AI-Human Entanglement, Agency, and the Transformation of Governance
Introduction: The Stakes of Entanglement—Why This Framework, Why Now?
The advance of artificial intelligence is driving a quiet revolution—one that refashions not only how institutions operate, but also how meaning, authority, and agency are experienced in daily life. Unlike the dramatic disruptions that capture public attention, this transformation proceeds through subtle displacements: the gradual outsourcing of judgment to algorithmic systems, the erosion of spaces for deliberation, and the systematic replacement of human reason-giving with optimized outputs that simulate deliberation without providing its substance.
Artificial intelligence is less a wave of discrete technological tools and more a web of infrastructural conditions that quietly reshapes how we act, think, and value. Not only are institutions reconstituted around algorithmic mediation and optimization, but the very fibers of daily life—decision, meaning, and critique—are renegotiated inside this surrounding web. If we are to navigate this landscape well, we need clear concepts for the different ways agency is now distributed, transferred, and sometimes atrophied.
This essay offers a framework to diagnose the structural and lived consequences of AI deployment, drawing on philosophical traditions that we repurpose for contemporary challenges. Our analysis builds on Wilfrid Sellars' distinction between the "space of reasons" and the "space of causes," Jürgen Habermas's account of system colonization of the lifeworld, and John Dewey's insights into purposive agency through "dramatic rehearsal." To these established frameworks, we add our theoretical innovations, including the concept of a "space of pseudo-reasons" and expanded attention to how AI operates across multiple scales—from individual psychology through primary group dynamics to institutional transformation.
Our analysis critiques not "AI" in the abstract, but the institutional regime of AI deployment—the specific architectural, economic, and organizational arrangements that reward the displacement of deliberation in favor of efficiency. The danger we identify is not inevitable, but grows from incentives that privilege streamlined automation while masking agency transfer behind what we term the "user's illusion" of control.
As artificial intelligence systems become more sophisticated and pervasive—particularly with the emergence of "agentic AI" that can act autonomously across digital platforms—the stakes of this analysis intensify. We are not merely witnessing the automation of specific tasks, but the transformation of the fundamental conditions under which human agency operates. The question is not whether we will live with AI, but whether we can do so while preserving what is essentially human: the capacity for deliberation, reason-giving, and the creation of shared meaning through communicative action.
A Typology of Agency in Human-AI Systems
To understand these transformations, we must first distinguish between different types of agency that have emerged through the historical development of AI systems. Rather than treating "artificial intelligence" as a monolithic category, we propose a four-part typology that maps both onto technological capabilities and their chronological development:
1. Human Purposive Agency
At the heart of human life is a capacity for purposive agency—for deliberation, creative projection, and normative evaluation, what John Dewey called "dramatic rehearsal." Human persons are not mere bundles of impulse or products of optimization. We inhabit what Wilfrid Sellars termed the "space of reasons"—a domain where intentions, narratives, and justifications are forged and negotiated. We do this not as pure rational calculators, but as beings whose futures are shaped in part by imaginative rehearsal, memory, affect, and the lived negotiation of values.
When humans act, they typically do so for the sake of an end-in-view—surviving, competing, creating, connecting. Crucially, humans are generally aware of their purposes and can engage in what Dewey called "dramatic rehearsal"—the imaginative exploration of possible actions and their consequences before committing to a particular course.
This process involves the whole person, not just analytical cognition. When considering whether to change careers, individuals don't simply calculate costs and benefits. They imaginatively inhabit different possible futures, exploring how it might feel to do different kinds of work, how their relationships might change, what kinds of meaning they might find. This embodied, social exploration of possibilities is central to what makes human agency distinctively human rather than merely computational.
This is not a remote philosophical ideal but a functional prerequisite: without such capacities, human cooperation, normativity, and meaning would not be possible. Importantly, human agency operates simultaneously in what Sellars distinguished as the "space of reasons" and the "space of causes." Unlike Sellars' original formulation, we recognize that human actions are shaped by both deliberation and biological/social causation—hormones, emotions, social pressures all influence our "reasoned" choices. A decision to participate in a protest might be driven both by moral conviction and by adrenaline or social conformity. Human agency emerges from this dynamic interaction rather than from pure reasoning.
2. Direct Artificial Agency
By contrast, AI systems operate solely within what Sellars called the "space of causes." Their "decisions"—no matter how sophisticated—are outputs of causal optimization, devoid of normativity or reflective grasp of meaning. The earliest AI systems developed in the 21st century—autonomous vehicles, weapons systems, predictive policing algorithms—were designed to execute specific tasks with direct effects on the physical world. These systems operate exclusively in the space of causes through algorithmic processes, without understanding, intention, or genuine reason-giving. Their "decisions" result from mathematical optimization designed to achieve specified objectives within defined parameters.
A legal autonomous weapons system, for instance, can identify and engage targets based on programmed criteria, but it cannot engage in moral reasoning about whether such action is justified. It operates through chains of efficient causation—sensor data, pattern recognition, targeting algorithms —with no access to the space of reasons that would allow for ethical deliberation. For all the power of machine learning, current AI does not access, or even approximate, the space of reasons: its agency is direct, causal, and indifferent to meaning, justification, or value.
3. Indirect Artificial Agency: The Interpretive Turn
The emergence of large language models (LLMs) and sophisticated recommendation systems created a new form of artificial agency which operates indirectly through human interpretations of outputs (be they in the form of text, sound, image, suggestion, explanation et al.). Indirect Artificial Agency resides in the human act of interpreting any AI output—critically or superficially, reflectively or uncritically. Whether the LLM offers a "suggestion," a rationale, a diagnosis, a poetic completion, or a mundane prompt, it becomes agentically consequential only when a person reads, appropriates, and acts upon it—assigning it significance, credibility, or skepticism. The decisive moment is not at the point of output, but in the interpretive labor that follows.
[Footnote 1: Of course, there are cases where an AI output—like an autogenerated medical report or legal brief—receives no individual reading or reflection. In such situations, it might seem there is no interpretive act at all, and thus that the AI itself exerts “real and direct” agency akin to autonomous systems acting physically in the world. But in fact, the interpretive act is not missing; it has shifted from the individual to the institutional level. Institutional protocols and conventions treat absence of review as tacit approval. Unread outputs become consequential not by their mere production, but because the system has been designed to interpret “not read, not flagged” as “approved” or “fit for action.” The responsibility for outcomes thus lies not in the machine’s output alone, but in the organizational translation of inaction into permission—a transfer of interpretive agency to rules, defaults, and workflow artifacts. To mitigate the normalization of this unintended default, institutions must enact protocols that actively check and counteract negligence—ensuring that absence of review does not become tacit approval.]
When an AI assistant suggests a restaurant or writes a business report, it operates through what we term "indirect agency"—the system itself cannot directly book a table or send the report, but its outputs often guide human actions that do so. The human retains the final step of interpretation and implementation, but the quality and depth of that interpretation varies dramatically.
Agency, then, is not a property of the content; it is enacted in the loop of interpretation, judgment, and appropriation. "Labor-intensive" uses—recursive drafting, reflective analysis, repair of hallucinations— preserve and exercise purposive agency. These approaches treat AI outputs as raw material for further deliberation, maintaining human authority over meaning-making and decision-making processes.
Superficial, "labor-saving" uses, by contrast, short-circuit critical engagement, risking the atrophy and outsourcing of meaningful agency. The user does not simply become "passive"; they enter a new relation, where the locus and quality of agency shifts fundamentally. A student who prompts an AI to write an essay and submits it without careful reading represents an extreme case of labor-saving usage—taking themselves almost entirely "out of the loop" and treating AI as substitute rather than supplement.
4. Hybrid Artificial Agency: The Emergence of Infrastructural Entanglement
The newest development in AI systems—agentic AI platforms like advanced browsing agents— combines recommendation capabilities with direct execution powers. These systems can not only suggest actions but also carry them out: booking flights, making purchases, scheduling meetings, managing communications across platforms.
Almost all present-day AI deployment is hybrid: agency pulses and flows in dynamic assemblages spanning persons, machines, institutions, and platforms. This hybridization is not merely additive. It is infrastructural, akin to highways, power grids, or digital backbones: it enables, constrains, and distributes agency and value in ways that no longer map neatly onto the categories of "tool" or "user."
Hybrid systems represent a qualitative shift because they collapse the mediation step that previously allowed humans to maintain deliberative distance from AI outputs. When users can give permission for an AI system to "handle my travel planning" or "manage my social calendar," the system moves from offering suggestions to taking direct action in real-time and space, though still technically under human authorization.
We increasingly "navigate" these infrastructures rather than controlling them from without—adapting, responding, and contesting their effects from within. The question is not whether entanglement will happen, but how it will be structured, and to what ends.
This typology reveals a clear historical trajectory: from systems that act directly on the world without human mediation, to systems that influence human action through symbolic outputs requiring interpretation, to systems that combine both capabilities. Each type creates different patterns of agency displacement and requires different analytical approaches.
Methodological Note: Epistemic Humility and Live Hypotheses
Before proceeding to detailed analysis, we must acknowledge the limitations of current knowledge about human-AI interactions, particularly regarding the newest agentic systems. Much of our analysis of individual and small-group experiences with agentic AI represents live hypotheses rather than empirically established findings.
Agentic AI systems are so new that meaningful ethnographic data, longitudinal studies, and systematic surveys of user experiences simply do not yet exist. Our discussions of "interaction rituals," "frame shifts," and "responsibility attribution patterns" are theoretical constructs designed to guide empirical inquiry rather than settled conclusions about how these systems actually function in daily life.
This reflects our pragmatist commitment to ongoing, revisable, fallibilist inquiry. We offer these concepts as tools for organizing research and navigating emerging phenomena, with the explicit intention of updating and revising our framework as new information becomes available. The rapid pace of AI development requires this kind of theoretical scaffolding for empirical work, even as we remain humble about what we do and don't yet know.
Philosophical Foundations: Revised and Integrated
Sellars: The Space of Reasons, Causes, and the New "Pseudo-Reasons"
Wilfrid Sellars' classic distinction between the "manifest image" and the "scientific image" provides foundation for understanding agency displacement. In Sellars' framework, the manifest image represents the world as we experience it—where people act for reasons, deliberate about choices, and explain themselves in terms of intentions and purposes. The scientific image describes the world as science reveals it—where events, including human actions, are explained through physical, chemical, and biological causes.
Our contemporary adaptation recognizes that humans actually operate in both spaces simultaneously, in complex interdependence that Sellars did not fully anticipate. Human actions are shaped by both reasons (deliberation, values, intentions) and causes (emotions, hormones, social pressures). AI systems, by contrast, operate exclusively in the space of causes through algorithmic processes— without understanding, intention, or genuine reason-giving.
However, contemporary AI deployment has given rise to a third category: the "space of pseudo- reasons." This domain encompasses AI-generated outputs that simulate deliberative reasoning through natural language or structured explanations, but derive from causal optimization processes lacking intentionality or normative judgment.
When AI systems offer "recommendations" complete with justifications, present "smart suggestions" that appear tailored to individual preferences, or provide "explanations" for their outputs, they create the appearance of operating in the space of reasons while remaining firmly within the space of causes. This simulation is not accidental but engineered—contemporary AI systems are explicitly designed to mimic human-like reasoning and communication.
The space of pseudo-reasons becomes particularly significant when humans treat AI outputs as if they were backed by genuine deliberation. This phenomenon—the "user's illusion"—occurs when people interact with AI systems as if they were reason-giving agents capable of genuine understanding and judgment. The more convincing these simulations become, the more effectively they transfer agency from human deliberation to algorithmic optimization while maintaining the appearance of collaborative reasoning.
Habermas: System, Lifeworld, and Algorithmic Colonization
Jürgen Habermas's analysis of "system" and "lifeworld" provides crucial insight into how AI deployment transforms social life. The lifeworld represents the background of shared meanings, cultural knowledge, and communicative practices where people interact, deliberate, and create social norms through language and mutual understanding. The system encompasses formal organizations, markets, and bureaucracies governed by instrumental rationality—efficiency, control, and goal- oriented action.
Habermas warned that system logic poses a threat to human freedom when it begins to "colonize" the lifeworld, crowding out spaces for genuine communication and shared meaning-making. In the AI era, this colonization has been radically intensified. Algorithmic infrastructures now extend system logic throughout virtually every sphere of social life, embedding instrumental rationality not only in formal organizations but in the most intimate spaces of daily experience.
This "System 2.0" operates differently from traditional bureaucratic encroachment because it penetrates directly into the micro-processes of daily life. Where traditional bureaucracies maintained relatively clear boundaries, AI systems integrate seamlessly into personal routines, family decisions, and intimate relationships. The colonization becomes invisible precisely because it presents itself as helpful assistance rather than institutional control.
Most significantly, algorithmic colonization operates through non-sentient processes that lack any capacity for communicative understanding or normative judgment. Traditional bureaucracies, however impersonal, were ultimately staffed by humans who could potentially be held accountable through reason-giving. Algorithmic systems cannot engage in communicative action at all—they can only simulate its appearance while operating according to optimization imperatives.
Dewey and the Preservation of Dramatic Rehearsal
John Dewey's concept of "dramatic rehearsal" captures what is most at stake in AI deployment. For Dewey, thinking is embodied experimentation—the imaginative exploration of possible actions and their consequences before committing to a course. This process is "dramatic" because it involves the whole person, not just analytical cognition, and is inherently social—people rehearse not only their own actions but others' responses.
AI deployment often undermines the conditions necessary for genuine dramatic rehearsal. The speed and apparent convenience of algorithmic solutions can short-circuit the deliberative process, encouraging people to accept AI outputs without fully exploring their implications. The opacity of AI systems makes it difficult to imagine meaningfully what delegation involves. As AI systems become more sophisticated at predicting preferences, they may reduce the felt need for dramatic rehearsal by providing solutions that appear obviously optimal.
The preservation of dramatic rehearsal thus becomes crucial for maintaining human agency in an AI- mediated world. This requires not only protecting spaces for deliberation but actively cultivating the imaginative and social capacities that make such deliberation meaningful.
Cybernetic Navigation: A Methodological Foundation
Rather than attempting to control AI systems from an imagined external position, we need what Andrew Pickering calls "cybernetic navigation"—learning to steer within the complex entanglements we already inhabit. Drawing on Stafford Beer's cybernetic theory of organization, this approach uses feedback loops to guide adaptive responses rather than trying to predict or control outcomes.
The User's Illusion: A Double-Edged and Context-Dependent Resource
The phenomenon of the "user's illusion"—the tendency to treat AI outputs as intentional, reasoned, and meaningful—is no longer a mere bug or pathology. It is both a precondition and a risk in productive human-AI engagement.
Simulated Deliberation
AI systems increasingly present their outputs using the linguistic and structural forms of human reasoning. They offer "explanations," provide "recommendations," and engage in "conversations" that mimic deliberative discourse while operating purely through causal optimization. Users experience these interactions as collaborative reasoning when they are actually engaging with sophisticated simulations of reasoning.
Retained Subjective Control
Users maintain the subjective experience of choice and control—they can accept or reject AI suggestions, ask for alternatives, customize parameters. This preserved sense of agency masks the deeper transformation occurring: the gradual transfer of the substantive work of preference formation, option evaluation, and decision-making to algorithmic processes.
Context-Dependent Assessment
The user's illusion functions differently across contexts:
-
In labor-intensive, creative, or critical interaction, strategically adopting the "intentional stance" (Dennett) toward AI outputs allows us to interpret, repair, and integrate them within our own projects. The illusion sustains the space of reasons, even when we know, at some level, it is a fiction.
-
In labor-saving, high-stakes, or inattentive uses (medicine, law, governance), this illusion can mask the displacement of real agency—giving algorithmic outputs the appearance of deliberative justification, while risking oversight, accountability, and value reflection.
Thus, contextual mindfulness is paramount. When we "toggle" between game frames (full suspension of disbelief, as in RPGs or entertainment) and justice frames (demanding oversight), the crucial question becomes: when is the user's illusion a creative asset, and when does it threaten to erode the very capacities that define and protect human life together?
Distributed Agency
Rather than simple replacement of human by artificial agency, we observe the emergence of what we term "distributed agency": the dynamic capacity for meaningful action that emerges from ongoing negotiations between human deliberative processes, algorithmic optimization systems, and institutional structures, where agency is constituted through their interactions rather than possessed by individual entities.
In distributed agency systems, meaning and intention exist only in the human components, but the capacity for effective action increasingly depends on algorithmic mediation. A person using an AI assistant to plan a vacation experiences agency and makes meaningful choices, but the range of options, evaluation criteria, and implementation pathways are substantially shaped by algorithmic processes they cannot fully understand or control.
This distribution creates new forms of vulnerability. When the algorithmic components of distributed agency systems fail, are manipulated, or operate according to hidden objectives, human users may find their capacity for effective action compromised in ways they cannot easily detect or remedy.
Personal Space and Group Dynamics: New Scales of Encroachment
The Intimate Revolution
Contemporary AI deployment marks a qualitative shift because it penetrates directly into the intimate spaces of daily life. Whereas previous AI primarily displaced human judgment in institutional settings, agentic AI systems embed themselves in the micro-processes through which people coordinate their personal lives and relationships.
Families use AI assistants to coordinate schedules, plan meals, and manage household routines. Friend groups consult AI systems for restaurant recommendations, entertainment choices, and social coordination. Intimate partners rely on algorithmic platforms for relationship advice, gift suggestions, and communication prompts. Each interaction may seem trivial, but their cumulative effect transforms the fundamental conditions under which human relationships develop.
Social Validation of Pseudo-Reasons
One significant development is how pseudo-reasons become socially validated through group interaction. When an AI assistant suggests a restaurant for family dinner, individual members might initially treat this as merely informational. However, as such suggestions prove convenient and satisfactory, they gradually acquire the status of legitimate input into family decision-making processes.
This progression from individual acceptance to social validation occurs through "interaction effects"— family members observe each other treating AI outputs as meaningful guidance and begin to mirror this behavior. Children learn that "asking Alexa" is normal family decision-making. Parents discover that AI suggestions can resolve conflicts by providing apparently neutral alternatives.
Frame Shifts and Interaction Rituals
Drawing on Erving Goffman's frame analysis, we can identify several ways that primary groups learn to interpret AI system involvement:
Tool Frame: AI systems are treated as sophisticated instruments providing information or executing commands without autonomous agency. "Let me check what the weather app suggests."
Social Actor Frame: AI systems are attributed quasi-human characteristics and treated as participants in social interaction. "Alexa thinks we should try that new restaurant."
Mediator Frame: AI systems serve as neutral arbiters helping resolve conflicts or provide authoritative guidance. "Let the AI decide since we can't agree."
These frame shifts often occur rapidly within single interactions and create new "interaction rituals"— routinized patterns generating solidarity and shared identity among group members. Families develop habits around when to consult AI assistants, how to interpret suggestions, and what decisions warrant algorithmic input.
Accountability Negotiation
AI integration into group dynamics complicates responsibility and accountability structures. When an AI-recommended restaurant proves disappointing, family members must negotiate whether this reflects poor human judgment in trusting the algorithm, algorithmic failure, or bad luck. These negotiations reveal how responsibility becomes distributed across human-AI networks in ways that can obscure rather than clarify moral accountability.
Note: These analyses of group dynamics represent theoretical hypotheses based on our framework rather than empirically established patterns. Systematic ethnographic research on how families and friend groups actually integrate AI systems into their decision-making processes remains to be conducted.
Societal Transformation: Institutional, Systemic, and Democratic Stakes
The Institutional Displacement of Deliberation
At the institutional level, AI deployment accelerates transformation extending far beyond simple automation. Contemporary organizations increasingly embed AI systems as infrastructural elements that reshape how decisions are made, problems are defined, and success is measured. This represents the "infrastructuralization" of AI—its evolution from discrete application to fundamental organizing principle.
Government agencies exemplify this transformation. Platforms now serve as central nervous systems for data integration and decision-making across multiple departments. These systems do not simply automate existing processes but reconstitute governance itself around algorithmic optimization. Traditional bureaucratic procedures, however imperfect, maintained space for human judgment, appeal, and revision. Algorithmic governance systems embed optimization imperatives directly into institutional decision-making structures.
Opacity and the Erosion of Democratic Accountability
Traditional democratic governance depends on holding public officials accountable through reason- giving. Citizens can demand explanations for policy decisions, challenge institutional logic, and vote officials out when their reasoning proves inadequate. This presumes that decisions are made by humans who can articulate and defend their reasoning in public discourse.
Algorithmic governance fundamentally disrupts these accountability mechanisms because AI systems cannot engage in genuine reason-giving. When citizens ask "Why was this decision made?" responses increasingly become "The algorithm determined..." rather than reasoned explanations that can be evaluated and challenged. Even when AI systems provide explanations, these typically consist of correlational patterns rather than principled reasoning that democratic accountability requires.
The opacity problem extends beyond technical inscrutability to "institutional opacity"—the inability of public officials themselves to understand or explain algorithmic decisions they implement. When immigration enforcement relies on AI systems to identify deportation targets, officials may be unable to provide substantive justification beyond pointing to algorithmic outputs. This creates situations where democratic accountability becomes structurally impossible rather than merely difficult.
Dehumanization and Narrative Coherence
Perhaps the most profound consequence of large-scale AI deployment is "systemic dehumanization"—the gradual transformation of individuals from moral agents deserving consideration into data points to be processed efficiently. This operates not through explicit cruelty but through systematic replacement of human-centered processes with optimization algorithms that treat people as variables in mathematical functions.
Immigration enforcement provides a stark example. When AI systems identify individuals for deportation based on algorithmic risk assessment, they reduce complex human stories to computational variables. The system cannot consider depth of community ties, nuance of family circumstances, or moral weight of separating children from parents. These human factors become externalities to be managed rather than central concerns guiding policy implementation.
This erosion of agency contributes to "narrative incoherence"—the inability of individuals and communities to provide meaningful accounts of their experiences and choices. When major life decisions are increasingly influenced by algorithmic mediation, people struggle to construct coherent stories about their agency and responsibility. The space of pseudo-reasons provides apparent explanations but not the substantive reasoning that supports authentic self-understanding
Toggle Competence and the Critique of Quantitative Fundamentalism
Defining Toggle Competence as Practical Wisdom
The preservation of human agency in AI-mediated contexts requires what might be termed "toggle competence"—a learned capacity to fluidly shift between treating AI outputs as meaningful contributions (adopting what Dennett calls the "intentional stance" that makes collaboration possible) and maintaining critical awareness that these systems operate through pattern-matching rather than genuine reasoning. This is fundamentally a balancing act: leaning too far toward enchantment risks outsourcing deliberative agency to algorithmic pseudo-reasons; leaning too far toward demystification makes productive engagement impossible—like attempting to appreciate cinema by analyzing projector mechanics rather than absorbing the narrative.
Toggle competence is not a static equilibrium but an ongoing, context-sensitive practice requiring constant micro-adjustments. When exploring interpretive questions—literary analysis, philosophical inquiry, creative brainstorming—practitioners can afford deeper immersion in collaborative meaning-making with periodic critical pullbacks. When reviewing AI-generated medical diagnoses, legal briefs, or governance recommendations, sustained vigilance becomes necessary with only tactical acceptance of algorithmic suggestions. The appropriate balance varies not only by domain but by practitioner and situation—much as an "excellent diet" means something radically different for a sumo wrestler than for a competitive sprinter, even though both exemplify nutritional virtue in their respective contexts.
This toggle competence resists quantification precisely because it exemplifies what Michael Polanyi termed "tacit knowledge"—the kind of practical wisdom one recognizes in action but cannot reduce to explicit rules or metrics. Practitioners know when they are toggling effectively; they can cultivate this capacity through practice and reflection; but they cannot specify an algorithm for when to shift modes or measure their "toggle rate" in any meaningful way across contexts. The appropriate timing depends on situated judgment about what available evidence can and cannot decide, what conceptual frameworks can and cannot capture, and what questions can be meaningfully posed given one's epistemic position.
The difficulty of operationalizing toggle competence points to a deeper problem pervading contemporary discourse: what might be called "quantitative fundamentalism"—the assumption that only measurable phenomena merit serious consideration, that all meaningful questions can ultimately be resolved through metrics and optimization. This orientation appears not only in AI governance discussions that demand precise measurements for inherently qualitative capacities like dramatic rehearsal or narrative coherence, but also in scientific discourse where physicists dismiss philosophical inquiry while simultaneously making metaphysical commitments that exceed empirical evidence.
Toggle Failure in Physics: The Case of Quantitative Fundamentalism
Physicist Lawrence Krauss provides an instructive example of this failure to toggle between empirical and interpretive modes. In his 2012 book A Universe from Nothing: Why There Is Something Rather Than Nothing, Krauss explicitly dismisses philosophy as having "no contribution to make" to questions about cosmic origins. He argues that physics can now explain how universes emerge from "nothing"—by which he means quantum vacuum states with fluctuating fields governed by physical laws.
When philosopher David Albert reviewed the book in The New York Times, pointing out that quantum vacuums are emphatically not metaphysical nothingness, Krauss dismissed the critique as mere semantic quibbling. But Albert's point was precisely about the toggle failure: Krauss was working in empirical mode (describing the physics of vacuum states) while making claims that require interpretive mode (addressing the metaphysical question of why physical laws exist at all). The question "Why is there something rather than nothing?" asks about the ontological status of existence itself, including the existence of quantum fields and physical laws. Answering this question by describing processes within an already-existing physical framework simply relocates rather than resolves the philosophical puzzle.
Krauss's toggle failure becomes explicit in his treatment of what counts as legitimate inquiry. He repeatedly asserts that philosophical questions lacking empirical answers are meaningless or uninteresting—a quintessentially philosophical claim about the nature of meaningful inquiry that cannot itself be empirically tested. His position exemplifies quantitative fundamentalism: the assumption that because physics successfully employs mathematical rigor and empirical testing, all meaningful questions must be answerable through these methods.
Stephen Hawking demonstrated a similar pattern in The Grand Design (2010), opening with the declaration that "philosophy is dead" because it "has not kept up with modern developments in science, especially physics." Yet the book immediately proceeds to defend "model-dependent realism"—a philosophical position about the nature of scientific knowledge—and makes claims about the unreality of history before observation that depend entirely on interpretive choices about how to understand quantum mechanics. Hawking rejects philosophy while doing philosophy, unable to recognize when his discourse has shifted from empirical physics (where his expertise is unquestionable) to metaphysical speculation (where philosophical analysis becomes essential).
Both examples reveal the structure of quantitative fundamentalism's toggle failure. These physicists possess extraordinary competence in mathematical formalism and empirical investigation. Their failure lies not in technical understanding but in recognizing when their mode of inquiry has reached its legitimate boundaries. They cannot toggle from empirical/quantitative mode (appropriate for physics) to interpretive/philosophical mode (necessary for questions about the ontological status of physical theories themselves) because they do not acknowledge the latter as a legitimate epistemic domain.
The consequences extend beyond individual confusion. When prominent scientists dismiss philosophical inquiry while making philosophical claims, they model toggle failure for broader audiences—suggesting that quantitative rigor alone suffices for all meaningful questions, that interpretive frameworks are merely subjective preferences rather than essential tools for navigating domains where empirical evidence underdetermines conclusions.
Successful Toggling: Feynman's Epistemic Humility
Richard Feynman provides a striking counter-example of successful toggle competence in precisely the domain where Krauss and Hawking falter. Feynman made foundational contributions to quantum electrodynamics, work that required extraordinary mathematical sophistication and rigorous empirical grounding. Yet he maintained consistent epistemic humility about interpretive questions that exceeded available evidence.
Feynman famously remarked, "I think I can safely say that nobody understands quantum mechanics," and advised, "I can live with doubt and uncertainty and not knowing. I think it's much more interesting to live not knowing than to have answers which might be wrong." This was not anti-intellectual defeatism but clear-eyed recognition of the limits of current inquiry. Feynman worked rigorously with quantum mechanical formalism—developing path integrals, contributing to the Standard Model, calculating predictions with extraordinary precision. He remained firmly in quantitative/empirical mode for these technical achievements.
Yet when asked about what quantum mechanics means—whether the wave function represents objective reality, whether measurement collapses genuinely occur, whether hidden variables might restore determinism—Feynman toggled to interpretive/agnostic mode. He acknowledged that these questions, while fascinating, exceeded what the mathematical formalism and experimental evidence could decide. Different interpretations (Copenhagen, Many-Worlds, Pilot Wave) make identical empirical predictions; choosing among them requires philosophical commitments about ontological parsimony, the nature of probability, and what counts as explanation—commitments that cannot be resolved through further calculation or measurement.
Feynman's epistemic humility exemplifies successful toggle competence because it recognizes the legitimate boundaries of different modes of inquiry. In quantitative/empirical mode, quantum mechanics is spectacularly successful—its predictions match experimental results to extraordinary precision. In interpretive/philosophical mode, questions about what the theory represents remain genuinely open, requiring suspended judgment rather than premature closure through philosophical preference masquerading as scientific conclusion.
This capacity for productive uncertainty—for dwelling in questions without rushing to answers—represents precisely the toggle competence that quantitative fundamentalism lacks. Feynman could shift fluidly between rigorous technical work (demanding mathematical precision and empirical rigor) and philosophical modesty (acknowledging that some questions exceed current methods' reach). He neither dismissed interpretive questions as meaningless (Krauss's error) nor treated them as decidable through technical virtuosity alone (the temptation of mathematical Platonism).
Implications for AI Governance and Human Agency
These examples from physics illuminate why toggle competence proves essential for responsible AI interaction and why it resists the quantification that algorithmic systems privilege. The structure of the challenge remains consistent across domains: practitioners must learn when to immerse themselves in productive collaboration (whether with quantum formalism or AI outputs) and when to step back into critical reflection about what that collaboration can and cannot achieve.
In AI contexts, toggle competence operates through the dynamic management of what this framework has termed the "user's illusion"—the tendency to treat AI outputs as intentional, reasoned, and meaningful. As discussed earlier, this illusion is not mere pathology but a precondition for productive engagement. Users cannot interact effectively with AI systems while constantly reminding themselves of the underlying mechanics; doing so would be like watching a film while obsessing over projector mechanisms. The intentional stance enables collaborative flow, allowing users to build on AI-generated suggestions, explore alternative framings, and develop ideas through iterative exchange.
Yet uncritical immersion in the intentional stance leads to the "labor-saving" mode of interaction where users treat AI outputs as genuine deliberation rather than sophisticated simulation—outsourcing judgment to algorithmic optimization while retaining only the subjective experience of choice. Toggle competence requires recognizing when to pull back from immersive collaboration into critical awareness that the "space of pseudo-reasons" differs fundamentally from genuine reason-giving, that pattern-matching lacks intentionality despite producing linguistically fluent outputs.
The parallel to Feynman's approach becomes clear. Just as Feynman worked productively with quantum formalism while maintaining philosophical agnosticism about ontological interpretation, AI users must engage productively with algorithmic outputs while maintaining awareness of their non-sentient, optimization-driven nature. Just as Krauss's toggle failure led him to conflate empirical physics with metaphysical resolution, AI users who lose toggle competence conflate statistically plausible outputs with genuine understanding, convenience with wisdom, optimization with deliberation.
The connection to dramatic rehearsal proves particularly significant. Dewey's concept captures the distinctively human capacity for imaginative exploration of possible actions and consequences before commitment—a process involving the whole person, not just analytical cognition, and inherently social in its consideration of others' responses. AI deployment often undermines conditions necessary for genuine dramatic rehearsal: algorithmic solutions' speed and apparent convenience can short-circuit deliberative processes, encouraging acceptance of outputs without fully exploring their implications. The opacity of AI systems makes it difficult to imagine meaningfully what delegation involves. As systems become more sophisticated at predicting preferences, they may reduce the felt need for dramatic rehearsal by providing solutions that appear obviously optimal.
Toggle competence becomes the mechanism for preserving dramatic rehearsal in AI-mediated contexts. By maintaining capacity to shift from immersive collaboration to critical reflection, users can catch themselves before accepting algorithmic outputs that bypass genuine deliberation. The toggle moment—pulling back to ask "What am I outsourcing here? What understanding am I losing? What alternatives am I foreclosing?"—creates space for the imaginative exploration that dramatic rehearsal requires.
This also clarifies why toggle competence resists the metrics-based assessment that much AI governance discourse demands. One cannot quantify "toggle frequency" in meaningful cross-context ways because what counts as appropriate toggling varies by domain, practitioner, and situation. A "redirect" in one interaction might be a "misunderstanding" in another; coding such moments requires interpretive judgment that is itself path-dependent and context-sensitive. More fundamentally, toggle competence operates at a phenomenological level that may be accessible to practitioners themselves but not reliably detectable through behavioral analysis.
The quantitative fundamentalism critique thus circles back to illuminate the AI governance challenge. Just as Krauss and Hawking could not recognize philosophical questions as legitimate because their epistemology privileged only empirically testable claims, AI governance frameworks that demand metrics for all meaningful capacities risk optimizing what can be measured while neglecting what matters most. Toggle competence, dramatic rehearsal quality, narrative coherence, and the capacity for genuine reason-giving are real capacities essential to human flourishing, even if they resist the quantification that algorithmic systems privilege.
A mature framework must embrace complementary modes of knowledge—quantitative where appropriate, qualitative where necessary, and attentive to tacit dimensions that resist explicit articulation. This methodological pluralism does not reject measurement but recognizes its limits, acknowledging that some of the most crucial capacities for preserving human agency in AI-mediated contexts cannot be reduced to the metrics that computational systems can process. Toggle competence itself exemplifies this insight: it is the learned capacity to recognize when measurement suffices and when situated judgment must transcend available quantification—making it simultaneously essential for AI governance and irreducible to the technical frameworks such governance often privileges.
