necessary capacities for the same basic reasons we should be wary of any essentialist view, in that it ignores the situated and interactive nature of existence. It is much better to talk not about what a system can do in the hypothetical sense, but what it actually does in the everyday sense. Intelligence isn’t just about having a capacity, it’s about using your capacities to do things; and we typically don’t know what we can do until we try. Anyone that can throw a ball can also open a jar; in this sense, our ancestors 50,000 years ago had the “capability” to open jars. But since they never found themselves in circumstances to exercise this capacity, it would be very strange if “jar-opening” (or car driving, or chess playing, or…) appeared on a list of their capacities.

The point here is not just that the list of necessary and sufficient capacities has changed over time; that’s obvious enough. The much deeper point concerns the nature of social cognition itself. The alternative to both essentialism and the naive capacities view, sometimes called Enactive Cognition, is that intelligence is a matter of participating in certain cooperative social practices, and that these practices are dynamical, interactive, and have flexible conditions on participation and engagement. There is no minimal agent, no necessary or sufficient capabilities required to be considered a thinking thing in the general case. There are only the particular games we play as a community, and the many overlapping games and communities with which we’re engaged. Ontologically, this framework covers “social” relationships at all scales of analysis (it is fundamentally just network theory, the ur-theory), so this can get confusing if we’re in the business of rigidly demarcating games and communities and scales. But in general it doesn’t matter what you are; it doesn’t even matter what you can do. What really matters is if you can play along.

The best example I know for making this “playing along” not confusing is the experiment that originally set off the enactivist paradigm, the perceptual crossing experiment. In this experiment they strip the social interaction to the bare minimum: a one-dimensional space of interaction, one bit of communication between objects. Despite the minimal environment, participants in the experiment were easily able to recognize other agents and distinguish them from objects that move indistinguishably from the humans in that space. They did this not by looking for any particular property of the agent in isolation, but instead by looking for the reactive, dynamical feedback between agents characteristic of agent-seeking behavior. Since this post is already long I’ll spare the details you can read above. The upshot can be grasped by imagining the following: Say you and I are standing in front of the Mona Lisa, and I whisper “her eyes follow you!”. What do you do to check if I’m right? You move around, usually quickly wobbling back and forth, watching to see if her eyes wobble along with you. The property you are looking for is dynamical responsiveness to your own dynamic behavior. You’re dancing, and looking to see if the system dances back. In this way, the recognition of intelligence is a fundamentally cooperative act. Greedy anthropomorphism leads you to extend the offer to more things than can properly dance back, hence pole dancing; but there’s a world of difference between dancing with the pole and dancing nearby it. That difference rests entirely in the interaction between the two systems.

Turing’s writing on AI is motivated by a worry that our prejudices against machines will never allow us to take them seriously as participants in our social practices, regardless of their performance. We expect perfection from a calculating machine and take its errors as evidence that it lacks even rudimentary intelligence, although we’d never hold a human to the same standard. Against this prejudice he advocates “fair play for machines”. He proposes the imitation game in order to level the playing field and use human prejudices to work in favor of the machine. The Turing test assumes that we’ll extend conversational cutesy to machines that ultimately don’t deserve it, in an effort to see how well they play along. The Turing Test is a way to check our greedy anthropomorphism, not to indulge it. He imagines that not only will computers improve their performance over time, but that playing these games will also result in a change in our social attitudes towards machines. As we become more familiar with the machines, we’ll set more reasonable expectations on what roles they might reasonably play in our social spaces. We’ll tune our prejudices away from wide-eyed science fiction and towards the particular machines that actually inhabit our world.

On this view, building a convincing chatterbot matters because of longstanding social practices that recognize the value of good conversation. We want a machine to play this social game with us because the game itself matters to us and there are a woeful lack of sympathetic ears. A chatterbot that remained convincing over longterm enagagement, one that people wanted to talk and listen to as they would a friend,  wouldn’t just be an awesome demonstration of AI, it would be an incredible tool with vast social use. It would change the face of clinical psychology and customer service and youtube comment threads. The goal of a chatterbot is not just to display an adequate level of intelligence; it is to fill the role of intelocutor, and unbelievably complicated social dance. The dynamical feedback and coordination required to successfully pull of a meaningful conversation is probably so close to the complexity required for full-blown moral personhood that I’m not sure it’s worth trying to distinguish the two.

We want machines that can play chess not because chess-playing has metaphysical significance but much more simply because we have traditions for recognizing success at chess with respect and deference. When the machine performs at similar levels, giving it similar sings of respect and deference isn’t a matter of anthropomorphization at all; it’s just fair play. That a machine can beat a human at chess is interesting for the same purely historical reason it is interesting that Obama is the first black president: it has never been done before. This doesn’t tell you much of anything about blackness or the office of the presidency; the interesting stuff isn’t in the categories, it’s in the history that brought us here. And AI is increasingly part of our history. It has consistently been brought in to participate in our reindeer games, and consistently shows us up across a variety of dimensions.

But intelligence isn’t a matter of besting us across all dimensions. Instead, the goal is is integration in the social fabric, “passing” in the queer theoretic sense. Passing isn’t deception, and it doesn’t trade on delusion. Passing is the outsider’s technique for navigating social environments that are hostile to difference. Turing’s imitation game is all about passing; the Turing Test is designed precisely for making it easy for machines to pass in this sense. This goal is made difficult by the persistent prejudices towards machines, unfortunately fueled by misunderstandings of Turing’s work, that insist on looking the gifthorse of AI in the mouth with a Voight-Kampff machine to make sure their dental records match ours. It is the AI equivalent of asking to look in your pants before I let you use the bathroom, when fair play dictates: “you gotta go, you gotta go.”

Submit a comment