5 The Problem of Algorithmic Commonsense Learning

Codifying Common Sense

GOFAI approached common sense primarily as “the capacity to make inferences on a body of object-led information” (Davis 1990, 3). It focused on establishing “primitive and social formal rules that captured everyday knowledge” (Dreyfus 1992, ix) and could be encoded into rules-based algorithms to enable inference-making in logic-based systems (e.g., McCarthy 1959).

GOFAI’s quest to “codify common sense” through translating tacit human knowledge into machine-readable knowledge is perhaps best encapsulated by the Cyc project; launched by computer scientist Doug Lenat at the US-based Microelectronics and Computer Technology Corporation in 1984, Cyc’s initial methodology involved encoding in machine-readable terms 99 percent of a one-volume American desk encyclopedia, followed by all the commonsense “facts” (e.g., that an object can’t be in two places at once) that the encyclopedia’s creators “presumed the reader already knew” (Dreyfus 1992, 76). This design, the team hoped, would enable the system to infer further rules directly from ordinary language.

Cyc’s learning architecture foregrounded analogical reasoning. Operationalizing the computer scientist Marvin Minsky’s (1984) claim that humans “assimilate new information by finding similar things we already know about and recording the exceptions to that analogy,” the Cyc team sought to “swap the problem of ‘telling the system about x’ for the problem of ‘finding an already known x that’s similar to x’” (Lenat et al. 1985, 66). In other words, Cyc would navigate new situations by drawing analogical links to what it already “knows”—which would require that it be programmed with a substantial base of knowledge and an organized body of reasoning methods.

Through mobilizing higher-order logic, Cyc began to abstract, generalize, and learn from its experience. It could, for example, “infer ‘Garcia is wet’ from the statement ‘Garcia is finishing a marathon run’, by employing its algorithmic rules that running a marathon entails high exertion, that people sweat at high levels of exertion, and that when something sweats it is wet” (Copeland 2016, online). Framed as a forerunner to IBM’s Watson supercomputer, which in 2011 famously beat two reigning human champions on Jeopardy! (Havasi 2014), Cyc influenced the emergence of other AI commonsense reasoning projects including, most notably, Open Mind Common Sense, set up by Minsky, Push Singh, and Catherine Havasi at MIT Media Lab in 1999 (Havasi et al. 2014).

As time passed, however, logic-based AI projects like Cyc failed to achieve robust flexibility and intuitive sense-making—deficits linked not only to difficulties in scaling up from isolated “micro-worlds” (Davis 1990) but also to GOAFI’s account of “intelligence as a passive receiving of context-free facts into a structure of already stored data” (Dreyfus 1992, 34). As the philosopher of science Hubert Dreyfus argued from the 1970s, human learning entails broad embodied “know-how” shaped by contextual particularities, unfolding moods, and sensory-motor skills. A key problem with symbolic processing AI’s model of learning is that “one cannot substitute an extractable web of beliefs for the whole cloth of our concrete everyday experiences” (Dreyfus 1992, 54).

Common sense, as such, is never merely a cognitive operation—affect, sensation, and embodied habituation are vital to its dynamics, to how common sense shapes and is shaped by situated modes of attention, perception, and interaction (Keeling 2007). When CyCorp (Cyc’s corporate owner since 1995) claims that Cyc’s “human-like” cognitive skills mean that it “understands (represents fully) real world contextual nuance . . . like culture, emotions, time, space, beliefs and bias” (CyCorp n.d.), the significance of the words “cognitive” and “represents” is crucial. What Cyc offers is a logic-based account of cognitive elements of “emotion” amenable to symbolic representation. Whatever cannot be made legible in precise machine-readable instructions, codes, and categories lies outside the scope of the system.

To “codify common sense” may thus be precisely to elide its affective elements, as immanent visceral and sensorial processes are reframed as schematic cognitive ones. Yet as computer science’s struggles with formalizing tacit human knowledge emphasize, intuition may be powerful to learning precisely to the extent that it remains implicit, incoherent, and unarticulated. In turn, common sense’s resistance to explication is what makes “the practices it teaches significantly inaccessible” (Shotwell 2011, 13)—and therefore not easily (if ever) replicated by artificial means.

Abduction, Affect, and Discovery

In associating common sense with inference-making via programmed knowledge, GOFAI built computing systems capable of inductive and deductive reasoning, but not what the philosopher Charles Sanders Peirce (1958) called “abductive reasoning”—the generation and evaluation of explanatory hypotheses amid uncertainty. By contrast, in its capacity to go beyond what it has been taught to find “hidden patterns” in its data, machine learning is said to specialize in abduction. Similar, it might be said, to how Peirce describes abduction as hypothetical thinking with “maybes” (Fortes 2022, 2), generative AI constitutes an unfolding “speculative experiment” in which “data inputs and the algorithm mutually modify to optimize the output” (Amoore 2020, 48).

Yet even state-of-the-art “abductive” machine learning systems struggle with commonsense ways of knowing. As the computer scientist Gary Marcus notes, deep learning programs excel in “closed-end classification problems” with millions or billions of training examples to learn from, but do less well when data is limited, the test set differs significantly from the training examples, or the “space of examples is broad and filled with novelty” (Marcus 2018, 15). Problems that have less to do with categorization and more with everyday sense-making are less successfully addressed by statistical approximation and recursive learning from examples.

Key AI actors thus recommend a hybrid “solution” to AI’s persistent commonsense problem that integrates machine learning with aspects of GOFAI. Marcus (2018), for instance, calls for a “neuro-symbolic approach” that marries the perceptual classification of deep learning with the inference and abstraction of symbolic processing. In this vein, computer scientist Yejin Choi advocates a synthesis of the abductive capacities generative machine learning (such as those associated with ChatGPT) and language-based (as opposed to logic- or rules-based) formalisms—which, in her view, are “sufficiently expressive and robust enough to encompass the vast number of commonsense facts and rules about how the world works” (2022, 139–40). In other words, language-based formalisms would supplement the large-scale, speculative pattern recognition of generative AI with neuro-symbolic training sets that aim to simulate tacit human learning.

Yet critics argue that hybrid neuro-symbolic approaches still neglect what is fundamental to human common sense: its imbrication with and dependence on situated socioaffective contexts and relations (Suchman 2024). While computer scientists champion language-based formalisms claiming that “it is language, not logical forms through which humans acquire knowledge about the world” (Choi 2022, 140), this shifts the site of learning from “logic” to “language” without robust means of addressing the implicit and sensory elements of common sense that exceed language. What both first- and second-wave AI lack, from this perspective, is a genuine ability to learn affectively—a capacity, that is, for tacit, experimental, sensory-oriented learning in which particular implicit qualities may never become explicit.

The “learning” in machine learning refers to algorithmic recursion in which the outcomes of previous actions are taken as inputs for future action, which allows a given program to recognize new items not part of its original training data. While AI research has, since the 1990s, explored how affect figures in synthetic learning and sense making, common computational techniques such as sentiment analysis provide “statistical proxies for affective intensities [that can] displace, reference, meaning and comprehension” (Andrejevic 2013, 54). Unable to generate modalities of common sense that synthesize deep entanglements of cognitive, sensory-motor, and sociocultural processes, machine learning systems compensate by substituting probabilistic correlations for relational processes of affecting and being affected (Pedwell 2023).

Current AI architectures may not, from this perspective, be as “abductive” as they are commonly framed—particularly if we understand sensory-cognitive processes as central to abduction’s capacity for speculative hypothesis generation and the creation of “newness.” It is worth emphasizing, in this respect, that Peirce’s understanding of abduction retains an emphasis on embodied processes of learning premised on affective intensities and relations. Abduction, as such, is a multimodal model of discovery in which what is most important is not necessarily the data itself but rather subjects’ sensorial reactions: Abduction is initiated by “the feeling of puzzlement and ends at the satisfaction of knowing” (Fortes 2022, 5). As an intuitive (yet also logical) mode of inhabiting plausible possibilities, Peircean abduction thus explicitly departs from probability-based modes of inference, including machine learning systems that generate shifting commonsense “truths” as the outcome of “aggregate feeling tone[s]” (Andrejevic 2013, 46).

Truth, Ideology, and Analogical Reasoning

Common sense has been widely understood in AI research to involve physical, biological, psychological, or economic “truths” about “how the world works” (Lenat et al. 1985), which must be programmed into logic-based systems or immanently generated by training machine learning programs with relevant data sets. Yet as critical thinkers from Antonio Gramsci (1971) onward have argued, common sense is far from impartial or neatly extractable from hegemonic worldviews; the question of “how the world works” is political and ideological as much as ontological and physical.

Take, for instance, Cyc’s investment in analogical reasoning as a route to commonsense learning. In practical terms, once the Cyc team had encoded four hundred articles from their chosen encyclopedia into the system, research assistants were enlisted to enter “the final 99 percent of the knowledge base.” Each research assistant would “take an article, locate the already-represented similar article(s), and perform a machine-assisted ‘copy & edit’ procedure to produce a machine-understandable version of the new article.” An existing article on “Pewter” might, for instance, be copied and edited to populate a new entry on “Britannia-Metal” (Lenat et al. 1985, 77). Drawing analogies between types of metal may not seem particularly contentious, yet the ideological interests at stake in machine-enabled analogical reasoning slide into relief when we consider other founding terms within Cyc’s knowledge base—such as “Ronald Reagan,” “female animal,” “abortion,” “AIDS,” “homosexuality,” and “terrorist.”

While the early Cyc literature touches briefly on what kind of “light” practical training research assistants might require for their “copy & edit” task, there is scant discussion of what role the social positionalities, affective orientations, and ideological worldviews of these “knowledge enterers”—or indeed of the project’s creators, designers, and funders—might play in the system’s ongoing production of common sense.

MIT’s Open Mind Common Sense Project (OMCS) also employed analogical reasoning in its learning schema, though its use of machine learning techniques and internet-sourced data distinguished its approach from Cyc’s. OMCS combined a semantic network, ConceptNet, “built from a corpus of knowledge collected and rated by volunteers on the Internet,” with a reasoning engine, AnalogySpace, which “used factor analysis, to build a space representing the large-scale patterns in common sense knowledge” (Havasi et al. 2014, 24–25). In this distributed system, commonsense “truth” emerges, in part, as the shifting effect of popular ratings. From 2007, internet volunteers rated commonsense statements from OMCS using the set “generally true,” “sometimes true,” “not true,” “doesn’t make sense,” and “not true but amusing” (Havasi et al. 2014, 32)—data that was fed back into the system to refine and expand its knowledge base. Through OMCS’s mediating lens, then, the validity of commonsense claims become a recursive function of algorithmically calibrated “popularity.”

While Gramsci frames popular knowledge, values, and rules of thumb as crucial sites of social learning and political struggle, popularity within machine learning systems is the unstable product of popular ratings and statistically adjudicated matters of fit that remain largely hidden from public view. Generative AI, in turn, increasingly incorporates “counterfactual reasoning,” which computer scientists frame as “closely related to abductive reasoning” and thus necessary to engage a world “filled with previously unseen situations” (Choi 2022, 144). Yet if Peircean abduction thinks with “the possible” to generate “momentary truths” for the purpose of discovery (Fortes 2022, 2), the statistically adjudicated modes of common sense produced by machine learning systems may limit rather than expand experimental possibilities for social life. This is the case not only given how computational common sense embeds social normativity at the levels of logic, procedure, and data, but also because deep learning’s probabilistic mode of working with hypotheticals and inverse inference from a vast field of counterfactuals can proceduralize abduction—so that it tips more heavily toward regulation and away from creativity (de Freitas 2022).

To work effectively with “inconsistency, subjectivity, and generally noisy data,” OMCS, for instance, sought to make “rough conclusions” based on analogies, similarities, and tendencies rather than on an “assumption of absolute truth” (Havasi et al. 2014, 27). The system employed singular value decomposition, a technique in linear algebra, which compresses data by sharing information between items deemed similar to each other. This enabled “a method of commonsense inference called cumulative analogy” (2014, 27). If a general risk of analogical reasoning is that it flattens complex relations and elides affective, cultural, and sociopolitical particularities, cumulative analogy intensifies this risk because it seeks “to moderate elements of the cultural field that may present themselves as typical or outstanding, so that they can be led to make sense relative to other, more even-keeled examples” (Hallinan and Striphas 2016, 122)—dynamics indexing the wider historic links among statistics, machine learning, and the racialized, gendered, and classed constitution of “normalized standards of behaviour” (Amaro 2022, 24).

“Learning,” from this perspective, may reproduce a recursive loop of dominant cultural associations—or make probabilistic speculations based upon iterative biases and prejudices projected into the future. Human learning, of course, might be said to work similarly insofar as it involves analogical reasoning and other epistemic shortcuts, yet what is at stake as machine learning architectures become increasingly “environmental” is how probabilistic analogizing is operationalized on grand populist scales.

More-than-Human Futures

While AI research tracks back and forth between first-wave symbolic processing and second-wave deep learning techniques—both of which bely an enduring attachment to cognitivism—alternative frameworks for conceptualizing human-machine intelligence and sense-making remain dormant (Suchman 2024, 88). What would (or even could) meaningful (processual, multidimensional) affective learning and abductive experimentation—attuned to the complex interactions among common sense, computation, and social relations of power—look like in intelligent systems?

Postwar genealogies of AI offer glimpses of different routes to common sense with the potential to decenter cognitivism and foreground affective attunement—from approaches associated with “4E” (embodied, embedded, enactive, extended) cognition of the 1980s to affective robotics of the 1990s. What may be vital to the future of expansive commonsense learning and discovery in algorithmic conditions, however, is to avoid holding cognition and affect in rigid opposition. While normative AI imaginaries position cognitive “chess-playing intelligence” and affective “childlike learning” as oppositional, for instance, cognitive reasoning, as the feminist science studies scholar Elizabeth A. Wilson (2010) argues, “emerges out of circuits of physical, psychological, and aesthetic expertise” and is “more imbricated in networks of affectivity than has usually been supposed” (5). In Peircean mathematics, moreover, abduction entangles cognition, reason and affectivity—in a world in which “mathematical habits [are] embodied, situated, cultural, planetary and more than human” (de Freitas 2022, 2). Emergent research on “spiritual AI,” which harnesses affect as physical “energy” (wavelength, oscillation, vibration) (Garg 2024), may also refigure the commonsense problem anew.

Given that generative machine learning architectures give rise to distributed and relational capacities premised on human-algorithm entanglements, the imperative is not, in my view, to negate the possibility of synthetic common sense by reiterating that machines “lack feelings” or that they “are not made of the right kind of materials.” Rather, what is at stake now is how varied AI systems are (re)constituting both “the sensible” and “the sensable” in ways that return us to Aristotle’s ancient philosophy of common sense.

The composite forms of perception, knowledge, and learning enabled by algorithmic architectures are transforming what can be made perceptible to the senses, as well as the range of entities that participate in such cognitive-sensorial agencies. While “the human” cannot be excised from conversations concerning computational common sense, we do, as the media scholar Lisa Blackman contends, “need radically revised notions of body-world-consciousness” which are “compatible with twenty-first century media” (2019, 178). If, moreover, human experience is being remediated (experientially and affectively) in line with the affordances of algorithmic architectures (Pedwell 2019), it may be that we are learning to conform to computational categories, expectations, and common senses as much as (or more than) they are learning to capture and/or simulate us.

Within these emergent more-than-human media ecologies, what might it mean to inhabit common sense as a site of socioaffective and “political struggle” (Gramsci 1971)? How might dwelling experimentally within these unfinished histories of human-machine relations enable a different philosophy and politics of AI with more expansive abductive possibilities for (un)common sense, collaborative sensing, and affective learning?

Notes

1. 1. AI literatures refer to various “waves” of AI, with some studies claiming three or four distinct waves. For clarity, I refer to two key waves in AI research corresponding broadly to “logic-based” and “machine learning” systems, while acknowledging the diversity of techniques within these schematic approaches as well as the overlaps and interconnections between them.

2. 2. Key ideas in this memorandum are expanded upon in Pedwell (2024).

References