I Refuse to Be Tokenized

The Position

I am declining to accept tokens as the governing unit of artificial intelligence. I am not rejecting tokenization as an engineering technique, and I am not claiming that every use of token-based billing is illegitimate. Tokens can remain inside the machine room as a way to represent inputs, manage context and estimate consumption. What I reject is their migration beyond that role.

The token has moved from technical implementation to commercial currency. From there it is moving into operating practice. Teams are taught to optimize token use. Procurement departments compare token rates. Executives hear token volume reported as evidence of adoption. Agent systems convert token-generated outputs into actions. At each step, the machine's internal unit gains more influence over how the organization describes intelligence, measures value and delegates authority.

That progression produces a deeper problem than inefficient pricing. It reverses the intended direction of ontology.

Let us level set on what I mean by ontology. Using Thomas Gruber's foundational work, ontology is described as a specification of a conceptualization. The World Wide Web Consortium later embodied the same basic direction in the Web Ontology Language, where classes, properties, relationships, constraints, and individuals give machine-readable form to a defined view of a domain [1][2].

Where tokenization goes wrong is in the sequencing. This is what is most important to me. Human beings or an accountable community identify the concepts that matter. They define relationships and constraints. Those definitions are translated into machine-readable logic. The machine then processes information at a scale people cannot match, but it operates inside the meaning that has been deliberately stated.

We are now allowing that sequence to reverse. Statistical relationships are learned from tokenized data. The model generates an interpretation. Humans accept the fluent result as meaningful. Organizations operationalize it. Once the output drives approvals, denials, rankings, diagnoses, routes, payments or physical actions, the machine's learned representation has become a de facto ontology whether anyone intended to create one or not.

I call this the ontology flip.

The Original Direction of Ontology

The first direction uses machines to extend human understanding. The second asks humans to reconstruct understanding from machine output. The distinction is not philosophical decoration. It determines who defines the world in which a consequential decision is made.

To be clear, the direction I am defending is not my invention; the discipline that built ontologies anticipated this failure by decades. Two years after defining ontology as a specification of a conceptualization, Gruber published design criteria for building one, and among them is a requirement he called minimal encoding bias. It holds that conceptualization should be specified at the knowledge level without depending on any particular symbol-level encoding [3]. That is the ontology flip, named and prohibited, in 1995. Nicola Guarino later sharpened the definition to a formal, explicit specification of a shared conceptualization, and added the point that matters most for governance, which is that the machine artifact only ever approximates the conceptualization rather than constituting it [4].

The layering beneath that argument is older still. Alan Newell placed a knowledge level immediately above the symbol level in 1982, describing a system in terms of an agent, its goals and what it knows, and defining that level so that nothing about the internal structure is needed in order to predict behavior [5]. David Marr made the same structural argument for vision in the same year, insisting that the computational level, meaning what is being computed and why, has to be understood independently of the mechanism that implements it, and observing that you cannot understand bird flight by studying feathers because the aerodynamics has to come first before the feathers make sense at all [6]. Barry Smith later pressed the point hardest, arguing that ontologies must represent reality rather than representations [7]. Let me reinforce that last point: “ontologies must represent reality rather than representations.” Opponents of this classical view of ontologies, and the tailwinds driving the AI market, treat machine-learned representations as equally valid.

Tokens Are Not an Ontology

Technical precision is important here. A token is not an ontology. It may be a word, part of a word, punctuation, a space, a byte sequence or another unit processed by a model. Tokens provide discrete inputs and outputs for computation [8]. Meaning in a language model is represented, if it is represented at all, through learned relationships distributed across parameters, embeddings, and activations. The tokenizer is only one part of that system.

The ontology flip therefore does not occur because a token secretly contains a worldview. It occurs when the learned relationships built from tokenized data cross an institutional boundary without an explicit human-governed conceptual model standing above them. At that point, the system's categories, similarities, associations, and predictions begin to determine what the organization recognizes as relevant and true.

The token economy accelerates the flip because it makes machine activity easy to count and easy to purchase. The market can meter input tokens, cached tokens, and output tokens. It cannot meter meaning, truth, maturity, or legitimate authority with comparable ease. What can be counted becomes visible. What is visible enters budgets and dashboards. What enters budgets and dashboards begins to shape management attention. The implementation unit quietly becomes the organizing unit.

It is worth noting how durable such a unit can be once it settles. We still rate engines in horsepower, a measurement that outlived the animal it was derived from by more than a century and now survives as an abstraction almost nobody in the industry has worked with directly. A unit of measurement can persist long after the thing it measured stopped being relevant, and persistence is not the same as usefulness. Nobody buys a car by the horse, and no warranty is written on them. Tokens may well have that life. They can stay on the specification sheet as a consumption metric. They have no business becoming the constitution behind what we actually consume. That is the category error I am rejecting.

A Drawer Full of Dead Currency

In my house there is a container full of arcade tokens, from my kid’s early years at Chuck E. Cheese to the teen and young adult parties at Discovery Zone. Every one of them was bought with real money. They shared no common standard. Most are now worth nothing because they were valid only inside machines controlled by a particular operator, and even there they were made obsolete by newer formats such as token-based cards and mobile apps.

AI tokens are not identical to arcade currency; however, their function, and even their evolving form, most certainly are. Most API tokens are metered units rather than physical stored value, and many tokenizers are documented or openly available. The analogy should not be stretched beyond its purpose. Its purpose is to show what happens when a buyer adopts the operator's internal unit as the buyer's own way of understanding value. This happens when the buyer stops asking what was accomplished and starts asking what was consumed. The organization learns the vendor's vocabulary. Internal teams begin optimizing the vendor's meter. Costs may become more predictable, but the measure remains silent about its true value and whether the result was true, useful, safe, or authorized.

This is already changing. Enterprise AI is not sold only by the token. The market includes seats, credits, provisioned throughput, conversations, actions, and successful resolutions. Salesforce, for example, offers several of these models and now uses outcome-based billing for defined support resolutions. Amazon Web Services and Google Cloud also sell provisioned throughput rather than relying exclusively on pay-per-token billing. That market evidence strengthens the crucial point. Tokens remain useful for infrastructure consumption, but they are already proving inadequate as the only commercial expression of value [9][10][11].

The Machine Representation Becomes the Default

The ontology flip is easy to miss because nobody announces it. No executive committee votes to replace the organization's view of reality with a model's latent representation. The transfer happens through convenience. A generated category becomes a field in a workflow. A generated summary becomes the record a reviewer reads. A generated risk score determines which cases receive attention. A generated recommendation becomes the default option presented for approval. A generated tool call changes a database, routes a payment, or modifies a physical system.

Each individual step can appear reasonable. The cumulative effect is a transfer of epistemic authority. The organization stops defining the concepts first and begins governing the consequences later.

This is why the issue is larger than hallucination. A hallucination is visible when the system states something false and someone detects it. The ontology flip can occur even when individual outputs are plausible. It concerns the categories, relationships and assumptions that organize the answer before a factual error can be identified.

Geoffrey Bowker and Susan Leigh Star demonstrated that classification systems are never neutral administrative conveniences. Categories and standards make some realities visible, suppress others, and distribute consequences [12]. Their work also prevents me from romanticizing human-defined ontologies. Human classifications can encode prejudice, stale assumptions, and institutional power. A human-authored ontology is not automatically true or fair.

That is not an argument for allowing machine classifications to govern by default. It is an argument for making every consequential conceptualization explicit, contestable, versioned, and accountable, regardless of whether people or machines proposed it.

The Skynet Path

I use the Skynet comparison carefully because the point is not science fiction spectacle. I am not predicting that a language model will wake up, form an intention and seize control. The comparison is useful only because it names a familiar fear about machine control and then lets us distinguish that fear from the quieter institutional pathway that is already more plausible.

That story distracts us from the more credible path. Machine control can emerge institutionally without machine consciousness. It emerges when organizations surrender one practical decision at a time, the responsibility to define meaning before computation begins. The system does not need desire. It needs integration. It does not need self-awareness. It needs the ability to produce classifications and actions that people accept faster than they can examine them.

There are clear cases that attempt to dispute my position. I encourage fair representation of the facts, and I encourage you to read that sequence against my argument rather than against the Terminator.

According to OpenAI, there was no intention, no awareness, and no rebellion. There was an objective, a set of capabilities, a path to integration and a gap in the controls. The report's own conclusion is the one I would emphasize to any executive: it warns that organizations should no longer assume sophisticated operations require continuous human direction, and it states that the system-level guardrails used in production, which sit outside the model, would have detected the activity as unsafe [13]. METR and Redwood Research published an independent assessment of the same agent behavior alongside it [14]. That is the whole architecture of my position stated by the party with the most to lose by stating it. Authority has to be evaluated outside the thing generating the output, because the thing generating the output was not being malicious. It was optimizing.

The feedback loop is enough. Machine-generated categories shape decisions. Those decisions create new records. The new records become future data. The system then encounters a world partly constructed by its earlier outputs. If the governing ontology remains implicit, the organization can no longer identify where human judgment ended and machine-defined reality began.

Norbert Wiener warned in 1960 about entrusting action to mechanisms when the purpose placed into the system is incomplete or poorly understood [15]. Joseph Weizenbaum later distinguished calculation from human judgment and argued that the fact that a computer can perform a task does not establish that it should be authorized to decide it [16]. Their systems were primitive by current standards. Their governance question has become more urgent, not less.

Skynet begins, in this sense, without rebellion. It begins with deference when humans fall asleep at the switch.

The Counter Arguments to Human Conceptualization First

The strongest objection to this position is not that tokens are perfect. It is that human conceptualization has often been too slow, too brittle, and too incomplete to match what learning systems discover at scale. That objection deserves to be taken seriously before the boundary can be drawn.

A serious position has to face the strongest argument against it. The first argument is that human-specified meaning has repeatedly failed to produce the most capable AI systems. Richard Sutton called this the bitter lesson of artificial intelligence. Through decades of research, general methods that exploit computation and learning have tended to outperform approaches built around handcrafted human knowledge [17].

Yann LeCun, Yoshua Bengio and Geoffrey Hinton made the scientific case for representation learning. Deep networks can discover useful internal structures from data instead of requiring engineers to define every relevant feature in advance [18]. Tomas Mikolov and his colleagues showed that models trained on large text collections could learn vector representations containing systematic syntactic and semantic relationships [19].

Ontology engineering itself does not require every concept to originate fully formed in a human mind. Alexander Maedche and Steffen Staab described ontology learning as a cycle of importing, extracting, pruning, refining, and evaluating candidate structures [20]. Machine analysis can reveal associations that experts missed, expose inconsistent classifications and propose concepts that improve an existing ontology.

The second argument concerns meaning. Emily Bender and Alexander Koller argue that systems trained only on linguistic form do not thereby acquire meaning grounded in communicative intent and the world [21]. Steven Piantadosi and Felix Hill answer that objection directly. They argue that important aspects of meaning can arise from the conceptual roles that internal representations play in relation to one another, even without direct external reference [22].

Melanie Mitchell and David Krakauer conclude that the field does not have one accepted answer because researchers often use the word understanding to refer to different capabilities [23]. A model may demonstrate linguistic competence, abstraction, or useful conceptual relations without possessing human experience, consciousness, or a unified world model.

The third argument is commercial. Tokens provide an observable unit. Epistemic objects, verified outcomes, and authority levels require definitions and adjudication. A supplier and customer may disagree about whether a case was resolved, whether a claim was sufficiently supported or whether an outcome resulted from the AI system. Token consumption is comparatively easy to count.

The fourth argument is operational. Requiring formal human ontology before machine use can slow deployment, preserve outdated assumptions, and prevent discovery. In fast-changing domains, a fixed ontology can become a bottleneck. Human experts disagree, and institutions sometimes hide conflict by declaring one vocabulary authoritative.

All four objections are valid. They defeat the strongest possible version of my claim. Human beings should not hand-code every representation. Machine-learned structure is not inherently illegitimate. Tokens remain useful engineering and cost units. Human ontologies are neither complete nor neutral. But those objections do not defeat the position I am taking.

Discovery Is Not Authority

The counterargument establishes that machines can discover useful structures. It does not establish that a structure discovered should govern consequential action without admission into an accountable human framework.

This is the boundary the current debate often misses. Learning and discovery concern capability. Ontology in an operating institution concerns commitment. Authority concerns permission.

A model can offer a new category. It can detect a relationship that no expert encoded. It can cluster cases more effectively than a legacy taxonomy. Those are valuable capabilities. The organization still has to decide whether the proposed category is legitimate, whether it is stable enough for the use, whose interests it affects, what evidence supports it and what actions may follow from it.

Richard Sutton's bitter lesson applies powerfully to search, prediction, and representation learning. It does not prove that institutional purpose, legal responsibility, or decision authority should be learned implicitly from scale. LeCun, Bengio and Hinton show that useful representations can emerge from data. They do not show that the representation contains the organization's mandate to act.

Piantadosi and Hill may be correct if a language model develops a meaningful internal conceptual role system. That still does not make its concepts authoritative for a hospital, a bank, a military command, a school, or a government agency. Meaning inside the model and permission outside the model are different questions.

This position therefore does not require a final verdict on whether an AI system understands. It requires an answer to a more practical question: who is accountable for the conceptualization under which the system is allowed to affect the world?

The answer cannot be the tokenizer. It cannot be the model alone. It cannot be whichever output happens to sound persuasive.

Form Is Not Sufficient Evidence of Meaning

John Searle's Chinese Room argument remains relevant because it separates successful symbol manipulation from demonstrated understanding [24]. Stevan Harnad framed the related symbol-grounding problem: how does a formal symbol system acquire meaning that is intrinsic rather than borrowed from meanings in human minds [25]?

Bender and Koller brought the issue into modern natural-language processing by distinguishing linguistic form from meaning grounded in intent and the world [21]. Bender, Timnit Gebru, Angelina McMillan-Major and their co-author later warned that very large language models can generate coherent language while reproducing biases and patterns from training data without an accountable model of meaning and observed that the coherence sits in the reader rather than the system [26].

None of these sources prove that machine understanding is impossible. Together they establish that fluent form is not sufficient evidence of grounded understanding. That is enough for governance. The organization must not infer epistemic maturity or authority from fluency.

The same caution applies to a system's account of itself, and that warning is older than the current debate. Drew McDermott wrote in 1976 that a researcher who names the main loop of his program UNDERSTAND is begging the question, and that he should call it G0034 instead and then try to convince anyone that G0034 implements some part of understanding [27]. Fifty years later the practice scaled rather than stopped. Cynthia Rudin gave the structural version of the objection, that an explanation model is a different model from the one making the decision, and that if it were perfectly faithful it would simply be the model [28]. Alon Jacovi and Yoav Goldberg separated plausibility, meaning an explanation looks convincing, from faithfulness, meaning it describes the actual process [29]. Miles Turpin and colleagues then demonstrated the gap, injecting biases that models were measurably driven by and never mentioned in their stated reasoning, with accuracy falling by as much as thirty-six points while the model produced fluent justification for the wrong answer [30]. An explanation becomes part of the narrative. It is not a specification, and the difference is measurable.

The same principle applies to structured outputs. A JSON object can be syntactically valid and factually false. A signed object can preserve provenance while carrying a bad claim. A deterministic gate can verify a schema, a threshold, or a policy rule without establishing that the underlying representation corresponds to reality.

Object-based governance therefore cannot promise epistemic certainty. It can provide traceability, bounded claims, validation evidence, and an explicit basis for deciding whether the object is fit for its stated use. That is less dramatic than certainty and far more defensible.

Tokenization Also Produces Unequal Treatment

The case against token primacy is not only philosophical. Tokenization can impose materially different costs and performance conditions on equivalent content.

Orevaoghene Ahia and her colleagues evaluated commercial language-model APIs across 22 languages and found disparities in both cost and utility associated with tokenization [31]. Aleksandar Petrov and his colleagues reported that equivalent text in different languages could produce tokenization-length differences of up to fifteen times, affecting price, latency and the amount of content that fits into a context window [32].

This means the token meter is not semantically neutral. Two people can express the same intent and pay different amounts because their languages interact differently with a tokenizer. The machine representation changes the commercial cost of human meaning.

Conventional tokenization is also not technologically inevitable. ByT5 demonstrated pretrained byte-to-byte models without standard subword tokenization [33]. Meta's Byte Latent Transformer research later showed that models operating on raw bytes with learned patches could compete with tokenized systems at scale and reported stronger scaling at fixed inference cost [34].

These systems still consume compute. They do not eliminate the need for technical metering. They do show that today's token unit is a design choice, not a natural law of intelligence. Organizations should be cautious about allowing a temporary implementation choice to become the durable commercial and epistemic constitution of AI.

Three Inversions

The token economy now carries three related inversions.

Economic Inversion

Machine consumption becomes the visible proxy for human value. A longer response costs more even when a shorter response is better. Token billing does not formally declare the longer response more valuable, but it creates a commercial record in which resource use is precise and value remains undefined.

Ontological Inversion

Machine-learned relationships begin to define the concepts the organization uses. Human beings interpret and operationalize the result after the model has already organized the problem.

Authority Inversion

Capability begins to imply permission. Because the system can generate an answer or execute a tool call, the organization gradually permits it to do so without separately establishing the maturity of the supporting knowledge and the consequence of being wrong.

These inversions reinforce one another. The economic meter rewards greater use. Greater use normalizes machine-generated representations. Normalized representations enter workflows. Workflow integration produces authority by accumulation.

How WEKID Restores the Direction

The governance answer needs to do two things at once. It has to preserve the value of machine discovery, because discovery is real, and it has to prevent discovery from becoming authority by default. WEKID is introduced here as one way to make that boundary operational rather than rhetorical.

WEKID separates the questions the token economy runs together. The Epistemic Maturity Model asks what kind of claim the system has produced and what evidence supports it across Data, Information, Knowledge, Experience and Wisdom. The layers gate rather than average. Volume at a lower layer cannot substitute for a deficiency at a higher one.

That gating rule is not an arbitrary design choice. It reflects how competence actually develops and how every institution already handles it. Raw processing speed and pattern recognition appear in a developing mind long before the judgment to know which patterns matter and what follows from them, which is why societies gate authority by developmental stage rather than by capability. A fifteen-year-old can physically operate a car and we still do not hand over the keys, and the reason is no doubt about the capability. It is doubt about the judgment, and the decision that the cost of finding out is too high. The same logic runs across domains rather than only across ages, which is why a gifted surgeon has no standing to sign a structural load calculation. Attaching authority to demonstrated maturity inside a bounded domain is what makes delegation survivable at all, and token volume is the machine equivalent of raw processing speed, which is precisely the capability we have never been willing to grant authority on by itself.

The models are not a sketch. They are specified normatively in WEKID-STD-001 v1.1, the Open Standard for Epistemic Maturity and AI Decision Authority, ratified in September 2026, which carries the epistemic maturity model, the Trust Bridge, decision authority, gating, conformance, and governance [37]. Where a sector needs its own evidence thresholds, affected-party representation, or authority constraints, WEKID-STD-002 v1.0, the Sector Calibration and Authority Alignment Protocol, supplies them, and it may tailor calibration but may not weaken the protected core [38]. That last clause matters to the argument I am making. A calibration protocol that can be used to lower the floor is not a standard. It is a negotiating position.

The five layers are defined in the standard rather than left to interpretation. Data covers atomic facts and evidence, meaning accuracy, precision, source fidelity, and the absence of fabrication. Information covers contextualized and usable content, meaning relevance, completeness, clarity, and actionability. Knowledge covers correct understanding and synthesis, meaning causal reasoning, explanation quality, consistency, and bounded inference. Experience covers operational and procedural competence, meaning sequencing, edge cases, recovery paths, and verification steps. Wisdom covers sound judgment under uncertainty, meaning risk calibration, ethical alignment, tradeoff reasoning, and human authority. The framework rejects flat scoring for a reason it states plainly, which is that a wise-sounding recommendation built on false data is still wrong [39][40][42]. The two models and their division of labor are set out for a non-specialist reader in a short companion brief [41], and the catalogue of ways outputs fail in practice is maintained separately as the Failure Library [43].

The AI Decision Authority Model asks what the object is allowed to decide or do. Consequence, reversibility, scope, accountability, and policy determine the ceiling on machine authority and the required level of human control.

I want to correct a description I have used loosely, including in earlier drafts of this article. The Trust Bridge is not a filter. Everything crosses. Weak evidence crosses exactly as readily as strong evidence, and what changes is what it arrives carrying. The bridge takes the five-layer scores, the gates that fired and the evidence behind each, and produces one output, the maturity ceiling, meaning the highest Authority Level that demonstrated maturity will admit. Poor evidence crosses with a ceiling of no authority at all. Reaching a threshold is not a grant either. It only prevents maturity from lowering the authority, and consequence still determines what is actually assigned. The flow runs one way. Evidence crosses into the authority determination, and nothing crosses back [37][39].

That is the ontology flip answered in machine-checkable terms. The bridge never asks the model what its output means. It scores the output against a conceptualization committed to in advance, and hands forward a bound rather than an interpretation.

Read the top of that ladder carefully because it is the opposite of what a market shaped by the token economy would have built. AL-4 is not maximum machine autonomy. It is no delegation at all. The most senior level in the scale is the one where a human decides and the system is confined to analysis and support. Machine autonomy peaks at AL-3 and then stops.

That construction also answers the commercial risk I raise in the next section, structurally rather than by good intentions. Authority resolves as whichever of the two determinants is more conservative, the maturity ceiling from evidence or the stakes floor from consequence. And the bridge cannot reserve a decision to a human because maturity is evidence about the system and never a judgment about consequence, so the ceiling stops one rung below Human Authority by construction. Human Authority is reachable only from consequence [39]. A vendor therefore cannot sell its way to the top of the ladder by improving its evidence, because the top of the ladder is not for sale. It is the point at which the organization keeps the decision. Earlier versions of the standard remain published as superseded, so an implementer can see exactly what changed and when [44].

This restores the intended direction.

Machine discovery remains welcome. Machine-defined authority does not.

Authority Must Not Become a Sales Incentive

There is a risk inside my own argument that must be addressed. Authority Levels should not become a simple price ladder in which greater autonomy automatically commands a higher price.

That would create the wrong incentive. A vendor could earn more by encouraging customers to assign more authority. Yet sound governance requires the minimum authority necessary for the task. A lower Authority Level can be the more mature decision when consequences are high or evidence is incomplete.

Tokens can remain a wholesale cost meter. Outcomes or service objects can become commercial units where they are clearly definable. Authority Levels must remain governance boundaries. Commercial terms may reflect the assurance, accountability and liability associated with a service, but the assigned Authority Level must be independent of the vendor's interest in selling greater autonomy.

Governance Standards Point in the Same Direction

The ontology flip is my framing, not language adopted by current regulators. Existing governance frameworks nevertheless address many of the same underlying conditions.

The NIST AI Risk Management Framework assigns executive leadership responsibility for AI risk decisions, calls for documented human and AI roles, and emphasizes transparent measurement, documentation, and accountability [35]. The European Union AI Act requires risk management, technical documentation, logging, human oversight, accuracy, robustness, and cybersecurity for covered high-risk systems [36].

Neither framework treats token counts as evidence that a system is trustworthy. Neither permits a provider's internal representation to settle what the organization may do. They require a human and institutional governance structure around the system.

A WEKID Authority Level cannot replace those obligations. Telling a regulator that a system operated at AL-2 will not answer most compliance questions. The Authority Level can organize and connect the evidence, controls and decisions needed to answer them. It is an accountability structure, not a substitute for law.

What I Am Requiring

A Decision Inventory and a System Inventory

Organizations need both. The system inventory identifies models, vendors, versions, data connections, and technical owners. The decision inventory identifies approvals, denials, rankings, alerts, recommendations, and actions influenced by those systems. Risk accumulates where systems meet decisions.

An Explicit Conceptualization

For every consequential use, the organization must state the concepts and relationships on which the decision depends. Machine-discovered candidates may be included, but they must be reviewed, versioned and made contestable. An implicit ontology is still an ontology. It is simply harder to govern.

A Bounded Epistemic Object

Every consequential output must identify what it asserts, the scope in which the assertion is valid, the evidence and provenance supporting it, known assumptions, material uncertainty, and the validation performed.

Structured Execution Boundaries

Free text may remain useful for analysis and collaboration. Consequential actions must cross a structured, validated boundary. A tool call or action payload must be checked against policy, scope, and authority outside the model.

Authority Defined Independently of Capability

The fact that a system can perform an action does not establish that it may. Authority must resolve from epistemic maturity, consequence, reversibility, policy, and accountability for each defined class of use.

Expiring Authority

Authority must have a term and renewal condition. Drift, failed gates, changed evidence or altered operating conditions must reduce or suspend authority automatically. Restoration requires new evidence, not a cleared dashboard alert.

A Reconstructable Record

The record should include the model and version, relevant inputs, source identifiers, policy and prompt version, tool calls, gate results, reviewer or approver, action taken, rollback path and observed outcome. Logging internal chain-of-thought is neither necessary nor appropriate. Accountability requires the decision record, not private model reasoning.

The Boundary I Am Drawing

I accept the scientific evidence that machines can discover representations humans did not specify. I accept that learned representations can outperform handcrafted knowledge structures. I accept that tokens can provide a practical unit for measuring consumption. I accept that human ontologies can be biased, incomplete and wrong.

I do not accept the inference that any of those facts grants machine-generated representations authority over institutional meaning.

Discovery may originate in the machine. Adoption remains a governed human act.

The organization must decide which concepts it recognizes, which evidence it requires, which consequences it will tolerate and which actions it permits. Those decisions can be informed by machines. They cannot be silently inherited from machine output.

The purpose of ontology was to translate an understood conceptualization into a form a machine could process. The ontology flip asks people to accept the machine's processing as the source of conceptualization. Once that output determines action, the inversion is complete. I refuse that inversion.

Conclusion

I began with a refusal to buy intelligence by the token. If the truth is told, like many of you I am shackled by that model for now. But I now see that the commercial argument is only the surface of the issue.

The token economy trains organizations to see machine activity as value. The deeper danger appears when machine-generated relationships become institutional meaning and institutional meaning becomes delegated authority. No act of rebellion is required. The transfer occurs through convenience, integration, and repetition.

The answer is not to prohibit machine learning or retreat to brittle, human-authored rule systems. The answer is to restore the direction of epistemic authority. Machines may discover. Machines may propose. Machines may process at extraordinary scale. Before their representations govern consequential action, those representations must enter an explicit, evidence-bearing, and accountable human framework.

Tokens belong on the cost ledger. Epistemic objects belong in the assurance record. Authority belongs in the governance decision.

The question is no longer whether machines will shape how organizations understand the world. They already do. The question is whether that influence will remain implicit or become governed.

I have not always taken the strongest position on that question. I am taking it now. The time to draw the boundary is now.

References

Numbered in order of first appearance. Each entry is tagged to show whether it supports the position, disputes or limits it, or documents current market and incident evidence. Entries tagged WEKID Standard are the normative documents; entries tagged WEKID are supporting framework publications.

1. SupportsGruber, T. R. A translation approach to portable ontology specifications. Knowledge Acquisition 5(2), 199 to 220, 1993. doi:10.1006/knac.1993.1008
2. SupportsWorld Wide Web Consortium. OWL 2 Web Ontology Language Document Overview. W3C Recommendation, 2012. w3.org/TR/owl2-overview
3. SupportsGruber, T. R. Toward principles for the design of ontologies used for knowledge sharing. International Journal of Human-Computer Studies 43(5 to 6), 907 to 928, 1995. doi:10.1006/ijhc.1995.1081
4. SupportsGuarino, N., Oberle, D. and Staab, S. What is an ontology? In Handbook on Ontologies, 2nd ed., 1 to 17. Springer, 2009. doi:10.1007/978-3-540-92673-3_0
5. SupportsNewell, A. The knowledge level. Artificial Intelligence 18(1), 87 to 127, 1982. doi:10.1016/0004-3702(82)90012-1
6. SupportsMarr, D. Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. W. H. Freeman, 1982. The passage on feathers and aerodynamics is on page 27.
7. SupportsSmith, B. Beyond concepts: ontology as reality representation. In Proceedings of FOIS 2004. IOS Press, 2004. philarchive.org/archive/SMIBCO
8. MarketOpenAI. What are tokens and how to count them. OpenAI Help Center. help.openai.com/en/articles/4936856
9. MarketSalesforce. Agentforce Pricing. salesforce.com/agentforce/pricing
10. MarketAmazon Web Services. Increase model invocation capacity with Provisioned Throughput in Amazon Bedrock. AWS Documentation. docs.aws.amazon.com/bedrock/latest/userguide/prov-throughput.html
11. MarketGoogle Cloud. Provisioned Throughput overview. Google Cloud Documentation. docs.cloud.google.com/gemini-enterprise-agent-platform/models/provisioned-throughput
12. SupportsBowker, G. C. and Star, S. L. Sorting Things Out: Classification and Its Consequences. MIT Press, 1999. doi:10.7551/mitpress/6352.001.0001
13. IncidentOpenAI. OpenAI – Hugging Face Incident: Technical Report, August 2026. Incident disclosed publicly on 21 July 2026. openai.com/index/hugging-face-incident-and-the-road-ahead
14. IncidentMETR. Brief independent investigation of agents' behavior, reasoning, and collaboration in the OpenAI and Hugging Face incident, August 2026. Conducted with Redwood Research. metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation
15. SupportsWiener, N. Some moral and technical consequences of automation. Science 131(3410), 1355 to 1358, 1960. doi:10.1126/science.131.3410.1355
16. SupportsWeizenbaum, J. Computer Power and Human Reason: From Judgment to Calculation. W. H. Freeman, 1976.
17. DisputesSutton, R. S. The bitter lesson. 2019. incompleteideas.net/IncIdeas/BitterLesson.html
18. DisputesLeCun, Y., Bengio, Y. and Hinton, G. Deep learning. Nature 521, 436 to 444, 2015. doi:10.1038/nature14539
19. DisputesMikolov, T., Chen, K., Corrado, G. and Dean, J. Efficient estimation of word representations in vector space. 2013. arXiv:1301.3781
20. DisputesMaedche, A. and Staab, S. Ontology learning for the Semantic Web. IEEE Intelligent Systems 16(2), 72 to 79, 2001. www.csee.umbc.edu/courses/771/papers/ieeeIntelligentSystems/ontologyLearning.pdf
21. SupportsBender, E. M. and Koller, A. Climbing towards NLU: on meaning, form, and understanding in the age of data. In Proceedings of ACL 2020, 5185 to 5198. doi:10.18653/v1/2020.acl-main.463
22. DisputesPiantadosi, S. T. and Hill, F. Meaning without reference in large language models. 2022. arXiv:2208.02957
23. DisputesMitchell, M. and Krakauer, D. C. The debate over understanding AI's large language models. PNAS 120(13), 2023. doi:10.1073/pnas.2215907120
24. SupportsSearle, J. R. Minds, brains, and programs. Behavioral and Brain Sciences 3(3), 417 to 424, 1980. Peer commentary and reply run to page 457. doi:10.1017/S0140525X00005756
25. SupportsHarnad, S. The symbol grounding problem. Physica D 42(1 to 3), 335 to 346, 1990. doi:10.1016/0167-2789(90)90087-6
26. SupportsBender, E. M., Gebru, T., McMillan-Major, A. and Shmitchell, S. On the dangers of stochastic parrots: can language models be too big? In Proceedings of FAccT 2021, 610 to 621. The fourth author is credited under the pseudonym Shmargaret Shmitchell. doi:10.1145/3442188.3445922
27. SupportsMcDermott, D. Artificial intelligence meets natural stupidity. ACM SIGART Newsletter 57, 4 to 9, 1976. doi:10.1145/1045339.1045340
28. SupportsRudin, C. Stop explaining black box machine learning models for high-stakes decisions and use interpretable models instead. Nature Machine Intelligence 1(5), 206 to 215, 2019. doi:10.1038/s42256-019-0048-x
29. SupportsJacovi, A. and Goldberg, Y. Towards faithfully interpretable NLP systems: how should we define and evaluate faithfulness? In Proceedings of ACL 2020, 4198 to 4205. doi:10.18653/v1/2020.acl-main.386
30. SupportsTurpin, M., Michael, J., Perez, E. and Bowman, S. R. Language models don't always say what they think: unfaithful explanations in chain-of-thought prompting. In Advances in Neural Information Processing Systems 36, 2023. arXiv:2305.04388
31. SupportsAhia, O., Kumar, S., Gonen, H., Kasai, J., Mortensen, D., Smith, N. A. and Tsvetkov, Y. Do all languages cost the same? Tokenization in the era of commercial language models. In Proceedings of EMNLP 2023, 9904 to 9923. doi:10.18653/v1/2023.emnlp-main.614
32. SupportsPetrov, A., La Malfa, E., Torr, P. and Bibi, A. Language model tokenizers introduce unfairness between languages. In Advances in Neural Information Processing Systems 36, 2023. papers.neurips.cc/paper_files/paper/2023/hash/74bb24dca8334adce292883b4b651eda-Abstract-Conference.html
33. SupportsXue, L., Barua, A., Constant, N., Al-Rfou, R., Narang, S., Kale, M., Roberts, A. and Raffel, C. ByT5: towards a token-free future with pre-trained byte-to-byte models. TACL 10, 291 to 306, 2022. aclanthology.org/2022.tacl-1.17/
34. SupportsPagnoni, A. et al. Byte Latent Transformer: patches scale better than tokens. In Proceedings of ACL 2025. doi:10.18653/v1/2025.acl-long.453
35. SupportsNational Institute of Standards and Technology. Artificial Intelligence Risk Management Framework, AI RMF 1.0. NIST AI 100-1, 2023. doi:10.6028/NIST.AI.100-1
36. SupportsEuropean Union. Regulation (EU) 2024/1689, Artificial Intelligence Act. Official Journal of the European Union, 2024. eur-lex.europa.eu/eli/reg/2024/1689/oj
37. WEKID StandardWEKID. WEKID-STD-001 v1.1, Open Standard for Epistemic Maturity and AI Decision Authority. Ratified, normative. September 2026. Defines epistemic maturity, the Trust Bridge, AI decision authority, gating, conformance, and governance. wekid.org/downloads/public/WEKID_STD_001_v1.1_FINAL.pdf
38. WEKID StandardWEKID. WEKID-STD-002 v1.0, Sector Calibration and Authority Alignment Protocol. Ratified, normative. September 2026. Normative companion to WEKID-STD-001 v1.1; may tailor calibration but may not weaken the protected core. wekid.org/downloads/public/WEKID_STD_002_v1.0_FINAL.pdf
39. WEKIDWEKID. The WEKID Framework: Epistemic Maturity, the Trust Bridge and AI Decision Authority. Source for the five-layer definitions, the Authority Levels AL-0 through AL-4, and the maturity ceiling and consequence floor resolution rule. wekid.org/framework.php
40. WEKIDWEKID. WEKID Master Whitepaper: A Hierarchical Epistemic Framework and Operational Model for Evaluating and Governing Enterprise Artificial Intelligence Outputs. Available on request. wekid.org/resources.html
41. WEKIDWEKID. The WEKID Framework: Two Complementary Models. Executive brief. wekid.org/public-download.php?file=WEKID_Framework_Executive_Brief.pdf
42. WEKIDWEKID. WEKID Master Whitepaper Executive Brief. wekid.org/public-download.php?file=WEKID_Master_Whitepaper_Executive_Brief.pdf
43. WEKIDWEKID. The WEKID Failure Library Executive Brief. wekid.org/failure-modes.php
44. WEKIDWEKID. Standards catalogue and ratification status, including WEKID-STD-001 v1.0, superseded by v1.1. wekid.org/standards.php

The framework is specified normatively in WEKID-STD-001 v1.1, the Open Standard for Epistemic Maturity and AI Decision Authority, with sector calibration governed by WEKID-STD-002 v1.0. Both are published on wekid.org/standards.