15 Responsible AI and Learning to Language

Sense-Ability

Learning is often described as a response to sensory experiences, and is linked to the development of a sense of responsibility around the consequences of action. Behaviorism ultimately describes learning as a series of adjustments guided by pain avoidance and reward. Radical constructivism describes how an organism responds when their senses contradict their previous understandings—assimilating new experiences into previous schema, or accommodating the schema. Social constructivism expands this approach, incorporating the experience of social structures in addition to sensory experience in the construction of knowledge. Enactivism, another offshoot of constructionist theories of learning, says that “cognition depends upon the kinds of experience that come from having a body with various sensorimotor capacities, and second, that these individual sensorimotor capacities are themselves embedded in a more encompassing biological, psychological, and cultural context” (Varela et al. 2016, 173).

These are all grounded theories of learning, as they build on sensory experience—experiences of pain and pleasure, but also observations and interactions that both challenge and affirm one’s understanding. This widespread approach to learning theory shares assumptions with cybernetics and information theory, which created a dominant cultural paradigm regarding emergent self-organizing systems that evolve or change based on cycles of response (e.g., Wiener 1948). These theories remind us that our human conceptions of learning are built around the centrality of sensing the world, and that the capacity to adapt and respond to sensory environments is key to learning.

It matters how we sense the world. We feel, hear, taste, smell, see, breathe, and digest things in our environments, and these experiences form the basis for our knowledge. Von Uexküll (1934 [2010]) used the word umwelt to describe an organism’s unique way of experiencing the world. Whales have a different umwelt than humans, so it is hard for humans to imagine what life would be like for a whale, harder yet to understand and communicate meaningfully with other organisms—algae, whales, humans, for example. But we assume that these nonhuman life forms achieve grounded knowledge through sensory experiences, and that their relationality, within their ecologies, entails sign-making and communication, a kind of grounded language. The umwelts of neural network AI would be very different from sense-based unwelts of organic beings (Hayles 2023). What could a neural network’s umwelt be? How can they have an umwelt if their material base is not organic, if they are not an organism? And without a relational world of sensory encounters, how could they possibly be held accountable for their language acts?

As an educator, I have an interest in developing students’ intelligence—their knowledge, their understanding, their decision making—but I cannot assess it directly. I can only assess what I perceive as action. I test students by giving them stimuli (language prompts usually) and assess their responses to these prompts. I test the way students respond, not their intelligence. Their abilities to respond to the stimuli are crucial in any process of learning, whether they respond to stimuli on tests or to interaction in dialogue—a feedback loop of responses to responses. Their experiences of pleasure, joy, dread, pain, triumph, fear, embarrassment and other feelings, all of them connected to sensual stimuli, play a significant role in human learning.

Response-Ability

Humans are, ideally, held to account for their words, and this accountability is linked to their embodiment. There are humans who do not bear responsibility for their language acts, or who may not think of the material consequences of their words. Children are not always held fully accountable for their words. Is an LLM like a child, permitted to speak without accountability? Are these machines taken as only learning the relationship between language and sense, and thus temporarily forgiven their indiscretions? Other humans insulate themselves from accountability; the relative wealthy among us often protect themselves with lawyers and layered bureaucracies that separate the investors from their exercises of power. Are LLMs like adults whose language production savvily positions others to take accountability for their language acts?

Responsibility means making considered choices and making oneself accountable for those choices. Posthumanisms (Braidotti 2017) and vibrant materialisms (Bennett 2009) remind us that the distinctions between nature and culture, human and machine, real and artificial, are rather fragile and provisional binaries that both help and hinder attempts to make sense of far more complex differences. Thinking with Donna Haraway (1985) about how to be “response-able” with different kinds of beings, we might comprehend ourselves as cyborgs: Perhaps humans are the sensory organs of our current “artificial” neural networks. The cyborg—a portmanteau word of cybernetics and organism—seems fairly appropriate. But which humans are entangled in this particular cyborg? How do the consequences of interactions with neural network AI circulate differentially across the cyborgian assemblage? There are indeed dangerous consequences of AI-generated texts and images, but it seems that marginalized humans are likely to bear the more destructive consequences. How does an AI network consider (in its training/development) the pains, pleasures, and sensory experiences of all humans, keeping in mind that categories of the human are often contested (Butler 2009; Jackson 2020).

Let’s imagine that generative AIs may be like humans as they “learn” and “develop” through responsive encounters with texts, growing their circles of relations. This response-ability to a growing relational ontology is an important feature of new materialist theory (see Haraway 2012). With increased relationality, one hopes an ethics of care and inclusion emerges across such AI, a response-ability for a heterogeneous collective that includes radical difference. But all the evidence points in the other direction; as AI enters the scene, “humans” seem increasingly confined to their online affinity bubbles, increasingly unable to even tolerate difference. Are we stretched too far, or not enough? How far can relations extend? Many of my relations are now mediated by written text: by email, and new mediums of language. I interact with people I have never met “in the flesh.” And yet I answer this question with language I learned from Indigenous knowledge keepers (elders); I live and work on land stolen from the Wolastokiyik people and I appreciate the wisdom of the local communities that have been living (grounded) with this land for millennia. Often, knowledge keepers will refer to their responsibilities and connections with “everyone, the four-legged ones, the ones that fly, the ones that swim, the trees, the rocks.” And they often close whatever they say with the phrase “all our relations,” which serves as a reminder that our circle of relationship and responsibility goes far and deep. Indeed, this is a cosmic relational ontology—an infinitely far and deep grounding of meaning and language action that seems out of reach to neural network AI and finite machines.

In drawing on this historical consciousness of past colonial violence, I choose language that entails relationships grounded in history and respect for place. And yet generative AI, perhaps computing more generally, is excessively dependent on costly energy and lithium and other extractive mining practices (Crawford 2021). In what ways are we accountable for the climate crisis as we engage with these resource-hungry models? We cannot forget our own response-ability for what is destroyed in the pursuit of new digital technology. When we scrape, puncture, cut, pound, and redistribute the land and other living beings around us with the aim of making a controlled world that fits some kind of ideal, we engage an earth that eventually strikes back. These climate strike-backs are wreaking havoc on our feeble and misguided attempts to conquer. Devaluations of place and perspective are another potential cost of distributing language interaction among ungrounded entities.

Words Matter

As a researcher, I think of the implications for how we attribute value to words manufactured by software. Within universities there is worry about plagiarism, and yet borrowed and commuted words are precisely what learning is all about. Does it matter where someone gets their words? Language is always mediated, borrowed, derivative. Bakhtin (1975/1981) described how all language comprises words, phrases and grammar borrowed from others. Barthes’s (1977) essay, The Death of the Author, showed how the idea of words belonging to an individual person is a recent development in history. The commodification of words connects with the excessive capitalist monetization of things that should belong to the collective. In traditional/Indigenous cultures, words and stories belong to communities, not to individuals.

We need to learn how to manage the language activity distributed across our circles of relation, including the more-than-human: agency is not one person’s decisive action alone, agency is an assemblage—a processual assembling of person with things, cultures, the land, and other creatures. As de Freitas & Sinclair (2014) showed, sensory embodiment is not an entirely individuated experience, and thus the question of grounding is far more complex than we know.

With the emergence of generative AI, there is the potential for deepening isolation as illustrated with the operating system “Samantha” in the film Her. The companionship performed by the gendered voice of Scarlett Johansson (another artifice of grounding) offered soothing words and intelligent conversation but not the range of sensory interaction to which humans are accustomed. Despite this cautionary tale, I would not want to discount the possibilities for generative AI to bring further richness to interaction in my circles of relation (Hayles, 2017).

This meditation on response-ability in language interaction identifies a need to evolve our standards for judging the words we encounter, whether we hear them or read them, as part of a transindividual experience. If we accept (or cannot detect) borrowed words from others, then how do we judge someone’s use of words? I suggest that originality (and independence from AI) is less important than that the speaker uses language in an accountable way, conscious of how words are part of a chain of responsibility. My words may sense-ably reflect my sensory experiences, or build on the sensing of others. My words can be aspirational—like the hyphenated word response-ability. I feel the weight of my words. My words can set a standard for others to hold me to account. Let us consider the impact of judging each other’s communication acts in terms of responsibility. What are our responsibilities in our language action? And how might “ownership” of words be transformed into responsibility for what is said? This would surely entail developing a language-learning AI that better addressed sense-ability and response-ability.

References