- The paper critiques ML’s foundational assumptions, arguing that data, randomness, and evaluation methods are inherently rhetorical practices.
- It reveals how benchmarking and performance metrics serve as ritualistic validations that obscure underlying epistemic and ethical issues.
- It advocates for transparency and intelligence augmentation, urging a shift from manipulative practices to user-centric, reversible ML systems.
The Rhetorical Foundations and Implications of Machine Learning
Introduction
Robert C. Williamson's "The Rhetoric of Machine Learning" (2604.06754) presents an incisive critique of the epistemic and practical assumptions underpinning modern ML systems. Drawing from philosophical, rhetorical, and socio-technical perspectives, the paper asserts that ML is not merely an objective or neutral technology for modeling the world, but an inherently rhetorical enterprise. Williamson meticulously dissects the "styles of reasoning," foundational primitives, and business models prevalent in ML practice, foregrounding the implications for autonomy, manipulation, and information flows.
Rhetoric as the Core of Machine Learning
Williamson situates ML within the broader intellectual tradition of rhetoric—the science of persuasion—noting parallels with scientific, statistical, and economic reasoning. ML systems, he argues, are cognitive technologies delegated to persuade users through outputs such as predictions, forecasts, and explanations. The prevailing narrative in ML—the denial of rhetoric or "anti-rhetoric"—claims objectivity and impartiality. Williamson forcefully critiques this stance, highlighting that choices about data, modeling, and deployment are deeply rhetorical, embedded in argumentation structures (e.g., enthymemes) that tacitly presuppose warrants and background knowledge.
Foundational Primitives: Data, Randomness, Aggregation, and Probability
The paper emphasizes the interface between ML and the world, disassembling foundational concepts:
- Data as Fact: ML treats data as axiomatic, seldom interrogating provenance or process, echoing the statistical tradition. Williamson challenges the notion of "intrinsic structure" in data, asserting that facts are constructed and chosen not to be questioned; demanding chains of reference is essential for scientific rigor.
- Randomness and Independence: ML's reliance on the assumption that data is independently drawn from a probability distribution is critiqued as a rhetorical convenience. Williamson makes the bold claim—contradicted by mainstream ML practitioners—that there is no good reason to presume the existence of a data-generating distribution, especially in social contexts. He links randomness to fairness, demonstrating the ethical and statistical ramifications of conflating independence with protected attributes [43].
- Counting and Categories: The actuarial stance of ML—viewing individuals as aggregates—leads to significant epistemic and ethical problems, including the ecological fallacy and inadequacies in reasoning from aggregate statistics to individual outcomes.
Williamson interrogates the rhetoric surrounding ML models as representations of the world. He contends that models are better conceived as "useful fictions" rather than objective mirrors of reality, noting substantial literature on the limits and pitfalls of representations. ML's claim to "extract information from data" is analytically linked to information theory; training objectives like perplexity are fundamentally compression metrics. Despite industry claims to the contrary, empirical work demonstrates LLMs retain substantial training data information, evidenced by successful extraction attacks [58] and equivalence between information and expected risk [55].
Method and Benchmarking: Rituals of Verification
The obsession with method—valorization of algorithmic benchmarking, SOTA leaderboards, and scalar performance metrics—is identified as a core rhetorical device. Williamson traces these practices to Ramist traditions that commodify knowledge and repress agonistic dialogue, reinforcing disciplinary boundaries and self-justifying styles of reasoning. He warns of the epistemic closure inherent in benchmarking culture, which renders alternative evaluative frameworks invisible or illegitimate. Recognized mitigation efforts tend to reinforce rather than transcend benchmarks, leaving underlying issues of representativeness and data provenance unresolved.
ML as Black Box: Documentation and Process
Contrary to the common complaint of algorithms as "black boxes," Williamson advocates for mature, reversible black boxes accompanied by comprehensive documentation, paralleling established engineering practices. He urges re-framing data as process—emphasizing provenance and warrant—rather than as static facts. Current ML practices rarely provide such transparency, with the focus on performance crowding out documentation and systematic data governance.
Autonomy, Manipulation, and Machine Learning Business Models
A central section addresses the distinction between persuasion and manipulation, with manipulation as "manipulation-as-a-service" (MAAS)—a prevailing ML-powered business model where targeted information flows alter individual behaviors without conscious awareness. The paper details how statistical modeling enables insidious manipulation, with privacy-preserving targeting that is nonetheless highly effective. Williamson discusses multiple vectors for disrupting these information flows, such as noise injection and ad-blockers, and speculates on future regulatory and technological interventions.
LLMs and Statistical Knowledge
Williamson offers a rigorous critique of LLMs, affirming that these models are statistical rather than epistemic repositories of knowledge. LLMs generate outputs by sampling from posterior distributions, lacking warrants for their claims and serving as complex "autocomplete" engines. The paper references empirical and philosophical literature positioning LLMs as cliché generators and "bullshit generators" [90, 91], warning against taking statistical summaries for substantive knowledge.
Alternatives, Conviviality, and Intelligence Augmentation
Drawing inspiration from historical perspectives on technology, Williamson suggests an alternative trajectory for ML: intelligence augmentation (IA) rather than automation. Tools for conviviality should facilitate human flourishing, promote autonomy, and enable richer conversation, rather than perform manipulative or persuasive tasks by default. He references practical and theoretical underpinnings for IA, emphasizing user empowerment and autonomy-preserving design.
Implications and Future Directions
The rhetorical critique presented has extensive implications for theoretical and practical AI:
- Epistemics: ML disciplines must extend beyond anti-rhetoric, interrogating foundational assumptions about data, randomness, categories, and aggregate/individual reasoning.
- Ethics and Autonomy: The conflation of persuasion and manipulation, and the business models built on these flows, demand new frameworks for personal and societal autonomy.
- Benchmarking and Evaluation: The discipline should develop evaluative practices that transcend narrow benchmarking, integrating provenance, context, and process.
- Technology and Design: ML systems should facilitate intelligence augmentation, supporting user autonomy, robust documentation, and reversible processes.
Williamson's analysis predicts increased scrutiny of ML interface design, expanded work on data provenance, and regulatory approaches to manipulation-as-a-service. The theoretical landscape may shift toward dialogic and participative models of ML interaction, and practical deployments may prioritize information hygiene and autonomy rather than optimization for profit.
Conclusion
"The Rhetoric of Machine Learning" presents a rigorous, multi-layered critique of the epistemic, rhetorical, and socio-technical foundations of ML systems. Williamson demands that ML practitioners and researchers revisit assumptions about objectivity, data, randomness, and evaluation, recognize the inherent rhetorical nature of ML, and reconsider ultimate ends vis-à-vis autonomy, manipulation, and augmentation. The piece serves as a foundation for future research and debate at the intersection of philosophy, rhetoric, and machine learning, advocating for convivial, autonomy-supporting systems that augment rather than replace human intelligence.
