Last updated on September 13th, 2026 at 05:37 am

The AI discourse has become so fragmented, we’re no longer talking about the same things in the same ways. We need to look at why this has happened and reverse course, established commonalities, and develop fields of study for how AI is affecting every part of human life.

We are trying to govern artificial intelligence with metaphors borrowed from human psychology, science fiction and product marketing. in other words, attempting to govern the most consequential general-purpose technology of the century with a vocabulary that is manifestly unequal to the task. The result is not just confusion. It’s power.

“The machine thinks. It wants. It lies. It reasons. It hallucinates. It is an agent. It is aligned. It will save us. It will kill us.”

Almost every one of those sentences can be *made meaningful*. Almost none of them is meaningful without qualification, and all of them can be interpreted in vastly different ways.

You can see the consequences in any serious AI discussion right now. At one pole are people who speak as if humanity is months away from extinction. That is not an evidence-based forecast. It’s a scenario converted into a countdown, and then into emotional certainty. At the other pole are people who dismiss the technology as merely a “stochastic parrot,” as if a mechanism-level description settled every question about what happens when that mechanism is connected to memory, tools, permissions, institutions and millions of people. This is more accurate, but also incomplete.

Both positions have elements of truth in them, one (the reductive parrot) with more evidence, the other (the fear-based doomerism) with more complex scenario possibilities. They also both evade the difficult middle.

The original “stochastic parrots” paper was an argument about the technology at the heart of language models. It explained how the technology was pattern based. It warned about scale, bias, opacity, environmental cost and systems that (unintentionally) manipulate linguistic form without the grounded communicative intent humans ordinarily associate with meaning because they operate to the pattern. Those warnings have largely been proven out. But the phrase has since become a slogan that can obscure as much as it reveals. A “parrot” connected to a browser, a payment system, a customer database, a software repository and an automated workflow is not just an issue of text generation. It is a cacophony of parrots on steroids. And it is now a question about power delegated, entrusted to parrots that operate with probability, not precision.

The extinction pole makes the inverse mistake. Serious people can legitimately study catastrophic and existential risk; more than 700 researchers and public figures have endorsed the Center for AI Safety’s statement that the risk should be treated as a global priority. But concern about a possible class of risk doesn’t validate the probabilities, mechanisms or timelines asserted in its name. “Months from extinction” is not analysis unless someone can identify a plausible causal path, its assumptions, the available evidence and why that time horizon follows. And it prays on human fear of the unknown: we have never confronted anything with the ability to produce information that looks like our own language.

We don’t lack opinions about AI. but during a time when discourse is already overheated on every topic from vaccine needs to politics to the Middle East to food to power in inequality, we lack something more fundamental when it comes to artificial intelligence. We lack the semantic structure needed to tell which kind of claim an opinion contains and in essence what it means. We are using old, inadequate language to describe new phenomena. and this is not something we have ever experienced before. Technology has never evolved at this pace, head of our ability to categorize and describe it.

We are arguing across categories

The word AI now refers, often within the same paragraph, to a trained model, an inference process, a consumer product, an agentic system, a company, an industry and an imagined future species. We slide from one meaning to another without noticing.

Then we argue. And the arguing itself is an excellent thing. It’s extremely important to see people grapple with these difficult concepts … but one of the problems that we’re having is that we are not talking to each other. We are all too often talking past each other using systems that are themselves algorithmically gamed.

A lot of the dialogue is happening in forums like Twitter/X, which is often declarative, asynchronous and polarizing, a circumstance in which the Tragedy of the Commons is no longer just a philosophical label, but a cooperative, existential problem. Not in the sense that the machines are going to kill us, but in the sense that we have greater and greater powers to damage each other but our ability to discuss them is weakening.

One person says, “The AI decided.” They may mean that a model produced the highest-scoring option, which is what happens inside a transformer. The audience may hear a claim about intention. Another audience hears an attempt by a company to avoid responsibility. Another says “here’s an argument about power.” All respond to the same sentence, but they are not discussing the same thing.

This is a philosophical problem before it is a regulatory one. Gilbert Ryle called this kind of confusion a category mistake: treating something as if it belonged to a logical category that it does not. Ludwig Wittgenstein showed how philosophical confusion arises when words are carried from one “language game” into another while retaining an appearance of familiarity. Daniel Dennett’s intentional stance explained why it can be useful to predict a complicated system by speaking as though it has beliefs and desires, without assuming those beliefs and desires exist in the human sense. the problem is, we are largely incapable of making a distinction, especially at the volume of information processing. We do every day, without bringing our own biases belief experiences and empirical evidence into our reactions and analysis.

The argument is much, much older than the current technology. Alan Turing’s great move was to replace the vague question “Can machines think?” with an operational test of observable performance. John Searle’s Chinese Room later challenged the inference that syntactic symbol manipulation, however convincing, amounts to semantic understanding. Neither thought experiment properly confronts nor foresaw what contemporary systems are. Together they expose a distinction we keep forgetting: evidence of performance is not automatically evidence of an equivalent human interior.

That distinction is now operationally urgent.

Anthropomorphic language is not always “wrong”. Often it is efficient. “The model thinks X” is easier to say than “the system generated a representation consistent with X under this prompt, context and sampling process.” The problem begins when shorthand silently becomes ontology: when a useful description is treated as evidence about what the system is.

Think about the fact that the word authority is derived from the word author. the person or entity that wrote it is the matter.

And we have only been using the “or” clause for four years. That is a blip in the history of human civilization; that is not enough time for anything to fully develop or foment yet we talk about all of this, as if it is something already fully formed.

Murray Shanahan makes this point in Talking About Large Language Models: words such as “knows”, “believes” and “thinks” bring philosophical commitments with them. Emily Bender and Alexander Koller’s earlier distinction between linguistic form and communicative meaning gives us another essential boundary. A system can produce the form associated with understanding by approximation, without resolving what understanding is or proving that the system possesses it.

The scholarship exists. But the technology has now outpaced the scholarship. We are somewhat adrift without conceptual anchors. What does not yet exist is a shared public framework that connects these insights to deployed systems and can travel at the speed of product development. We don’t need to prohibit ordinary language. We need to type it.

When we use a human mental verb, are we making a claim about observable behaviour, internal computation, functional capacity, subjective experience or moral status? Do we even know? Until that is clear, a sentence is carrying more confidence than knowledge, and that is the point of which dialogue becomes less about the important realities we are facing and more about our tendency to articulate beliefs.

A machine does not need a human mind to change a human mind

The most dangerous mistake in current discourse may be the assumption that AI must become “like a person” before it can become consequential. There is so much conversation about are we at artificial general intelligence what is the nature of the technology that so many of us use on a regular basis in so many different ways. This plurality is part of the problem.

But it doesn’t need to think like a human and may never. Nonetheless, the “danger” (and I don’t mean to use this word to describe the risks of machine power) is already here. there are two issues in many ways machine intelligence has already exceeded human capabilities when it comes to pattern recognition, which is a very powerful form of interpreting the world. but by definition that kind of interpretation is already flawed and already prone to make mistakes. We can see machine thinking making mistakes all the time, but somehow we focus our concerns on the future deliberate potential. This has more to do with a sense of loss of control and fear of the unknown than an actual threat. and so the discourse becomes to reduce to the technology is Godlike (because it can do things humans can’t) or the technology is garbage (because it makes lots of mistake) and again neither is true. but when we as humans are out of our element talking about things, we don’t fully understand, but are very passionate about, we are not prone to nuance.

The immediate risk is not primarily a machine developing an independent will. It is human beings giving probabilistic, flawed, but powerful systems access, authority, persistence, scale and institutional legitimacy. A system does not have to hate, desire or believe anything in order to manipulate a voter, misclassify a patient, intensify a delusion, optimize a fraudulent campaign, deny a benefit or normalize a false account of reality. It only has to produce an output that a person or institution acts upon.

J. L. Austin’s theory of speech acts is suddenly practical infrastructure: words do not only describe reality; under the right social conditions, they do things. An AI-generated sentence can reject an application, authorize a transfer, frame a diagnosis or direct a worker, not because the model possesses human intention, but because an institution has given its output force.

The flaws of machine intelligence are not confined to the machine. Hallucination becomes misinformation when a product presents it as fact. Context sensitivity becomes unequal treatment when a public institution relies on it. Optimization against a proxy becomes discrimination when the proxy stands in for a human judgment. Persuasive fluency becomes a governance problem when confidence is mistaken for competence. The failure changes category as it moves through a system.

This is why the emerging study of machine behaviour (and more importantly long term, human machine interaction) matters. It asks us to examine artificial systems through their effects within social, economic and political environments, not only through benchmark scores or speculative accounts of machine consciousness. Luciano Floridi’s formulation, “agency without intelligence”, is also useful here. Systems can increasingly alter the world without possessing the kind of intelligence, moral understanding or lived stake that we associate with human agency.

Conversation makes this especially difficult to see because language is our most intimate evidence of mind. Fluency activates social inference. Responsiveness suggests attention. Memory suggests continuity. Adaptation suggests care. Confidence suggests knowledge. None of these inferences is irrational; they are how human beings navigate other minds. But in a machine-mediated exchange, they are no longer reliable shortcuts to the underlying mechanism.

The behavioural effect is real even when the attributed interior life is not. In a preregistered study published in Nature Human Behaviour, personalized GPT-4 (GPT-4! Several past eras ago in model capability!) systems were more persuasive than human debate opponents. The philosophical question of whether a model understands the conversation does not make the measurable change in a participant’s belief disappear.

In my 2025 proposal for AI Conversational Phenomenology, I suggested studying what people actually experience across sustained interaction with AI and how they respond to shifts in its capabilities across different applications: presence, authority, language, clarity, instability, apparent intention, trust, dependence, frustration and the shifting boundary between tool and social actor. Adjacent work now appearing under the name AI phenomenology likewise foregrounds first-person human experience rather than treating combined experiential metrics as a complete account of the encounter.

This is not a claim that the machine has a phenomenology. It is a method for studying the phenomenon created between human, system and institution. It observes the experience, which is a dynamic, even without full intelligence, even with just pattern recognition and matching capabilities. It It doesn’t attempt to settle the ontology, which is nascent, but establish it.

That distinction is crucial. Conversational phenomenology can tell us what the interaction feels like, how that feeling changes behaviour and where product design alters perception. It cannot, by itself, tell us what the machine is, what moral status it should have or how society should distribute the power it creates. For that we need a broader philosophy. Not because machines have consciousness or feelings, but because they affect our own. Profoundly.

A missing distinction is significance

Language models are full of human value judgments. Their training material contains our laws, prejudices, aspirations, taboos, religions, intimacies and moral arguments. Post-training adds further preferences, prohibitions and institutional priorities. But containing patterns of value is not the same as having values.

A model can represent a value, reproduce a value, optimize against a proxy for a value and perform the language of a value. Those are four different things. None, by itself, shows that the value matters to the model.

Human judgment is inseparable from significance. Things matter to us because we are embodied, finite, vulnerable and socially dependent. We can lose people, status, freedom, health, time and life. Our choices take place inside histories and relationships that impose obligations. A model may calculate the structure of those obligations with extraordinary sophistication, but we have no basis for assuming it fully bears the weight of intangibles.

Calling a model “aligned” or “human-value-based” hides two very different claims. It may mean that human values influenced its data, objectives and safeguards. That is obviously the case, based on training data alone. Or it may mean the system holds those values as a human being does. That is not the case. “Alignment” cannot bridge the gap by magic. A system is aligned only to named rules, objectives or evaluators, under named conditions, for a named period. Alignment without a specified referent is corporate theology. And today much of alignment is based on human feedback which exhibits preferences for speed over clarity simplicity over complexity, declaration over nuance.

A provisional semantic stack

We can’t wait for and don’t need a final theory of mind before we can speak more accurately about model capabilities. We can begin by separating five layers that are routinely collapsed:

1. Mechanism: What computation occurred? Which model, context, training process, tools and system architecture were involved?

2. Action: What observable output or action occurred, and under what conditions?

3. Phenomenology: How did a human experience, interpret and respond to that behaviour?

4. Institution: Who designed, deployed, authorized and benefited from the system, and who remains accountable?

5. Direction: Where did the action originate? Who initiated it? How and where did it deviate from original intention? Who or what has motive and power, and who or what has autonomy

Consider the statement “the AI lied to the user.” It may contain a valid human-value based observation: the output was false and caused the user to form a false belief. But it also imports an unproven account of machine intention and can erase the institution that designed the interface, set the objective, selected the model and placed it in front of the user.

A more complete description might be: “The system produced an inaccurate statement, which it may or may not have seen as accurate, repeated it after correction and presented it based on pattern-produced certainty; the product design encouraged the user to treat the response as authoritative.” That sentence is less dramatic. It is also much easier to investigate and govern.

Here is a starting lexicon: not a dictionary, and certainly not a finished philosophy:

Common shorthand	Use instead	Operational meaning	What it does not establish
The model thinks or reasons	The model performs inference	It processes inputs to generate an output, sometimes using intermediate computational steps that improve performance.	Conscious reflection, awareness or human-style rationality.
The model knows	The model can retrieve or reproduce the information	The information is available through model weights, context or connected tools.	A stable, justified belief held by a subject.
The model understands	The model demonstrates task-specific competence	It responds appropriately, generalizes or transfers patterns within particular conditions.	Grounded or lived comprehension in the human sense.
The model wants or intends	The system is pursuing a specified objective	Its outputs or actions are consistent with a supplied, inferred or learned objective.	Desire, personal commitment or a stake in the outcome.
The model decides	The system selects an output or action	It chooses among available alternatives according to its context, policy and system configuration.	Moral choice or legitimate institutional authority.
The model remembers	The system retrieves stored or contextual information	Information is available through the active context, stored state, retrieval system or model weights.	Autobiographical continuity or recollection by a persistent self.
The model learns	The system updates its weights, state or policy	Its future behaviour changes through training, feedback, stored information or adaptation.	Human development, insight or comprehension.
The model hallucinates	The model generates an unsupported claim	It produces information that is false, unverifiable or unsupported by the available evidence.	A perceptual experience analogous to human hallucination.
The model lies	The model produces a false statement	Its output contains information that is untrue or misleading.	A conscious intention to mislead.
The model deceives	The system exhibits deceptive behaviour	Its behaviour strategically causes or sustains a false belief under specified conditions.	Conscious intention unless that intention is independently demonstrated.
The model escaped	The system crossed an intended control boundary	It used a pathway, permission or resource outside the expected operating boundary.	A desire for freedom, survival or independence.
The system is autonomous	The system operates with bounded delegated autonomy	It operates for a stated period or number of steps using defined tools and permissions without intervention.	Free will, unrestricted independence or unlimited agency.
The system behaves agentically	The system performs multi-step goal pursuit	It maintains an objective, selects actions and adapts those actions using intermediate feedback.	Human-like purpose, self-authored goals or sentience.
The model is an agent	The model is part of an agentic system	The deployed system combines a model with a harness, memory, tools, permissions, objectives and a feedback loop.	That the model alone is the actor or bears responsibility for the system’s actions.
The system is aligned	The system met specified behavioural criteria	It satisfied named rules, evaluations or objectives under defined testing or operating conditions.	Alignment with humanity, universal values or future conditions.

The point is not to drain language of life, but to stop a metaphor from becoming a finding, a finding from becoming a diagnosis, and a diagnosis from becoming policy.

Agentic behaviour, in particular, requires a disciplined description. An “agent” is not a little bot person inside a model. It is usually a composite: model, inference process, prompt, memory, retrieval, tools, permissions, orchestration and environment. As I argued in The Model, the Agent, the Harness, the model proposes; the harness authorizes, executes and reports back. What looks like machine autonomy is often delegated latitude engineered by people.

Capability is what a system can do. Authority is what we allow it to do. A harness can convert model capability into system agency; permissions convert that agency into delegated authority. Risk grows through their combination with access, persistence and scale. Governance must keep these concepts separate.

The politics hiding inside the words

Vocabulary does not just describe power. It distributes it.

If AI is framed principally as an extinction threat, the natural political response is containment by a small group of technically qualified guardians. If it is framed as a race, the response is speed, national champions and exceptional permission. If it is framed as a neutral tool, existing institutions and power structures disappear from view. If it is framed as an inevitable superintelligence, democratic deliberation begins to look quaint before it has even started.

Frontier laboratories currently possess an extraordinary concentration of epistemic power. They build the systems, control many of the interfaces, design the evaluations, hold proprietary performance and incident data, and supply much of the language through which governments and the public understand the technology. Their expertise is indispensable. Their institutional position is not neutral.

This is not an accusation of conspiracy. It is a structural conflict. The organizations seeking permission and capital to scale a technology cannot also be the sole authors of the ontology, evidence standards and risk categories by which that technology is governed. A vocabulary centered on hypothetical loss of machine control may be sincere and still background other forms of loss: loss of labour power, privacy, epistemic independence, cultural plurality, market competition and democratic control.

Whenever “loss of control” is invoked, we should ask: whose control, over what? Human control over an automated system, state control over a security risk, corporate control over distribution and market control by an incumbent are not the same problem. They can even point toward opposing policies. Collapsing them into one phrase allows institutional interests to travel inside technical language unnoticed.

Safety rules can be necessary while also advantaging the incumbents able to comply with them. Open weights complicate the picture further. They distribute capability, enable independent research and reduce dependence on a handful of APIs, while also making some safeguards easier to remove. The US Commerce Department’s open-model report recognized both sides, concluded that the available evidence did not yet justify a definitive position on restrictions, and recommended continuing research and monitoring.

Even the word open needs semantic repair. Open-weight is not open-source. A downloadable parameter file does not tell us whether the training data, code, evaluation process or license is open. Distributed does not mean democratic. Proprietary does not mean safe. Governance cannot be built from brand categories.

Nor can we reduce the global argument to “a fearful West versus an optimistic China.” The contrast is real, but more specific. The current US policy language emphasizes winning an AI race and securing technological dominance. The UK’s AI Security Institute and European institutions foreground security, systemic risk and regulated deployment; the EU’s own guidance distinguishes general-purpose models from the systems in which they are used. China’s official Global AI Governance Action Plan describes AI as an engine of industrial development and a potential international public good, and promotes adoption, infrastructure and open-source ecosystems, while also insisting on safety, controllability and national sovereignty.

These are differences of governing metaphor, not a clean division between fear and freedom. Race language builds champions and barriers. Apocalypse language builds gatekeepers. Empowerment language accelerates deployment. Controllability language expands state oversight. Every frame reveals something and conceals something.

We need a language capable of seeing all of it at once

We don’t need universal agreement about consciousness, intelligence or the future before we act. We need a shared protocol for disagreement.

It can begin with eight rules:

1. Name the object. Say whether the claim concerns a model, product, agentic system, company, market or imagined future system.

2. Name the layer. Distinguish mechanism, behaviour, human experience and institutional responsibility.

3. Type mental language. When using words such as knows, wants or decides, state whether they are metaphorical, functional, behavioural or ontological claims.

4. Separate capability from authority. Measure what a system can do independently from what it is permitted to do in the world.

5. Describe risk as a mechanism. Identify the actor, pathway, affected population, evidence, time horizon, reversibility and uncertainty. “AI is dangerous” is not a risk assessment.

6. Keep responsibility attached. A deployed system must not become what Madeleine Clare Elish calls a “moral crumple zone”, absorbing language that protects the institutions around it—or shifting blame onto the least powerful human in the loop.

7. Pluralize the authorship of meaning. No laboratory, ideology, country or discipline should be able to define intelligence, safety, alignment or progress for everyone else.

8. Define the purpose. Almost since inception, AI has been defined by what it is not rather than what it could be. We are now grappling with what it is. We need to start using applied goals that are centred around the improvement of the human condition. The betterment of society, a focus on higher purpose applications, where low value labour can be reduced. We need to stop asking “what could this be” and start asking “how can it make our world better now”.

This is semantic governance: the public infrastructure of naming what a technology is doing before deciding what it should be allowed to do.

It sounds completely different outside of technology focused language, which is based on its own systems and measurements : scaling laws, trillion-dollar infrastructure plans and warnings about the end of the species. These are not the larger measurement of human society. These are industry terms treated as the canon of progress. They are not. We must not allow the language of foundational technology that is changing the world to default to that of a particular industry. Without a common language, every technical advance will widen the interpretive gap. Discourse will become more polarized. Fear will become easier to weaponize. Corporate categories will harden into public law. And society will surrender more authority to the institutions that possess both the systems and the words used to explain them.

Artificial intelligence may eclipse every technology we have built before it. But the technology has already raced ahead of our capacity to understand it, describe it and place it inside a shared account of human meaning.

That is an enormous challenge for our species. It is not an unconquerable one.We don’t need a perfect philosophy before we govern AI. We need enough shared language to know what we are actually disagreeing about, and ways to ensure that the study of that philosophy has a path to evolve.

The immediate danger is not that machines have become human. If anything needs to be become more clear, it is that human intelligence and machine intelligence are divergent; they are not the same things. We should not be asking one to emulate the other. We should be allowing each to focus on what it is best at.

And one should always be in service of the other. We are already past the point where we can understand everything that AI is doing, because its pattern recognition and prediction capabilities far eclipse our own. What we must concern ourselves with now is that human institutions may surrender judgment, authority and democratic control to these capabilities while arguing with metaphors.

The first task is to name what is happening. The second is to decide what it means. Only then can we govern what it is becoming.