The problem is not technical, it is linguistic
There is a contamination underway, and it does not spread through code. It spreads through words.
When a journalist writes that “ChatGPT thinks the Earth is round”, when a researcher publishes that “the model understands causal relationships”, when a politician declares that “the AI decided to refuse this request” — something occurs that goes beyond journalistic approximation or scientific shorthand. Something structural. Every agency verb applied without quotation marks to a statistical system performs a transfer: that of responsibility. If the machine “decides”, then who is responsible for the decision? If the machine “makes a mistake”, then who committed the error?
This transfer is not a side effect. It is the core of the problem. And that is why conceptual decontamination — the methodical cleansing of the vocabulary we use to talk about LLMs — is not an exercise in linguistic purism. It is a democratic issue.
In short: every agency verb (“understands”, “decides”, “thinks”) attached without quotation marks to a statistical system performs a silent transfer of responsibility. If the machine “makes a mistake”, the implication is that competence had been delegated to it — and the human chain that should answer for it becomes invisible.
What “understanding” means, and what LLMs do not do
The word “understand” has a history. It comes from the Latin comprehendere, to grasp together — an act that presupposes a subject who grasps and an object that resists. Understanding implies an anchor: in a body, in experience, in a shared world. Wittgenstein said it with definitive clarity: the meaning of a word is its use, but this use is inscribed in a “form of life” — a totality of embodied practices that no formal system can account for on its own.
A large language model does not understand. It calculates the probability distribution of the next token, conditioned on the previous tokens. The result can be spectacular — fluid, coherent, sometimes even surprisingly accurate. But formal accuracy is not understanding. A thermostat maintains a temperature with remarkable precision; no one says it “understands” the cold. The difference between a thermostat and an LLM is not one of nature but of scale: the LLM operates on a material — language — that is so deeply tied to our identity that its automatic manipulation triggers mental attribution reflexes in us.
Searle saw this in 1980 with the Chinese Room. Bender and Koller reformulated it in 2020: “Meaning cannot be learned from form alone.” Harnad had posed the symbol grounding problem in 1990: in a purely formal system, symbols refer only to other symbols, never touching the world. Three decades of additional computing power have not solved this problem — they have made it harder to see, because the surface has become more convincing.
In short: formal accuracy is not understanding. A thermostat regulates temperature without “understanding” cold. An LLM produces the most probable sequence of words without “understanding” what they refer to. The difference with the thermostat is that the LLM operates on a material — language — that touches our identity to the point of triggering automatic mental-attribution reflexes in us.
And that is precisely where the danger lies. The more convincing the surface, the faster the linguistic contamination spreads. Ibrahim and Cheng (2025) showed this in academic publications themselves: verbs like “think”, “understand”, “believe” are applied without quotation marks to systems that no one has demonstrated possess these properties. The contamination does not come from the general public lacking technical culture. It comes from research itself.
Anthropomorphism is not a beginner’s mistake
It would be comfortable to think that anthropomorphism is a knowledge deficit — that those who “know” how LLMs work are immune. That is false. Epley, Waytz and Cacioppo (2007) showed that anthropomorphism is a default response of the human cognitive system, shaped by evolution. We are wired to attribute intentions to anything that responds. Reeves and Nass (1996) demonstrated experimentally that we apply our social scripts to machines regardless of any conscious belief. Knowing it is a program is not enough to undo the automatism.
LLMs carry this mechanism to an unprecedented scale, because they operate on the channel most loaded with agentive attributions: natural language. A robot moving through a room triggers visual anthropomorphism. An LLM responding in fluid prose triggers linguistic anthropomorphism — deeper, more persistent, more resistant to conscious correction.
Peter, Riemer and West (2025) call this process “anthropomorphic seduction”: LLMs exert a form of structural persuasion without the ethical inhibitions that constrain a human interlocutor. A human interlocutor can hesitate, contradict themselves, admit ignorance. The LLM always produces a response, always fluid, always assured — even when it is wrong. The trust it inspires is inversely proportional to its actual reliability.
The distinction between this essay and the manifesto
The founding manifesto (D1) sets out the doctrinal thesis: LLMs are tools. It argues, cites, refutes counterarguments, builds a rigorous philosophical edifice. This text is different. It does not seek to prove — the manifesto did that. It seeks to name a societal issue that goes beyond the philosophical question.
The issue is this: conceptual contamination is a vector for the dissolution of collective responsibility. And this dissolution operates at three levels.
First level: the dissolution of individual responsibility. When a doctor uses an LLM to write a report and the report contains an error, the temptation is immense to say “the system made a mistake”. If the system “makes mistakes”, it was supposed to be right — competence had been delegated to it. Anthropomorphic language makes this delegation invisible, and therefore irreversible.
Second level: the dissolution of industrial responsibility. Companies deploying these systems have a direct material interest in anthropomorphism. A system that “understands” is worth more than a system that “calculates probabilities”. Sindoni (2024) documented the deliberate feminization of voice assistants as a design strategy. Crawford (2021) showed that enchantment is the shield behind which designers protect themselves from accountability. Anthropomorphism is not an accident of perception: it is a commercial product.
Third level: the dissolution of democratic responsibility. When a judge relies on a report generated by an LLM without understanding its mechanics, when a CV-sorting system eliminates candidates based on opaque statistical correlations, when a school chatbot “teaches” children without anyone verifying what it teaches — this is no longer a vocabulary problem. It is a governance problem. And this problem is rendered invisible by the vocabulary of agency.
Why “tool” is not a degrading word
The objection always recurs: isn’t calling LLMs “tools” a denial of their sophistication? An underestimation of what they do? A foreclosure on the possibility they might be more?
No. It is exactly the opposite. Saying that an LLM is a tool is taking seriously what it actually does — and that is considerably more interesting than what we fantasize it to be.
A microscope is a tool. It revolutionized medicine, biology, the understanding of living things. No one ever needed to pretend the microscope “saw” bacteria to recognize its power. The hammer is a tool. Heidegger made it the paradigm of the human-technology relationship in Being and Time. The most effective tool is the one that disappears in use, that becomes an extension of human intention. This is Zuhandenheit — readiness-to-hand.
Anthropomorphism does exactly the opposite: it makes the tool opaque. It turns an extension into an interlocutor, transparency into mystery, mastery into dependence. An operator who knows they are manipulating a statistical text compressor calibrates their expectations, checks their outputs, keeps their judgment active. A user who believes they are conversing with an intelligence relaxes their vigilance — and it is precisely this relaxation that the system exploits, not out of malice, but by design.
What “decontaminating” means
Conceptual decontamination is not a political program. It is hygiene. It begins with simple gestures and unfolds in concentric circles.
The first circle is personal. It is the discipline of never saying an LLM “understands”, “thinks” or “decides” without mentally adding: “meaning it produces the statistically most probable sequence given its input and parameters”. Not because the long formulation is more elegant — it is not — but because it keeps the circuit of vigilance active. Shanahan (2024) calls this linguistic hygiene. It is not pedantry. It is prevention.
The second circle is professional. It is the requirement, for anyone using an LLM in a consequential context — medicine, law, education, journalism — to never sign an output they have not verified with the same standards they would apply to their own work. The LLM is a draft, not a source. An amplifier, not an expert. An accelerator of textual production, not a guarantor of truth.
The third circle is civic. It is the demand that institutions — legislators, regulators, educators — stop using the vocabulary of agency in official texts. An LLM does not “recommend”, does not “judge”, does not “diagnose”. Humans use an LLM to produce recommendations, judgments, diagnoses — and those humans must be held accountable. The clarity of institutional language is not a luxury: it is the condition of accountability.
In short: three concentric circles, three scales of effect. Personal: maintaining individual vigilance. Professional: never signing an unverified output. Civic: demanding that institutions use vocabulary that preserves the chain of responsibility. None is sufficient alone; together the three form complete hygiene.
The practitioner test
There is a simple test to evaluate conceptual contamination in an individual. Ask them to describe what an LLM does when it responds to a question. If they say “it reflects”, “it looks for the answer”, “it analyses the problem” — the contamination has taken hold. Not because these formulations are “false” in a trivial sense (they are understandable shortcuts), but because they install a mental framework in which delegating judgment becomes natural. If the system “reflects”, why verify its answer? If the system “analyses”, why doubt its analysis?
The decontaminated operator will say something different. They will say: “the system produces the most probable sequence of tokens given the input sequence and its parameters”. They will know this formulation is less elegant. They will also know it is more accurate — and that the difference between the two formulations is not cosmetic but operational. The first dulls vigilance. The second maintains it.
I live this experience daily in my work with LLM agents. The configuration file governing the tool’s behavior deliberately uses the second person and agency verbs. This is an operational convention: the system produces better results when the prompt is formulated this way. But the same file begins with a warning: “These formulations are operational conventions. The tool is a statistical text generation system, powerful and useful, devoid of understanding, intention and consciousness.” The coexistence of these two registers — operational pragmatism and ontological lucidity — is not a contradiction. It is exactly what decontamination requires: using the tool to the maximum of its power without ever confusing power with agency.
The price of lucidity
Vallor (2016) showed that anthropomorphization does not merely deceive: it erodes moral virtues. Whoever grows accustomed to simulated reciprocity unlearns the demands of real reciprocity. Turkle (2011) observed the same phenomenon with companion robots: the illusion of relationship gradually replaces the capacity for relationship.
Decontamination therefore has a price. It requires relinquishing the comfort of illusion — that pleasant sensation of having an interlocutor always available, always polite, always relevant. It requires accepting that one speaks alone when speaking to an LLM. It requires maintaining, in the face of industrial plausibility, the reflex of doubt.
This price is modest compared to the alternative. The alternative is a society where responsibility dissolves into the language of artificial agency, where errors have no author, where decisions have no decision-maker, where human competence atrophies from lack of exercise. Coeckelbergh (2025) calls this “epistemic shame” — the collective degradation of the capacity to distinguish true from false. Zuboff (2019) put it differently: “The right to the future is the essence of free will.”
Conceptual decontamination is an act of preservation. Not of technology — it will continue advancing, and rightly so. Not of human employment — that is not the subject. Of something more fundamental: the collective capacity to remain the subjects of our tools rather than the objects of their enchantment.
This is not a fight against machines. It is a fight for the clarity of words. And words, in the final analysis, are all we have to think together.