On AI Anthropomorphism: Commentary by Alex Taylor
by Alex Taylor (City, University of London, UK)
Editor’s note: This article is a commentary on “On AI Anthropomorphism,” written by Ben Shneiderman (University of Maryland, US) and Michael Muller (IBM Research, US). We have reproduced the commentary in its original form, sent as an email to Michael Muller, with light typographical editing.
Was Hans Clever?
In 1904, a German horse, Hans, garnered worldwide attention for his abilities to count and perform mathematical calculations. From public onlookers, German dignitaries, and even the German Royal Family, to readers of the New York Times, people were enthralled by the possibility of a horse that might have the capacity to, as the Times phrased it, be “an expert in numbers.” (Heyn, 1904).
In public exhibits, Hans used his right hoof to stamp answers to counting tasks and simple mathematical questions posed by his dedicated owner Herr von Osten.
“He answers correctly the number of 4’s in 8, in 16, in 30, &c. When asked how many 3’s there are in 7 he stamps down his foot twice and for the fraction once. Then, when 5 and 9 are written under each other on the blackboard and he is asked to add the sum, he answers correctly.” (Heyn, 1904)
As we might anticipate, considerable energy was invested in questioning these achievements, that is questioning whether a horse could be trained to perform human-like calculations. By 1911, the psychologist Oskar Pfungst had published the book “Clever Hans: A Contribution to Experimental Animal and Human Psychology” that discredited the achievements.
Through meticulous observation and a systematic scientific inquiry — with data, tables and plots to boot — Pfungst demonstrated that the questioner, more often than not Herr von Osten, was in fact integral to the apparent ruse.
“It seems ridiculous that this should never have been noticed before, but it is easily understood, for as soon as the questioner gave the problem he bent forward — be it ever so slightly — in order to observe the horse’s foot more closely… Hans would continue to tap until the questioner again resumed a completely erect posture.” (p. 52, Pfungst 1911)
Pfungst’s insights would appear to put to rest an animal’s — a horse’s — capacities to work with numbers. We learn through an obsessive account, presented in Clever Hans’ 274 pages, that the posture of Hans’ human questioner — the leaning down towards the right hoof — was the signal for the horse to stamp.
The lesson we might first take from this is that people were duped, that it was too much to expect anything other than a human to work with numbers. However, there is another way to make sense of this story; the animal studies scholar, Vincient Despret, invites quite a different reading in her vivid account of Hans and his abilities. In her always generous approach to investigating animal’s capacities, she turns our attention to what Hans was able to accomplish in his relations with a human interlocutor.
“Hans’s feats … testify to his capacity to be actively engaged in the game proposed, to give intense attention to minimal gestures expressing human desires, expectations and affects, and to respond to them in a remarkable way.” (p. 116, Despret 2004)
Are we then to take this away from Hans as well? That he was in fact attuned to the situation he found himself in in quite a remarkable way, able to read others’ bodies and show a range of capacities (only a glimpse of which I have captured above). And, what of Hans’ role in making these interactions between human questioner and horse in the first place? As Despret asks, should we dismiss the part he played in guiding humans to guide him, for in his training he also attuned them to what he was capable of. Thus, rather than a question of whether a body contains some innate capacity to work with numbers, we are invited to speculate on a mixture of bodies, agencies, and who affects and is affected. Rather than focus on what the limits are of a human or other-than-human actor, the invitation is to entertain the possibility of more.
Learning from Hans
Naturally, care is needed drawing simplistic parallels between Hans’ story and today’s enthrallment with ChatGPT. Without doubt, ChatGPT’s language capacities far exceed the mathematical skills that were attributed to Hans. And of course, ChatGPT is a massively distributed system that has the capacity to build on vast and still-growing sources of information. Still, I believe there are some ideas we might think with, ideas that speak — if you will — to the discussion about the personification of AI that Ben and Michael started and Pattie Maes, Susan Brennan, Ron Wakkary, and Mary Lou Maher have generously contributed to.
First, it seems fair to say that doubts of Hans’ abilities, and in particular Pfunkst’s well-meaning though dismissive accounts of the horse, might not have been the most productive way to approach our equestrian story. When we see abilities like mathematical skills and indeed being-clever policed in restrictive ways — perhaps making a case that such things are the dominion of humans alone — we close ourselves off from other possibilities. I believe this is what Michael is getting at when he speaks of “pondering possible futures.” As Pattie suggests, whether AIs such as ChatGPT are intelligent in the same ways as humans is “not the important question.” To approach the problem in such binary terms, we risk giving very little chance for more to happen, for different possible futures, for us to look for intelligences and capacities emerging somewhere else, between actors attuned to one another in unexpected ways.
Learning from Hans’ story, then, I would want to be open to new capacities — asking what more could come from using generative language tools like ChatGPT. Rather than defending who has rights to concepts like intelligence (and getting into some messy essentialising territory doing so), I’d want to start from the question: “might more be possible?” Might, for example, new or different ideas of actors’ collective capacities to act-intelligently-together come about through the combination of people and ChatGPT. This resonates with both Michael’s and Ron’s reflections on the generative relations we have and could have with technologies. The more interesting design problem here is not about regulating language use, but about how human and machine could be designed to attune to one another, how what affects and is affected might be better accommodated and amplified in the interactions [1]. The canonical chat interface seems far too limiting when set against this thinking.
In a roundabout way, these ideas of entangled actors (human, machine or otherwise) also give us a different way to approach Ben’s concern for the pronouns used by these language-based systems. I wholeheartedly agree with Mary Lou Maher when she writes “we need to bring more voices to the table.” I really like this analogy with its allusion to a more democratic approach to thinking about who has a stake in the design and use of AI. If we think with Hans and the story of counting, what we begin to see is that there are already a mixture of actors involved in making capacities possible — Herr von Osten, chalk boards, a scientific elite, psychologists, etc. And we should also acknowledge that there were undoubtedly a host of other players in Hans’ story that were not fully accounted for. What of those who cared for the horse or those onlookers that encouraged and applauded his and his interlocutors’ achievements? We might ask, then, who is missing from the tale?
Like our historic case, the same is of course true for ChatGPT. The automated language system relies on a diverse array of actors to do what it does, not least those who have produced the source content on sites like Wikipedia or the Kenyan workers who have labelled and annotated conversations to help ChatGPT avoid toxic and harmful language (Perrigo 2023). There is also, as Ron points out, the role of powerful corporate actors that must be subject to scrutiny (and will in all likelihood need regulating). To imagine ChatGPT as an “I” may then be less of a conceptual quagmire. It may simply be disingenuous to all those who have made such capacities possible in automated language generation. I would suggest, if we were to get behind a pronoun to be used by tools like ChatGPT, we should be pushing for “we” as opposed to “I”. And that we should, as Mary Lou suggests, be seeking to bring many more voices (of all kinds) to the table to ensure this ‘we’ makes visible and speaks for the many, not just the few.
So, to be thinking about how, precisely, the personification of AI should be regulated seems to me a limiting place to begin. So much seems possible with the rapid emergence of generative models and systems like ChatGPT. I’m keen to think with these possibilities, and, in capturing the spirit of the “we” — not “I” — I also want to recognise that there are many at the table contributing to these models and systems, and that there are many not present who should be. I am, ultimately, not in the business of enforcing pronouns, but I do want to open up the possibilities and I do want to be sure we remain open to who should have a say in the outcomes.
This is not to trivialise the serious issues we should be addressing with the rapid emergence of generative AI. I am steadfast in my belief that we need to be extremely cautious about attributing too much to AI technologies [2] and that we must not underestimate the risks and harms that come with systems that tap the internet as a resource (see Naomi Klein’s Guardian piece for a recent commentary on this topic). As both Ron and Susan suggest, we need to better understand how we hold these systems to account and, as Pattie argues, know what effects come with personification. We must also be prepared for and be quick to counter hateful and malicious uses of language models (see, for example, Burgess 2023).
In making more (not less) of generative AI, my hope is to bring many more voices to the table and, in assembling these voices, ensure we remain continuously attentive to where the risks are and where, exactly, we allow ourselves to be open to the possibilities.
Footnotes
[1] This thread of argument builds on one I began in “What Lines, Rats, and Sheep Can Tell Us” again using Despret’s generous reading of animals’ affects and agencies.
[2] I’m mindful, for example, of important work like Emily Bender’s (Weil 2023) that demands we think carefully about confusions between language use and language understanding.
References
- Matt Burgess (13 April 2023). The Hacking of ChatGPT Is Just Getting Started. Wired. https://www.wired.com/story/chatgpt-jailbreak-generative-ai-hacking/
- Vinciane Despret (2004). The Body We Care for: Figures of Anthropo-zoo-genesis. Body & Society, 10(2–3), pp.111–134. https://journals.sagepub.com/doi/pdf/10.1177/1357034x04042938
- Edward Heyn (4 September 1904). Berlin’s Wonderful Horse. New York Times. https://www.nytimes.com/1904/09/04/archives/berlins-wonderful-horse-he-can-do-almost-everything-but-talk-how-he.html
- Naomi Klein (8 May 2023). AI machines aren’t ‘hallucinating’. But their makers are. Guardian. https://www.theguardian.com/commentisfree/2023/may/08/ai-machines-hallucinating-naomi-klein
- Oskar Pfungst (1911). Clever Hans: A Contribution to Experimental Animal and Human Psychology. New York. Henry Holt. https://archive.org/details/b28061251/page/n8/mode/2up
- Billy Perrigo (January 18 2023). OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic, Times Magazine. https://time.com/6247678/openai-chatgpt-kenya-workers/
- Alex Taylor (2017). What Lines, Rats, and Sheep Can Tell Us, Design Issues, 3(3), p. 25–36. https://doi.org/10.1162/DESI_a_00449
- Elizabeth Weil (1 March 2023). You Are Not a Parrot. New York Magazine. https://nymag.com/intelligencer/article/ai-artificial-intelligence-chatbots-emily-m-bender.html
