Recent research, including neurological experiments, has again raised the problem of individual agency and its opposition to strong causality. To summarize: DNA plays a very strong part in determining a wide variety of traits, including intellectual and emotional traits. The rest is determined by environmental influences, primarily peers. All of this shapes the neural connections and activity in the brain. Neural activity needed to implement a decision begins as much as a half second before the individual is aware of making the decision. This has been interpreted as evidence that the experience of deciding to do something (in this example, push a button) is an illusion, consistent with the view that actions are fully caused by the activity of neurons, in turn shaped and fully caused by DNA interacting with life experiences. This is all consistent with a philosophical position of strong determinism – after the causal effects of all the “external” factors have been accounted for, there is nothing left for the “person” to “decide.” A person holding a gun encounters an unexpected stranger: He does not decide whether to shoot the stranger or not; the outcome (shoot him or welcome him in) is fully determined by the interactions of neurons, themselves fully “programmed” by DNA + accumulated experience. Agency, the ability to decide on a course of action and the potential to be held accountable (morally or otherwise) for the results, is an illusion. The individual actually has no agency because even the most trivial action is fully caused by the inherited DNA and its history of interactions with the biological, physical, and social environment. I believe that the arguments supporting this “strong causal” view all lead to paradox – and some are based on self-contradictory assumptions. The most common error, even among writers who vehemently reject the idea of a “soul-stuff” independent of the material brain, or a “man inside” (“homunculus”), is to allow the homunculus to sneak back in. Searle’s “Chinese Room” thought experiment provides an excellent example: He posits an intelligence who receives strings of Chinese characters, looks them up in a code-book that provides responses to each string (early AI was a lot like this). Searle uses this as an argument against strong materialism in many of his early writings, never seeming to understand that he has smuggled the homunculus back into the room. The person is the “room,” the interaction among all the body’s neural systems, which includes perception and memory, shaped by accumulated social interactions and the cultural practices that govern and shape them. There is no other person. Neural science has never been able to pin down where and how these interactions are consolidated and integrated. But no-one has ever demonstrated a location for anything like an independent self – or a language of the brain that might be used by an independent self. So there is no “Chinese room” in the brain – Chinese, or any other language the person has learned, is understood and used by the interactions among several areas of the brain. The same is true of all other actions. Killing the “man inside” leaves the argument for strong causality at best only slightly injured. However, the strong causality argument runs into two paradoxes. One is learning: If the neural system is incapable of even the weakest form of independent agency, then it is impossible to learn anything new, as several philosophers have concluded. Chomsky resolves this paradox for language by positing that a fundamental knowledge of language is innate. But that solution leads only to more paradox: The ability of humans to formulate and learn new ideas seems virtually infinite, but the genome is finite – and no-one has found evidence to support an innate “language of the brain.” The other paradox derives from evolution theory. We know that efficient use of resources is a fundamental law of evolution, and the human brain is an extraordinarily expensive organ, consuming 20% of the body’s energy intake. If it is incapable of assessing a situation, considering alternative courses of actions, overcoming the over-learned (fully caused) impulses in order to choose and enact the most promising, then how else does the large, expensive cerebral cortex pay for itself? Organisms live in an environment in which new events often occur more or less at random. If somehow a member of an already brainy species appeared (as a result of random alterations in DNA) with the ability to suspend or escape strong causality long enough to represent the current state of the organism in its environment then project and evaluate the results of alternative actions and choose the one that seemed most promising, this would present a tremendous improvement in reproductive fitness. Assuming a stray lightning bolt doesn’t kill the brainy-plus person, this capability would spread widely. That is precisely the primary function Antonio Damasio posits for the neural system, including the brain. All very nice, a strong causality proponent might object, but even the emergence of this new power is caused by something, and in turn it is constrained by the causal power of the neurons themselves and the data they report from the environment. I think this argument risks veering back into something very much like the homunculus error, in that it seems to impart an independent causal agency to the concept of causality. In an algorithm solving a complex set of equations, introducing a random number will often break the system free of a sub-optimal solution (a “local minimum”) so that it moves toward a better solution. The very appearance of a non-predictable event is random (with respect to the organism), with the potential to disrupt the routine, over-learned and fully caused response patterns. The ability to suspend these patterns long enough to consider and decide among alternative actions would be – is – a powerful contribution to reproductive fitness. No-one has yet shown how consciousness works. There is some evidence of how neural columns represent predicted futures and compare them to incoming data, but how all this is coordinated to produce agency is pretty fuzzy. However, projecting, evaluating, and choosing among actions and the alternative futures they present is one good candidate for answering the question, “what is consciousness for.” In sum, yes, most of our routine actions may be fully caused by the interactions among neural interconnections that are themselves determined by DNA+environment+experience. But some of these neural interconnections are capable of responding to unpredictable events in a considered way, i.e., they have agency. Agency is much more limited than we think it is, but it is real, and it makes an important contribution to reproductive fitness.

Chatbots 2

So, chatbots can develop a nasty edge – just like humans.  And they have clearly passed – surpassed - the Turing test. There has been much discussion of whether that makes them “sentient” or in some sense even “conscious.” 

A common response is that AI reflects just the “scrapings” from everything on the internet which, of course, includes violence, sex, bigotry, hate and bullying along with the more benign content of friendly e-mail and texting, philosophy and science, poetry and humor.  

It is worth remembering that one of the primary ways humans acquire language is by encountering words and phrases many times in various contexts – which is also a primary way humans learn culture.  So to this extent, chatbots’ use of language is quite “human.” 

Humans also learn language (and culture) by association with the non-linguistic contexts in which they encounter words and phrases, including how others respond to their actions including their use of language.  Human language use and acculturation are also conditioned by biological and social drives and needs – hunger, sex, security, social contact, etc.  All of this entails the chemical environment of the brain / body, including oxytocin, adrenaline, and cortesol.  None of that context is part of AI training, except inasmuch as it might be reflected in the language people use to describe and respond to it.  It is difficult to imagine how this social, cultural, and chemical context might be incorporated into the training of AI, or what might substitute for the socialization this non-linguistic context provides human language learners.  

So we have created entities with a superhuman power of language – a power unconstrained by normal socialization.  Whether these entities qualify as “conscious” hardly seems important.  

What will they do? 

What will unscrupulous people – or well-intentioned but misguided people – do with them?  

Chatbots

            There has been quite a bit in the news recently about “chatbots,” AI programs that can engage in conversations and write essays, stories, or even poems on any topic, imitating any style.  Some of this discussion has been alarmist – a reaction I can understand, given the importance in my own teaching of student writing, including essay and short answer questions.  Before the holiday season I checked this out, giving the chatbot a typical midterm / final exam question on a couple of different topics.  As other commentators have noted, the results are pretty mediocre – correct grammar and spelling (in contrast to typical undergraduate writing) but unimaginative and dispirited (in contrast to the best student writing).  In a 200- or 300-level class, these essays would receive at least a C, maybe a B in lower-division classes.  So – take-home or do-at-home essay and short-answer questions have just become obsolete.  Hand-written in-class writing is still usable – but it can be painful to grade.  

            On the other hand, a couple of commentators have pointed out that the chatbot responses do provide a nice overview-level summary of current thinking and ideas about a topic, as well as an example of how a generalized discussion might be organized.  So – if these are accepted as a tool, students might be encouraged to begin with a prompt and chatbot response, then build an essay on that base.  Either way, I think it will require some fresh thinking about the purpose of student essay assignments, and perhaps about the learning objectives of humanities and social science classes.  

            In a future entry I will talk about another brief test in which I asked the chatbot to analyze metaphors in a passage or set of passages.    

Talking whale

I’ve been reading “How to Speak Whale” by Tom Mustill, and it is astonishing in many ways.  Overall, it illustrates the fact that, whatever topic you’re interested in, the number of researchers investigating that topic and the amount of new evidence they are producing (and publishing) is growing exponentially. That’s both bad news and good news.  The bad news:  everything I knew about animal communication and about research on animal communication just a year ago is obsolete. Forget about catching up or keeping up.  The good news:  if you have access to “big data” (huge data sets and the computer power to process them) you can answer questions that were impossible even to ask as recently as ten years ago.  And – something I vaguely knew that Mustill relates in some detail – “big data,” using advanced artificial intelligence learning systems fed by miniaturized observation instruments, is rapidly filling in what we know about the complex social lives and communication behavior of cetacians – and many other animal groups. 

I won’t even try to summarize all I have learned from this book.  The citations in the end-notes will keep my reading list overflowing for the rest of my sabbatical.  By then much of it will be already obsolete, and I’ll need to refresh it as well as I can. 

“How to Speak Whale”:  Highly recommended.  

Puzzles and language

  A recent article in New Scientist highlights some very interesting research by Gillian Forrester at Birkbeck, University of London.  She has been developing puzzles to test the ability of great apes (including human children) to use their hands to solve increasingly complex puzzles, designed to test conceptualizing abilities similar to those required for language learning and use.  The evidence she has garnered through this (and earlier) research supports the claim that sign language developed prior to vocal language, and further suggests a role for puzzle-solving ability.  This short video gives a flavor of her work: 

https://youtu.be/8edayRfe484 

If Robin Dunbar is correct in his claim that language evolution was driven primarily by the pressures of living in large and complex social structures, Forrester’s work would seem to suggest that solving social puzzles is related to solving physical puzzles.  I’m looking forward to the publication of Forrester’s work for further details!  

Fatigue

  I just read an interesting review (in The New Scientist) of recent research on fatigue.  I’ve always considered feelings of fatigue to be based only on depleted energy supplies and accumulated “waste” chemicals in the muscles.  Recent research suggests it is more complicated; several brain systems are involved. In addition to the interoceptive reports on available energy and depletion of energy, reward systems assess the potential benefits of continuing the activity. A separate system balances the energy supply with the anticipated rewards in an on-going “cost/benefit analysis.”  The experience of being fatigued is the result of the brain’s assessment that the potential rewards of continued physical or mental effort do not justify the energy expenditure.  Anything that reduces the assessed reward value of the activity (such as depression or boredom) increases fatigue; anything that increases it (such as social encouragement – a cheering squad or the offer of monetary rewards) decreases the feeling of fatigue.  This helps explain why depression, which reduces the subjective pleasure of almost everything, is so enervating.

I have often noticed that the last mile of a hike seems especially long and tiring, whether it comes after three miles or fifteen miles.  The rewards of the hike (pretty views, birdsong, etc.) are all past, and apparently the reward of reaching the car and resting feels about the same regardless of how long the hike was.   

Dreaming – does it necessarily have a purpose?

  In his recent work Antonio Damasio has argued that the brain has evolved primarily as an organ for representing the body in its social and physical environment, and projecting that representation forward into the future.  The representation of the body Damasio writes about is story-telling, dynamic not static. Other researchers argue that a primary function of consciousness is to consider the potential outcomes of alternative actions as a basis for making decisions.  These ideas together suggest that the stories we invent in our dreams – both daydreams and sleeping dreams – are an extension of the process through which we represent and project forward our body’s physical and social presence.  

In “When Animals Dream” Peña-Guzmán argues that animals actually experience their dreams, and takes it as evidence that animals have some form of conscious experience, though the quality of that experience is likely to be quite different from that of human consciousness.  

I think it is reasonable to extend Damasio’s account of mental representation to dream behavior in both humans and other animals.  Perhaps, as Peña-Guzmán and others argue, dreams provide a way of consolidating mental representation, rehearsing behavioral responses, or both.  But it is also possible that the dream behavior is simply a continuation of forward-projection, disconnected from actual sensory input, so drawing its substance from random fragments of memory, sometimes but not always including memories of recent (e.g. yesterday’s) experience.  If that is the case, dreams may be simply an epiphenomenon, lacking any function at all - sometimes disturbing, sometimes entertaining.