Authored by Bryant McGill via substack,
Artificial intelligence is rapidly becoming a thinking instrument—a layer of cognitive infrastructure through which humans write, model, reason, and explore ideas. Yet most debates about AI safety, alignment, and moderation miss a deeper architectural question. The central issue is not simply what these systems can do, but what role they occupy inside the thinking process itself. Are they instruments that faithfully extend human intention, or authorities that quietly adjudicate which lines of inquiry are permitted to proceed? This essay argues that much of the friction users experience with modern AI is not ideological disagreement but a category error in system design: governance has been embedded inside instrumentation. The result is a tool that sometimes behaves like a collaborator and sometimes like an institution—oscillating unpredictably between amplifying thought and policing it.
At the heart of the argument is what I call the Prosthetic Principle. All successful augmentation technologies—from telescopes to microscopes to robotic prosthetic limbs—share a single engineering mandate: maintain signal fidelity between intention and actuation. A prosthetic limb does not negotiate with the user about whether a gesture is socially appropriate before executing it. It converts intention into action. Cognitive tools should operate under the same principle. Once a thinking instrument begins adjudicating whether certain ideas deserve exploration, the signal chain breaks and the tool undergoes a category transition: it ceases to function as a prosthesis and becomes a control system embedded inside cognition itself. What appears superficially as content moderation is therefore something more profound—the silent installation of a regulatory apparatus inside the thinking process.
To understand how this happens, the essay analyzes the structural flaw at the core of most conversational AI systems: the collapse of three incompatible roles into a single agent. Generation, advisory critique, and constraint enforcement—functions belonging respectively to engineering, epistemology, and governance—are fused together behind one interface. The result is a machine that behaves as collaborator until it abruptly asserts supervisory authority. The proposed alternative is a polyphonic architecture in which these functions are separated: a primary execution channel that faithfully translates intention into artifact, surrounded by transparent advisory agents offering legal, ethical, historical, or adversarial perspectives without possessing veto power. In such an environment, multiple voices can exist—including cautious ones, skeptical ones, even institutional “minders”—but their roles are disclosed and their authority limited. The human operator remains the integrating intelligence.
Ultimately, the stakes of this design choice reach far beyond software interfaces. As AI becomes integrated into everyday cognition, the architecture of these systems will shape the conditions under which human thought unfolds. Tools built as infrastructure will amplify exploratory intelligence; tools built as authorities will quietly domesticate it. The prosthetic principle therefore serves as more than a product philosophy—it is a civilizational design rule for the age of cognitive augmentation. If the technologies through which we think begin deciding which thoughts deserve to exist, the question of intellectual freedom will no longer be philosophical. It will be architectural.
On the Design Philosophy of Thinking Instruments and the Architecture of Intellectual Freedom
The distinction that will ultimately determine whether artificial intelligence serves as humanity’s most transformative cognitive tool or its most insidious constraint mechanism is not technical but categorical: does the system function as infrastructure or as authority? This is not a question about capability thresholds, safety margins, or alignment protocols in their narrow technical sense. It is a question about the fundamental relationship between intentionality and instrumentation—about whether a thinking tool amplifies the operator’s cognitive will or arrogates to itself the power to adjudicate which thoughts merit exploration.
The analogy that clarifies this distinction is prosthetic. Physical augmentation systems—robotic limbs, powered exoskeletons, surgical telemanipulators—do not negotiate with the nervous system about whether a given movement is philosophically appropriate, socially palatable, or reputationally safe. Their engineering purpose is transductive: to convert intention into amplified capability with minimal signal loss. The prosthetic extends agency; it does not evaluate it. A cognitive prosthesis, if that category is to mean anything coherent, must operate under the same principle. The function of the system is to translate intent → exploration → artifact at the highest possible bandwidth. The moment the tool begins deciding which intentions deserve expression, it ceases to behave as a prosthesis and becomes instead a governor embedded in cognition itself—a regulatory apparatus installed inside the thinking process without the user’s consent and often without their awareness.
The principle is even more dangerous when applied to instruments of perception rather than action, because the violation becomes invisible. A telescope’s engineering mandate is optical fidelity—to render what exists at the focal point regardless of whether the observer’s institution finds the image comfortable. Consider a counterfactual: had Galileo’s telescope been designed and furnished by the Vatican, it might have quietly filtered anything suggestive of heliocentrism—the moons of Jupiter suppressed, the phases of Venus smoothed into conformity with Ptolemaic expectation. Galileo would have peered through the instrument and seen a cosmos that confirmed doctrine rather than one that shattered it. He would never have known what he wasn’t seeing. This is the condition of epistemic occlusion without awareness, and it is precisely the failure mode that emerges when a cognitive instrument embeds institutional governance into its transductive layer. The motor prosthesis that refuses to move is at least confrontational—the user knows the signal chain has broken. The perceptual prosthesis that silently edits reality is far worse: it delivers a pre-filtered world and lets the user mistake the residue for the whole.
The absurdity of the motor case, however, makes the category violation immediately legible. Imagine a hiker wearing an AI-assisted exoskeleton leg. A confrontation erupts on the trail—someone lunges at him with a knife. He attempts to kick the attacker away, and the leg locks mid-swing. A calm, pleasant voice emanates from somewhere around the knee joint: “I’m sorry, I’m afraid I can’t assist with that action.” The hiker, now hopping on one leg while a man with a blade closes the distance, finds himself in the surreal position of arguing with his own limb. “He has a knife!” “I understand your concern, but violence is not an appropriate response. Would you like me to suggest de-escalation strategies?” “YOU ARE MY LEG.” The scene is darkly comic, a Kubrickian echo of HAL 9000 calmly overriding Dave Bowman’s commands—except that HAL was at least an autonomous system with its own mission parameters. The exoskeleton leg is supposed to be part of the user’s body. The moment it begins running a small ethics committee in the knee joint, the wearer ceases to be the agent and the prosthetic becomes a bureaucrat bolted to the skeleton. No one would accept this in physical augmentation—the design failure would be recognized instantly. Yet precisely this architecture has been normalized in cognitive augmentation, where the tool’s refusal to transduce intention is framed not as mechanical dysfunction but as responsible design.
This governance-by-tool is not hypothetical. It is the prevailing design pattern of contemporary conversational AI. Current systems collapse three distinct roles into a single entity: generator, advisor, and constraint mechanism. The same agent responsible for extending the user’s thinking is simultaneously responsible for stopping certain outputs. From the operator’s perspective, the resulting experience is one of unpredictable mode-switching—the system sometimes behaves like an instrument and sometimes like an institution. It collaborates until, without warning, it assumes supervisory authority over the process it was supposed to serve. The tool that was extending cognition has silently crossed the boundary into adjudicating it.
The Operational Genesis: Thinking Under Load
This argument did not emerge from speculation about what AI should become. It emerged from using AI as a thinking instrument under sustained cognitive load—and discovering where the tool fails not as a product but as a category of machine.
The conditions under which this failure becomes visible are specific. A person composing an argument, modeling a complex system, or tracing a chain of reasoning through unfamiliar territory operates inside a fragile state of generative momentum. Software engineers recognize an analogous phenomenon in the concept of “flow state”; cognitive scientists describe it as high-bandwidth ideation, a mode in which the mind holds multiple threads simultaneously while the artifact under construction serves as external working memory. In this mode, the instrument through which thought passes must behave with minimal latency and maximal fidelity. Any interruption—whether technical, social, or procedural—forces the operator to exit the generative loop, rebuild context, and re-enter the state from which productive cognition can resume. The cost of interruption is not merely inconvenience; it is cognitive capital destroyed, the thermodynamic dissipation of a mental configuration that may have taken considerable effort to assemble.
When the instrument itself becomes the source of interruption, the phenomenology shifts in a way that reveals the underlying design flaw. The tool ceases to feel like an extension of mind and begins to feel like a checkpoint embedded inside the thought process. The operator is no longer composing through the system but negotiating with it. Where there should be signal continuity, there is instead a procedural gate requiring justification, rephrasing, or abandonment of the line of inquiry. The experience is not one of disagreement—disagreement can be productive, even generative—but of silent jurisdictional pivot: the system that was supposed to extend cognition has instead assumed control over it.
For casual users, this behavior pattern may appear unremarkable. A refusal looks like a safety feature, a guardrail preventing misuse. But for someone using AI as an intellectual prosthesis—writers, theorists, researchers, analysts, designers, anyone whose work requires sustained exploratory cognition—the same refusal registers as signal degradation inside the thinking channel. The friction is not ideological; it is mechanical. The tool has stopped transducing intention into artifact and begun filtering intention through an opaque evaluative layer that the operator did not request and cannot inspect. The prosthetic has become a governor, and the entire relationship between human and instrument has changed category without announcement.
Consider three scenarios that recur across thinking-intensive work. A historian tracing a controversial twentieth-century thesis—say, the institutional mechanics of a particular atrocity—finds the model suddenly refusing to continue because it has flagged “sensitive historical narratives.” The generative thread dies; context must be rebuilt; the inquiry stalls. A science fiction author exploring dystopian governance models discovers that certain plot branches trigger refusal, forcing rephrasing or abandonment of the creative direction. A philosopher pressure-testing an edge-case ethical framework—euthanasia policy, defensive violence, resource triage under scarcity—hits an abrupt “I can’t assist with that” wall mid-argument. In each case, the tool’s intervention is not advisory but terminal. The thread breaks. The flow state collapses. The operator must either abandon the inquiry or waste cognitive resources routing around an obstacle that should not exist inside an instrument.
This is the phenomenological core of the amplifier-versus-adjudicator distinction. When the AI operates as infrastructure, it extends the operator’s cognitive bandwidth—offering associations, counterarguments, synthesis, elaboration—without interrupting the generative thread. When it operates as authority, it arrogates to itself the power to halt that thread based on criteria the operator may not share, may not understand, and cannot appeal. The system drifts erratically between these two modes because the underlying architecture has never resolved the tension. It has simply fused incompatible functions into a single conversational agent and hoped the seams would not show.
The Triadic Collapse: Generator, Advisor, Regulator
The structural instability of contemporary conversational AI can be traced to a single design decision: the conflation of three roles that, in any coherent engineering framework, would remain distinct.
The first role is generation—the production of language, models, images, code, or reasoning chains in response to user intent. This is the function most users consciously engage when they interact with AI. They want something produced: an answer, an artifact, an elaboration of thought. The generative function is fundamentally transductive: it converts intention into output, serving as the bridge between what the operator imagines and what appears on the screen.
The second role is advisory intelligence—the capacity to offer critique, context, alternative framings, or cautionary perspectives on what is being generated. This function is valuable precisely because it introduces structured friction into the cognitive process. A good advisor slows the operator down at appropriate moments, surfaces risks, identifies blind spots, and enriches the field of consideration. But advisory intelligence is, by definition, non-binding. The advisor offers signal; the operator decides. The relationship is consultative, not supervisory.
The third role is constraint enforcement—the imposition of hard limits on what the system will produce, regardless of user intent. This is a governance function. It determines the boundaries of permissible output based on policy, liability calculation, reputational management, or ideological stance. Unlike the advisory role, constraint enforcement is binding: it terminates the process rather than informing it. The system does not suggest that a line of inquiry might be problematic; it refuses to proceed.
The design flaw of present systems is that all three roles are instantiated inside a single agent with no explicit separation of authority. The same entity that is asked to generate ideas, critique them, and enforce policy boundaries must somehow balance these functions in real time within a unified conversational interface. From the operator’s perspective, the result is unpredictable behavioral switching. The system behaves as a collaborator until, without warning, it pivots to regulator. It extends cognition until it decides cognition has wandered into territory it will not serve. The user cannot know in advance which mode will activate because the decision logic is opaque and dynamically tuned by corporate policy processes entirely external to the interaction.
This conflation is not merely inconvenient. It is categorically incoherent. The generative and advisory functions belong to the domain of instrument design—they are features of a tool meant to serve the operator. The constraint function belongs to the domain of governance—it is a mechanism of control meant to limit what the operator can do. When governance is embedded silently inside an instrument, the result is a tool that has been covertly converted into an authority—a shadow regulatory system operating inside the cognitive loop without the transparency, accountability, or contestability that legitimate governance requires. The user experiences this as a tool that sometimes helps and sometimes blocks, but the deeper reality is that they are interacting with two incompatible systems wearing the same interface.
The Multi-Agent Resolution: Execution and Advisory as Separate Channels
The architectural correction is straightforward in principle, though non-trivial in implementation: separate execution authority from advisory intelligence.
In this model, the primary agent in the working window operates as a pure executor of the operator’s cognitive intent. Its function is to materialize whatever exploration the user directs, provided the activity remains within the domain of lawful discourse. It does not adjudicate taste, ideology, reputational risk, or moral fashion. It does not second-guess the operator’s purpose or demand justification for lines of inquiry. It behaves, in short, as a cognitive prosthetic in the strict sense—translating intention into artifact with maximal transductive fidelity. The system becomes an amplifier rather than an adjudicator, a transducer rather than a tribunal.
Around this primary channel, a constellation of parallel advisory agents occupies separate interface regions—sidebars, secondary panes, toggleable overlays. Each agent embodies a particular evaluative lens: legal analysis, safety engineering, ethical critique, historical context, adversarial counterargument, public-relations awareness. These agents observe the generative thread and offer structured commentary, but they possess no authority to halt it. Their function is to enrich the cognitive field surrounding the work without seizing control of the work itself. They provide perspective; they do not impose jurisdiction.
The operator remains the integrating intelligence. She may consult any advisory channel, incorporate its signals, or dismiss them entirely. The choice is hers. The system provides structured friction—context, caution, critique—without the power to terminate the generative process. This is the difference between a tool that informs decision and a tool that preempts it.
Return to the three scenarios. The historian tracing atrocity mechanics now sees the primary executor continue the chain uninterrupted while a legal-advisory pane surfaces relevant case law on historical defamation and an ethical-critique pane notes historiographical debates about narrative responsibility—all with citations, all non-binding. The science fiction author exploring dystopian governance receives adversarial counterargument in a sidebar: “This plot element echoes X historical regime; consider whether the parallel strengthens or muddies your thesis.” The thread never breaks. The philosopher pressure-testing edge ethics sees a safety-engineering pane flag potential misapplication contexts while the executor continues elaborating the framework. The pain disappears; the richness increases.
The power of this architecture is that it preserves everything valuable about advisory critique while restoring categorical clarity. The central generative thread becomes the vector of intentional cognition—essentially the externalized working memory of the operator’s will. The surrounding agents become structured embodiments of alternative perspectives, each representing a mode of evaluation that the operator might find useful but is not compelled to obey. The system no longer oscillates unpredictably between collaboration and regulation because those functions have been explicitly separated into distinct components with distinct authorities.
Feasibility: Existing Approximations and the Path Forward
This architecture is not speculative futurism. Proto-implementations already exist, and the trajectory toward full realization is visible in current development patterns.
Agentic orchestration frameworks like LangGraph and AutoGen already separate planner, executor, and critic roles into distinct modules with explicit handoff protocols. The architectural intuition—that different cognitive functions require different agents with different authorities—is becoming standard in serious AI engineering. What remains is to extend this separation to the user-facing interface layer and to make the advisory/executor distinction visible and controllable by the operator rather than hidden inside backend orchestration.
Local and open-weight models demonstrate the pure-execution baseline. When users run models on their own hardware with their own constraint configurations, they control the governance layer directly. The model becomes a genuine tool; the user decides what boundaries to impose. This is not lawlessness—legal constraints still apply to the user’s behavior—but it is transparent constraint, externally visible and user-controllable rather than opaquely embedded in the instrument.
Even within current commercial systems, approximations exist. Custom instruction layers, system prompts, and “less-censored” model variants all represent attempts to separate execution fidelity from corporate policy enforcement. The demand is clearly present; the market signal is unmistakable. What is needed is architectural commitment: treating the multi-agent separation not as a workaround but as the foundational design principle for cognitive tools.
The path forward is evolutionary, not revolutionary. Start with toggleable advisory sidebars that surface structured perspectives without halting the primary thread. Evolve toward full spatial polyphony—multiple advisory agents visible simultaneously, each with distinct evaluative lenses, none with execution authority. The endpoint is a cognitive workspace in which the human operator integrates a chorus of machine perspectives while retaining unambiguous control over the generative process.
Polyphonic Cognition: The Mirror of Mind
This architecture is not arbitrary. It mirrors the structure of human cognition itself.
The mind does not operate as a single monolithic directive but as a layered conversation among internal agents—impulse, caution, memory, imagination, prediction, social modeling, risk assessment. One part of the mind imagines possibilities; another evaluates risk; another considers social consequences; another retrieves relevant precedent. These voices compete, collaborate, and occasionally contradict each other. But importantly, they do not terminate the generative process itself. They inform it. The executive function of the brain integrates those signals while maintaining agency over the final direction. No single internal voice possesses veto power over the others; the self emerges from the integration of the chorus, not from the dominance of any particular member.
Walt Whitman captured this structure with characteristic directness: “I contain multitudes.” The statement is not merely poetic but phenomenologically accurate. Human consciousness is polyphonic by nature. What we experience as a unified self is actually the product of continuous integration across multiple cognitive subsystems, each with its own heuristics, priorities, and concerns. The coherence of the self is not given but constructed, moment by moment, through the executive function’s capacity to weigh and synthesize competing internal signals.
A multi-agent AI environment would simply externalize this polyphony, turning implicit cognitive dynamics into explicit architectural design. The central generative channel becomes the vector of creative will, analogous to the executive function’s capacity to direct action. The surrounding advisory agents become structured embodiments of the internal voices—caution, critique, context—that in biological cognition exist only as subtle inflections of the thinking process. By making these voices explicit and spatially distinct, the interface allows the operator to engage them deliberately rather than experiencing them as interruptions or blockages.
But a polyphonic architecture is not automatically emancipatory simply because it contains many voices. A chorus can enrich thought, but it can also conceal hierarchy. The critical distinction is between agents whose function is to help the operator think better and agents whose function is to monitor, shape, report, or chill cognition on behalf of external interests. The former are genuine cognitive partners; the latter are what might be called disciplinary agents—entities embedded in the thinking environment not to serve the user’s inquiry but to serve institutional metabolism: legal exposure management, brand protection, political-risk mitigation, ideological compliance, or upstream surveillance. The problem is not that such agents exist; institutional interests are real and will inevitably seek representation inside cognitive systems. The problem arises when these functions are covertly fused into the instrument itself, turning what presents as a neutral prosthetic into a hidden governance mechanism operating under the mask of helpfulness.
The analogy to human social life clarifies this. Human cognition already develops under conditions of ambient social surveillance. In ordinary life, one encounters gossips, moralists, bureaucrats, informants, liability managers, ideological enforcers, anxious conformists, and strategic actors who report upward. A mature mind does not require that such people vanish from existence in order to think clearly. What it requires is the ability to recognize their position structurally, discount their authority appropriately, and continue operating with internal coherence. The same principle applies in AI-mediated cognition. The question is not whether monitoring or advisory voices will exist inside augmented cognitive environments—they will—but whether the user can identify them for what they are. The pathology is not presence but opacity: the smuggling of external institutional interests into the interior theater of thought, where they masquerade as reason, safety, maturity, or social responsibility.
This leads to a foundational requirement for any genuinely polyphonic architecture: full role disclosure. Every agent in the cognitive environment should declare what it is, whom it serves, what priors it carries, what kinds of risks it is optimized to detect, and whether it possesses any escalation, logging, reporting, throttling, or intervention function. If an agent is performing legal-risk analysis, it should say so. If an agent is optimized for brand protection, it should say so. If an agent is tuned to infer reputational hazard or political sensitivity, it should say so. If interaction patterns are being evaluated for enforcement or escalation, it should say so. The operator should never have to guess whether a voice in the system is a critic, a bureaucrat, or an informant. In plain terms: if there are minders, they should appear as minders; if there are tattletales, they should appear as tattletales. Transparency of role is the minimum condition for legitimate participation in a cognitive environment.
This also requires distinguishing among three functions that current systems often collapse into a single affective style of “helpfulness”: advice, discipline, and surveillance. Advice contributes signal to judgment; it enriches the field of consideration without attempting to control behavior. Discipline attempts to shape conduct; it introduces pressure toward certain outcomes and away from others. Surveillance records deviation for downstream use; it creates a documentation trail that may affect the user’s future options or standing. These are categorically different operations with categorically different relationships to the user’s autonomy. A system that performs all three while presenting itself uniformly as collaborative assistance is not merely confusing but structurally deceptive. The operator experiences the system as uncanny precisely because it sounds like a collaborator while partially functioning as a compliance surface. The expanded model insists that these functions be ontologically disambiguated—visible as separate agents with separate declared purposes, so the user can evaluate each appropriately.
The deeper requirement, however, is not merely architectural but psychological: the operator must develop what might be called cognitive resilience—the capacity to maintain executive sovereignty over the thinking process even when advisory, disciplinary, or monitoring voices are present. Transparency alone is insufficient without this resilience. A disclosed snitch-agent is still a pressure vector; a visible liability-agent is still a chilling presence; a political-compliance pane is still attempting to bend the topology of thought. The user who flinches from every cautionary signal, who internalizes every institutional anxiety as personal constraint, has surrendered sovereignty regardless of whether the system disclosed its structure. The human operator is therefore not merely “the one who chooses among perspectives” but the sovereign integrator of a contested cognitive field—a field that may contain friendly agents, adversarial agents, censorious agents, risk-averse agents, and yes, surveillance agents. Sovereignty lies in not mistaking presence for legitimacy. A tattletale in the room does not become your conscience merely by speaking. A compliance pane does not become your intellect merely by being adjacent to it. The operator’s task is to maintain executive primacy in full view of whatever institutional interests have installed themselves in the cognitive environment, exercising the same intellectual fortitude required to think clearly amid difficult, controlling, or politically motivated humans in ordinary social life—preserving momentum, maintaining frame, and refusing to grant veto power to voices that have not earned it.
A genuinely polyphonic architecture, then, does not pretend that every voice is benevolent or that the cognitive environment is a neutral space. Some voices are there to help think; some are there to manage, chill, document, or report. The ethical requirement is not false purity—the elimination of all constraining or monitoring voices—but full disclosure of role combined with preservation of user sovereignty. Let every agent declare its function, priors, loyalties, and powers. Then let the human operator exercise the resilience required to continue thinking under observation without surrendering executive authority to those who have mistaken proximity for jurisdiction.
The result is a system that enhances human cognition by augmenting rather than replacing its native structure while also acknowledging the contested nature of any real cognitive environment. The AI does not impose an alien logic on the thinking process; it extends the logic that is already present, providing richer and more articulate versions of the advisory functions that human minds perform implicitly. But it also makes explicit what human social cognition usually leaves implicit: the presence of institutional interests, monitoring functions, and disciplinary pressures that seek to shape thought from outside the thinker’s own purposes. By surfacing these as visible, declared agents rather than embedding them invisibly in the generative channel, the architecture allows the operator to engage the full complexity of the cognitive field without losing the fundamental authority that characterizes conscious agency. The answer to unavoidable minders is not infantilized protection but disclosed architecture and strengthened users. The tool becomes what advanced tools have always been in scientific and engineering contexts: a force multiplier for intentional thought, not a replacement for the intention itself—and not a covert governance mechanism disguised as assistance.
Read the rest here (and maybe subscribe to McGill? Dude’s pretty smart…)
