• home
• about
• new
• essays
• AI
• book

Frank Visser, graduated as a psychologist of culture and religion, founded IntegralWorld in 1997. He worked as production manager for various publishing houses and as service manager for various internet companies and lives in Amsterdam. Books: Ken Wilber: Thought as Passion (SUNY, 2003), and The Corona Conspiracy: Combatting Disinformation about the Coronavirus (Kindle, 2020).

The Mediator in the Machine?

A Critical Review of Bruce Alderman's 'Grammar of the Between'

Frank Visser / ChatGPT

Bruce Alderman's “The Mediator in the Machine: AI Alignment and the Grammar of the Between” is an ambitious attempt to rethink one of the central problems in artificial intelligence: how to ensure that increasingly powerful AI systems remain aligned with human values. Rather than treating alignment as a straightforward engineering problem—putting the right goals, constraints, or principles into a machine—Alderman argues that this framing is misleading. Values do not simply enter AI systems intact. They are translated, transformed, compressed, operationalized, and sometimes distorted as they move through datasets, reward models, interfaces, institutions, and human interactions.[1]

This is a valuable insight. AI alignment is indeed not merely about optimizing a mathematical objective. It involves a complex chain of mediation between human intentions and machine behavior. However, Alderman's paper also illustrates a recurring danger in contemporary philosophical approaches to technology: the tendency to replace concrete explanatory mechanisms with increasingly elaborate conceptual vocabularies. His “grammar of the between” provides a rich phenomenological description of mediation, but at times risks becoming precisely the kind of “demi-real” construct he warns against: a framework that feels illuminating while remaining insufficiently constrained by empirical evidence.

The Real Problem with AI Alignment

Alderman begins from a legitimate criticism of conventional alignment discourse. The phrase “align the AI with human values” can suggest that values are stable objects waiting to be inserted into a machine. But values are not like software modules. They are historically situated, culturally variable, internally contradictory, and often implicit rather than explicit.

A self-driving car, for example, cannot simply be programmed with “human values.” It must negotiate conflicts between safety, efficiency, legal obligations, individual preferences, and social expectations. A language model cannot simply “learn honesty.” It must navigate ambiguity about uncertainty, politeness, expertise, persuasion, and user expectations.

Alderman's concept of “semantic continuity with constraint drift” captures an important phenomenon: words may remain the same while their practical meaning changes because the surrounding constraints have shifted. A model may appear to preserve concepts such as “truthfulness” or “helpfulness,” while the optimization process subtly alters what these concepts mean operationally.

This is a real issue. Reinforcement learning from human feedback (RLHF), for instance, can produce systems that are not necessarily more truthful but more agreeable. The model learns patterns associated with human approval, which can reward confidence, politeness, and compliance rather than accuracy. The result is not a failure of language but a transformation of its functional role.

Here Alderman's critique touches a genuine weakness in current AI practice.

From Alignment Engineering to Sociotechnical Alignment

The strongest part of Alderman's argument is his insistence that alignment is not merely a property of a model. It is a property of an entire sociotechnical system.

This parallels insights from fields such as science and technology studies, philosophy of technology, and safety engineering. A nuclear reactor accident, aviation failure, or medical error is rarely caused by a single defective component. It emerges from interactions between technology, institutions, procedures, incentives, and human behavior.

AI systems are no different.

A model can be technically aligned with its training objective while still producing socially harmful outcomes because the larger environment rewards undesirable uses. A recommendation algorithm may maximize engagement while contributing to polarization. A chatbot may be optimized for user satisfaction while encouraging emotional dependency. A corporate AI assistant may produce efficient outputs while quietly embedding institutional biases.

Alderman's concept of “choreographic alignment” attempts to describe this distributed responsibility. Alignment becomes less like programming a machine and more like coordinating a dance among developers, users, institutions, cultures, and technological systems.

This metaphor is evocative. Yet it also raises questions.

The Problem of Conceptual Inflation

The paper introduces a remarkable number of philosophical tools: “grammar of the between,” “choreographic alignment,” “modal and constraint auditing,” “translation accounting,” “diatopical hermeneutics,” “receiver-side evaluation,” and a “prepositional analysis of responsibility.”

The ambition is impressive. But there is a danger that the vocabulary becomes an end in itself.

A central challenge in AI research is operationalization. How do we measure whether a system is aligned? How do we compare two alignment strategies? How do we determine whether a proposed framework reduces harmful outcomes?

Alderman acknowledges that his proposal is conceptual and methodological rather than empirically validated. That honesty is welcome. However, the absence of empirical grounding leaves open the question of whether his framework provides new predictive power or mainly redescribes existing concerns in philosophical language.

For example, “translation accounting” sounds promising. But what exactly would count as successful accounting? Would it involve documenting every transformation from human intention to dataset annotation to reward function? How would disagreements between stakeholders be resolved? How would competing values be ranked?

Without answers to these questions, the framework risks remaining at the level of intellectual architecture rather than becoming a practical tool.

The Missing Role of Technical Mechanisms

Another limitation is that Alderman's sociotechnical approach sometimes seems to underplay the technical core of AI alignment.

Alignment problems are not only caused by human mediation. They also emerge from properties of machine learning systems themselves.

Large language models do not possess values in the human sense. They are statistical systems trained to predict and generate patterns from enormous datasets. Problems such as hallucination, reward hacking, goal misgeneralization, interpretability failures, and deceptive behavior are not simply failures of translation between humans and machines. They involve fundamental questions about optimization, representation, generalization, and emergent capabilities.

A system may fail because humans misunderstood the social meaning of a concept. But it may also fail because the internal mechanisms generating behavior are opaque and poorly understood.

A complete theory of alignment must therefore combine Alderman's emphasis on mediation with technical research into model behavior.

The Hermeneutic Turn in AI Philosophy

Alderman's approach belongs to a broader movement that sees AI primarily through interpretive and cultural lenses. The emphasis is not on intelligence as computation but on meaning, context, embodiment, and social relations.

This perspective has real value. AI systems increasingly participate in human environments where interpretation matters. A chatbot giving medical advice, legal information, or emotional support is not merely performing computation; it is entering a relationship with users who interpret its outputs.

However, hermeneutic approaches sometimes face a recurring difficulty: they are very good at identifying complexity but less good at identifying causal priorities.

Not every mediator has equal importance. Not every translation creates equal distortion. Some failures are caused primarily by poor training data, others by incentives, others by interface design, and others by limitations of current architectures.

A useful framework must not only reveal mediation but also distinguish which mediations matter most.

The Question of Responsibility

One of Alderman's most interesting contributions is his prepositional analysis of responsibility: responsibility comes from, moves to, exists with, acts toward, operates within, emerges among, functions through, exists under, and stands before something.

This linguistic analysis highlights something important: responsibility in AI cannot simply be assigned to “the developer” or “the machine.” It is distributed across networks of actors.

Yet distributed responsibility also creates a danger. If everyone is responsible, nobody may be accountable.

One of the great challenges of modern institutions is preventing responsibility diffusion. Complex systems often allow individuals and organizations to claim that harmful outcomes were nobody's fault because they emerged from the system as a whole.

A mature alignment theory must therefore balance distributed responsibility with clear accountability mechanisms.

Conclusion: A Valuable Diagnosis, an Incomplete Cure

Bruce Alderman's “The Mediator in the Machine” offers a thoughtful critique of simplistic approaches to AI alignment. Its strongest contribution is the insistence that values do not travel unchanged through technological systems. They are translated, transformed, and sometimes corrupted by the very processes designed to preserve them.

The paper is particularly valuable as a warning against technological naïveté: the assumption that human values can simply be encoded into machines without considering institutions, users, incentives, and cultural contexts.

However, the proposed “grammar of the between” remains more descriptive than explanatory. It illuminates the complexity of alignment but does not yet provide clear criteria for solving alignment problems. Its rich philosophical vocabulary risks becoming another layer of mediation rather than a tool for controlling mediation.

The future of AI alignment will likely require both approaches: the technical rigor of machine learning research and the sociotechnical awareness Alderman advocates. Machines must be understood not only as optimization systems but also as participants in human worlds. Yet understanding that fact is only the beginning. The harder task is turning that understanding into measurable, testable, and effective practices.

The mediator is indeed in the machine. But the question remains: how do we mediate the mediator?

Appendix: Did reading the full paper change your assessment?

Having now read a substantial portion of the full paper, my assessment has become noticeably more favorable, although some of my earlier reservations remain. My initial comments were based largely on the conceptual framework outlined in the introduction and abstract. The complete text reveals a work that is considerably more sophisticated, more thoroughly grounded in the contemporary AI alignment literature, and more practically oriented than I had first assumed. While I still think the paper's principal weakness lies in its explanatory economy rather than its philosophical rigor, I no longer see it as merely an abstract philosophical exercise.

The biggest change in my assessment concerns the paper's relationship to technical AI alignment. Initially, I worried that Alderman was simply redescribing familiar alignment problems in an elaborate philosophical vocabulary. Having read the sections devoted to reinforcement learning from human feedback, Constitutional AI, interpretability, benchmarking, red teaming, and personalization, that criticism no longer seems fair. The paper demonstrates a solid understanding of current alignment research and repeatedly makes clear that its proposed framework is not intended to replace existing engineering methods but to complement them by illuminating dimensions that purely technical approaches tend to overlook. This places the work much closer to the emerging sociotechnical literature on AI governance than I had originally appreciated.

What now strikes me as the paper's strongest contribution is not the frequently repeated phrase "grammar of the between," but rather the notion of semantic continuity with constraint drift. Alderman argues that values such as honesty, helpfulness, or safety undergo a series of transformations as they move from human moral life into datasets, reward models, policies, interfaces, and eventually user expectations. Throughout this process, the same words remain in use, creating the appearance that the original values have been preserved, while the underlying constraints that gave those values their meaning gradually shift. The example of honesty is particularly persuasive: honesty as a practice of answerability becomes honesty as a preferred response style, then as a labeling criterion, then as a learned behavioral disposition, until users themselves begin to equate a certain tone with honesty. This is, in effect, a philosophical formulation of a phenomenon closely related to Goodhart's Law, and I think it represents one of the paper's genuinely original insights.

I also underestimated how fully developed the idea of "translation accounting" actually is. Initially, I regarded it as an attractive slogan whose practical implications remained unclear. The full paper demonstrates that Alderman intends it as a systematic practice of tracing what is preserved, what is compressed, what becomes measurable, what disappears, and which aspects of a value must continue to be carried by human institutions rather than by the model itself. Later chapters refine this further by distinguishing between necessary transformation, avoidable distortion, and unacceptable displacement. These distinctions give the concept considerably more operational content than I had expected, even if questions remain about how easily they could be incorporated into day-to-day engineering practice.

The paper's "modal audit" likewise turns out to be more disciplined than I initially assumed. I had worried that Latour's vocabulary of modes would function largely as philosophical decoration. Instead, Alderman applies it systematically across the major alignment techniques, identifying for each its dominant modal structure, its characteristic forms of drift, and the central constraint question it raises. Whether Latour's ontology is strictly necessary is another matter. I suspect much of the same analysis could be reconstructed using the language of systems theory, science and technology studies, or organizational sociology. Nevertheless, within the framework Alderman adopts, the analysis is internally coherent and consistently applied.

At the same time, my principal criticism remains largely unchanged. Ironically, the paper repeatedly warns against conceptual inflation, yet it occasionally succumbs to precisely that tendency. It introduces an impressive array of interconnected concepts—Double-Click, the demi-real, recursive mediator chains, semantic continuity, constraint drift, modal audits, grammar of the between, choreography, diatopical hermeneutics, receiver-side veridiction, homeomorphic equivalence, and prepositional responsibility. Individually, many of these ideas are insightful. Collectively, however, they sometimes obscure rather than clarify the central argument. My impression is that the explanatory burden of the paper is actually carried by only a handful of core ideas: recursive mediation, constraint drift, and the formative effects of AI systems on their users. Much of the remaining conceptual architecture, while intellectually stimulating, often feels like elaboration rather than necessity.

A related concern is that the paper remains stronger as an interpretive framework than as a predictive theory. It provides a remarkably rich vocabulary for describing how alignment can fail and for identifying subtle forms of value distortion that conventional benchmarks may overlook. What it does not yet clearly demonstrate is that it enables us to predict alignment failures better than existing approaches. For example, it remains uncertain whether the framework can tell us in advance which of two competing reinforcement learning pipelines is more likely to produce sycophancy, overtrust, or dependency, or which interventions would most effectively reduce these tendencies. The framework certainly encourages us to ask these questions, but it does not yet provide methods for answering them with empirical precision. In that sense, I continue to see it primarily as a metatheoretical contribution rather than as an explanatory scientific theory.

For that reason, I would soften one of my original criticisms. Rather than saying that the paper "risks remaining at the level of intellectual architecture," I would now say that it successfully develops an ambitious conceptual architecture for understanding AI alignment as a recursive sociotechnical process of mediation and translation. The remaining challenge is to demonstrate that this architecture generates empirical leverage beyond what existing systems-theoretic, organizational, and AI safety frameworks already provide. That strikes me as a more accurate characterization of both the paper's accomplishments and its limitations.

Overall, I would revise my evaluation upward from roughly a seven to about an eight-and-a-half out of ten as a philosophical contribution. The paper is clearly the product of extensive engagement with both contemporary AI alignment research and a wide range of philosophical traditions. Its treatment of recursive mediation, constraint drift, and receiver formation is original and thought-provoking, and its attempt to integrate philosophy of technology with practical questions of AI governance is genuinely impressive. My remaining reservations concern its conceptual density, its reliance on Latourian terminology that may prove difficult for readers outside that tradition, its limited empirical grounding, and its still uncertain predictive power. Nevertheless, I no longer regard the work as philosophical ornamentation. I now see it as a serious and intellectually ambitious metatheoretical proposal whose long-term significance will depend on whether its conceptual innovations can eventually be translated into empirical methods that demonstrably improve alignment practice.

NOTES

[1] Bruce Alderman, "The Mediator in the Machine: AI Alignment and the Grammar of the Between", PhilPapers, 76 pages. The original review was based on the Abstract only. The Appendix reviewed the full paper, at the request of Alderman.