On Apes and AI
The common thread between “Ape Language” and AI is a striking human ability to engineer the circumstances in which self-deception thrives.
By Vincent Carchidi
Koko and Ape Languages
Do you remember Koko the Gorilla? The name might sound familiar, but here’s a refresher: Koko was a female western lowland Gorilla who had the unusual privilege of being taught a human language, namely American Sign Language (ASL). Her caregiver and teacher, Francine Patterson — an animal psychologist — undertook with her colleagues “Project Koko” which ensured that Koko was exposed to human spoken language from six months of age and to ASL from one year of age. The researchers talk of Koko’s acquisition of ASL as one might when studying a human child, noting how Koko began learning and using signs for “food,” “drink,” “more,” and signs for “large,” “small,” “same,” and “different” via “the sizes of the glasses she used at meals and snacks.”
Koko’s penchant for food-related signs is no accident — the poor Gorilla was living in an entirely different cognitive reality than her human caretakers. Her caregivers and researchers did not see Koko this way, though, and were frequently quite taken with Koko’s apparent abilities, believing the gorilla really had acquired ASL.
Yet, as linguist Geoffrey Pullum so nicely put it, “Koko never said anything: never made a definite truth claim, or expressed a specific opinion, or asked a clearly identifiable question.” Koko did not use language so much as she “[flailed] around producing signs at random in a purely situation-bound bid to obtain food from her trainer.”
Much to the chagrin of linguists like Pullum, Sherman Wilcox, and Noam Chomsky, Koko’s inability to actually acquire human language made little difference to her public image. In 1988, Koko warmly greeted Mr. Rogers; in 2001, she signed “tickle” when meeting Robin Williams, who was, of course, happy to oblige; and by 2015, well into her old age, Koko appeared in a video put out by The Gorilla Foundation for presentation at that year’s Paris climate change conference, imploring the world through sign language to save the planet, telling humans she has “love” for them but that they are “stupid.”
Seeing Humanity
Koko might be on to something there.
More seriously, there is something happening here — and it’s not animal psychology. As Robert Sapolsky explains, Patterson, when communicating with Koko, would frequently attribute meaningful language use to Koko’s signs even if Koko herself produced the wrong signs in response to a question. Importantly, the parameters of Koko’s testing were often shockingly lax, allowing the gorilla to fumble through signs again and again until something approximating a correct, human-appropriate gesture was made — or overinterpreted by the humans.
Herbert Terrace, who played a major role in the downfall of “ape language” research, noted something interesting along these lines: when human caretakers would interact with and test the sign language abilities of one notable chimpanzee, “Nim Chimpsky,” the humans would subtly prompt Nim “with appropriate signs, approximately 250 milliseconds before he signs” in order to provide the chimp with the contextual resources it needed to sign in response.
(Fun fact: Terrace was initially determined to prove Chomsky‘s approach to language — that it is a distinctively human possession — wrong. He later came to be at odds with ape language researchers like Patterson.)
Consider what this means: for years, the study of ape language proceeded in part because human researchers — intelligent, educated people — were imposing onto chimps like Nim and gorillas like Koko meaning that simply was not there. Human language was reduced in the course of this research from a rich system of rules that non-arbitrarily link sound (or signs) with meaning; the spontaneous and principled expression of new thoughts that may or may not hold any identifiable connection to one’s immediate environment, yet correspond with the thoughts of others.
Indeed, this reduction of human language extends well beyond these apes’ immediate circle of human caretakers. In one 2018 obituary for Koko, the author expressed how ‘remarkable’ and ‘poignant’ it was for Koko to sign the words, “you key there me cookie.” Yet, only if one squints does “you key there me cookie” express a mentally represented and human-like thought, rather than a jumble of just-close-enough gestures that have historically earned the ape some more food.
AI Is All About Us
To be abundantly clear, Koko herself was doing something notable with her signs and likely improved the public image of apes— but whatever Koko was doing, it wasn’t human language. More than this, it is difficult to believe that we could ever know what Koko understood herself to be doing. Other species have cognitive realities we do not share, and cannot access. Even with animals with whom humans quite clearly can communicate in some way, like dogs, what a dog understands a howl to be — from the dog’s perspective — is off-limits to us (even though it often seems like we do know).
At the risk of trotting out more than my fair share of clichés in this essay, I believe something similar to the reduction of human language and distortion of our cognitive reality is happening in Artificial Intelligence (AI) and Large Language Models (LLMs).
There have been a lot of attempts to explain why AI models like LLMs are so readily anthropomorphized; that is, why some people — who are intelligent and educated — observe or interact with such models and come away believing they have witnessed something decidedly human-like or on the verge of humanness. My remarks here are not to disparage these folks — we are all susceptible to this in one context or another.
I follow, for example, the commentary of Wharton School Professor Ethan Mollick, who is notably bullish on generative AI and its future. Even a brief perusal of his Twitter feed will give you the (mostly correct) impression that virtually every day Mollick promotes yet another fantastical application of generative AI to one’s daily life.
Interestingly, Mollick is among many online AI enthusiasts who frequently lament that not enough people, “from business executives to scientists,” have “even tried a GPT-4 class model. Less than 5% has spent the required 10 hours to know how they tick.”
It’s easy to dismiss Mollick’s comment if you are not continuously wowed by GPT-4 and comparable models. What, though, accounts for the disparity between those who are and those who are not?
The answer, I believe, is that the same phenomenon is at work in both the study of “ape language” and the belief that LLMs exhibit human-like intelligence.
Put another way: AI, as it exists today, is all about us.
The idea of trying to teach a non-human animal a human linguistic ability that it has never expressed any interest in or propensity to develop on its own is downright strange when you stop to think about it. Would it be amazing if a gorilla could construct new thoughts and express them to us in an intelligible manner? Absolutely! Yet, it is akin, using Noam Chomsky’s favorite example, to teaching humans the communication system of another organism, like bees. I’ll let him explain:
It’s an insult to chimpanzee intelligence to consider this their means of communication. It’s rather as if humans were taught to mimic some aspects of the waggle dance of bees and researchers were to say, “Wow, we’ve taught humans to communicate.”
He continues:
Humans can be taught to do a fair imitation of the complex bee communication system. That is not of the slightest interest to bee scientists, who are rational, and understand something about science: they are interested in the nature of bees, and it is of no interest if some other organism can be trained to partially mimic some superficial aspects of the waggle dance…We would learn nothing about apes from the fact that grad students can be trained to more or less mimic them…just as we learn nothing about humans from the facts that apes can be trained to mimic humans in some respects.
Chomsky concludes:
This is all sentimentality of the worst sort.
I believe Chomsky is correct about this, right up until he reaches his conclusion: this is more complicated than sentimentality.
When enthusiasts (for lack of a better word — I count myself as an AI enthusiast, after all) see humanity in LLMs, they often construct the circumstances in which they allow themselves to see it. They do not, however, necessarily do this consciously, and are baffled and even frustrated when others do not see the obvious humanity of an LLM’s responses. There is little doubt that ape language researchers were sincere in their efforts, too — but it was they who cast the spell, not the apes who quite innocently wanted food and tickles (and cats!).
The Psychic’s Trap
Ape language researchers — at least the ones who kept the project going for as long as it did — were constantly pushing their ape subjects in the direction they wanted: nudging them through contextual clues, including their own gestures, to produce signs that would indicate the chimp or gorilla is constructing a new thought through human language. When the signs were incorrect, the human researchers readily imposed their own meanings on them, filling in the conceptual gaps where the ape fell short.
The researchers were, in effect, fooling themselves.
An essay making the rounds in recent days by software engineer Baldur Bjarnason (written back in July 2023) offers what I consider the most compelling explanation for this tendency to fool oneself about the presence of human-like intelligence that I have personally come across.
Bjarnason’s essay is fantastic, and I encourage you to go read it separately, as I cannot do it justice here. As I read it, I immediately got the sense that this was written by someone who finally gets it; why some are captivated by LLMs and others are not. Others, like Grady Booch and Gary Marcus, have seized on Bjarnason’s piece recently as an explanation for the divergence we are concerned with here.
His central argument is that some individuals’ interactions with LLMs involve a dynamic that is conspicuously similar to the dynamic between a psychic con artist and their marks. In observing the remarks that enthusiasts make about LLMs (e.g.,“There really is something there. Not sure what to think of it, but I’ve experienced it myself.”), Bjarnason says: “This specific blend of awe, disbelief, and dread all sound like the words of a victim of a mentalist scam artist — psychics.”
What does he mean by this?
His argument overlays the meticulous process by which psychic con artists study and manipulate their marks using statistically likely information onto the process by which individuals come to view LLMs as human-like. Psychics do this, for example, by monitoring the social media behaviors of event attendees (the marks), having staff during pre-event socializing collect tidbits of information about the attendees for later use, etc. During the event itself, psychics engage in “cold reading” in which “many people will interpret even the most generic statement as being specifically about them if they can relate to what was said.” The trick of course is that, with a little help from the background research, the psychic makes statistically likely comments or questions for the marks, and the marks — eager to believe in the psychic’s power — will tolerate inaccuracies and convince themselves that the psychic is truly reading their minds.
The most fascinating thing about this process is that those who see themselves are more intelligent are more likely to fall victim to the psychic’s tricks. As Bjarnason explains:
Somebody raised to believe they have high IQ is more likely to fall for this than somebody raised to think less of their own intellectual capabilities. Subjective validation is a quirk of the human mind. We all fall for it. But if you think you’re unlikely to be fooled, you will be tempted instead to apply your intelligence to “figure out” how it happened. This means you can end up using considerable creativity and intelligence to help the psychic fool you by coming up with rationalisations for their “ability”. And because you think you can’t be fooled, you also bring your intelligence to bear to defend the psychic’s claim of their powers. Smart people (or, those who think of themselves as smart) can become the biggest, most lucrative marks.
People who believe themselves to be more intelligent are more likely to aid the psychic in their attempt to deceive them, using their own intelligence to complete the deception even though they sought initially to deconstruct the psychic’s alleged powers.
Below is Bjarnason’s graphic illustrating this process from start to finish in the AI context:
From this vantage point, the remarks of someone like Mollick make sense. Mollick is an educated, intelligent person who seems to find it outright confusing that others do not see the intelligence embedded in GPT-4. The tendency to toy with LLM-powered chatbots until the desired output is attained is strikingly similar to the mark’s tolerance of the psychic’s vagueness. We have all seen, by now, the back-and-forths on social media (especially Twitter/X) about this — “if you ask GPT-4 this way, it gets it right!”
Yet, those in this position may have selected themselves for the psychic’s trap. In doing so, they make AI a matter of intimate and unearned familiarity — it’s all about them.
What the Future Holds
Ape language research persisted for years. And, if I am correct that the same phenomenon undergirds this research and the perception of human-like intelligence in LLMs — and if Bjarnason is correct that this phenomenon is the psychic’s trick — then there is little reason to expect that disappointing developments in future research within this vein will break the spell.
It virtually guarantees that debates about human-like intelligence and the possibility (or actualization) of “artificial general intelligence” will continue without resolution or hope of commensurability with the release of models like GPT-5, no matter how much of an improvement this model is on GPT-4.
I close by emphasizing that, just as I have no problem with Koko or Nim, I also have no problem with LLMs. What I do take issue with is hype — endless, boundless, indulgent hype that likely does more harm to the trajectory of AI than good, just as excessively fixating on teaching chimps and gorillas human languages harms the study of non-human apes. There is plenty in humans’ primate cousins to appreciate without forcing ourselves into it. The same is true of AI — and I hope becomes more true over time as we shed our anthropomorphization of it.
