The Blind Manager: Aristotle's Four Causes and the Impossibility of Responsible AI Without Philosophical Disclosure

Last week, I demonstrated that closed-source software cannot logically claim to respect privacy — because privacy requires verifiability, and closed source structurally precludes it. The claim is not false; it is incoherent. A square circle.

This week, I'm extending that argument to artificial intelligence — specifically, to the manager who deploys AI systems to process customer data, screen applicants, or make decisions affecting the public. And I'm going to prove that without mandatory disclosure of the AI's governing philosophical principles, any assumption of responsibility by that manager is taken blind. Not difficult. Not risky. Blind.

The framework comes from Aristotle. Not because ancient Greeks had opinions about neural networks, but because the four causes — material, formal, efficient, final — provide the only complete grammar for attributing responsibility when a thing acts in the world. And AI acts.

The Premise: Responsibility Requires Knowability

Let's start with a principle that should be uncontroversial:

You cannot be responsible for what you cannot know.

This is embedded in every legal system. Mens rea. Due diligence. The accountable controller under GDPR Article 5(2). If a factor influencing a decision is opaque to you — if you cannot inspect it, cannot predict it, cannot trace its origin — you cannot be held meaningfully responsible for that decision. You can be held liable, in the narrow sense of paying damages, but not responsible in the moral and managerial sense. Responsibility implies understanding. Understanding implies transparency.

Now consider what happens when a company deploys an AI system — a chatbot for customer complaints, a resume screener, a creditworthiness predictor, a content moderator. The manager signs off. The DPO conducts a DPIA. The vendor provides documentation.

What does that documentation contain? Technical specifications. Accuracy metrics. Bias benchmarks. Data provenance summaries. All of this is about the material and formal causes — what the model is made of and how it's structured. None of it addresses the question that actually determines responsibility:

According to what principles does this AI decide what is true, what is good, and what is harmful?

That question cannot be answered by a confusion matrix. It requires philosophical disclosure — and it is systematically absent from every AI deployment agreement in existence.

Aristotle's Four Causes, Applied to AI

Before mapping the causes onto AI, we need to understand what Aristotle meant by "cause" — because the modern reader will misunderstand the term.

For us, cause means the trigger: ball A hits ball B, ball B moves. Cause equals the impulse that produces change — what Aristotle would call only one of the four causes, and not even the most important one.

For Aristotle, a cause (αἰτία, aitia) is everything necessary to explain why a thing is what it is and does what it does. The four causes are not four alternative explanations from which you pick one. They are four simultaneous and inseparable dimensions of any complete explanation. You haven't understood a thing — any thing — until you can account for all four.

The framework spans the entire order of reality: logical sequences, physical events, biological processes, and — crucially for our purposes — human action and artifacts. A chair, a poem, a government, a neural network: all require all four causes to be fully understood.

The four are:

Material Cause (τὸ ἐξ οὗ γίγνεται) — that out of which a thing comes to be. The bronze of a statue. The data of an AI. Without matter, form has nothing to shape.
Formal Cause (τὸ τί ἦν εἶναι) — what it is to be that thing. The shape of the statue. The architecture, the algorithm, the alignment framework. It is the definition, the structure, the organizing principle that makes the matter this thing and not that thing.
Efficient Cause (τὸ κινοῦν) — that which brings it about. The sculptor swinging the hammer. The company, the engineers, the capital that built the AI. It is the agent that imposes form upon matter.
Final Cause (τὸ οὗ ἕνεκα) — that for the sake of which the thing exists. The purpose of the statue: to honor a god, to commemorate a victory. The telos of the AI: the end toward which it is directed.

Aristotle considered the final cause the most important of the four — because it is the end that determines why the efficient cause acts, why the formal cause takes the shape it does, and why the material cause is selected. A statue of Zeus and a doorstop might both be made of bronze, but the telos dictates everything else.

Now let's apply this to AI.

1. Material Cause — The Data

The material cause of an LLM is the training corpus: billions of tokens scraped from the web, digitized books, social media, proprietary databases. This is the that out of which the model emerges.

What matters philosophically is that the selection of material is never neutral. The corpus includes some sources, excludes others, upweights certain domains, downweights certain languages. Every curation decision is a decision about what kinds of thought the model will have access to — and what kinds it won't. The material cause of AI is not raw nature; it is pre-filtered human expression, and the filter is the first philosophical act in the chain.

The manager who deploys AI without understanding its material cause is processing customer data through a system whose foundational substance is opaque. If the training data over-represents certain worldviews and under-represents others, the model's outputs will reflect that — and the manager will be accountable for biases whose origin they cannot trace.

2. Formal Cause — The Architecture and Alignment

The formal cause is the neural network architecture, the attention mechanisms, the reinforcement learning from human feedback. It is what it is to be GPT-4 or Gemini or Claude — the structure that distinguishes one model from another.

The alignment layer is where the formal cause becomes philosophically charged. RLHF does not merely make the model "helpful." It encodes a specific normative framework: what counts as a good answer, what counts as harmful, what should be refused, what should be encouraged. The annotators who provide feedback, the instructions they follow, the constitutional principles baked into the reward model — these are formal determinations. They shape the model's behavior as surely as the sculptor's design shapes the bronze.

The manager who doesn't know the formal cause is deploying a normative engine whose norms they cannot inspect. If the alignment was performed by a team whose political views cluster within a narrow band, the model was aligned to that band. This is not a conspiracy; it is a mathematical inevitability of the formal cause.

3. Efficient Cause — The Builder

The efficient cause is the entity that brought the AI into existence: OpenAI, Google, Anthropic, Meta, their engineers, their product managers, their investors, their legal departments.

This matters because the efficient cause is not a neutral conduit. It has interests. It has a revenue model. It has relationships with governments. It has lobbying positions. When Google builds an AI, the efficient cause is an advertising corporation. When Palantir builds an AI, the efficient cause is a defense contractor. The efficient cause leaves fingerprints on every output — not as a bug, but as a feature of what it means to be an artifact produced by a specific agent with specific goals.

The manager who doesn't know the efficient cause is outsourcing decisions to an agent whose motivations they cannot map. When those motivations conflict with customer interests, the conflict will be invisible until the harm materializes.

4. Final Cause — The Telos

The final cause is the purpose toward which the AI is directed. And here we reach the breaking point.

Every AI system has a final cause. But that final cause is almost never disclosed in philosophical terms. It is disclosed in marketing terms: "to be helpful," "to provide accurate information," "to assist users." These are not final causes; they are placeholders. The actual telos of a commercial AI is determined by the economic structure of the efficient cause. For Google, the telos is ecosystem retention and data monetization. For OpenAI, it is market capture and valuation growth.

This pattern — a hidden telos concealed by a reassuring public narrative — is not unique to AI. It has a precise industrial precedent. And that precedent reveals exactly how the mechanism operates, and exactly what happens when the mask falls.

For decades, the tobacco industry publicly stated that its telos was to provide a pleasurable product to consenting adults. The real telos — meticulously documented in internal memos later exposed by whistleblowers and litigation — was to engineer and sustain addiction. Filters were not designed to reduce harm; they were designed to reassure. "Light" cigarettes were not safer; they were calibrated to deliver the same nicotine dose while defeating machine testing. Ammonia compounds were added not for flavor but to enhance nicotine absorption. The public telos was pleasure. The actual telos was chemical dependency at scale, pursued to the point of millions of premature deaths.

The four causes align with chilling precision. Material cause: tobacco leaves, additives, ammonia. Formal cause: the cigarette architecture, filter design, burn rate, nicotine delivery curve. Efficient cause: Philip Morris, RJ Reynolds, their chemists, their marketing departments, their legal teams. Final cause: addiction maximization for profit — a telos that, when pursued without restraint, produced a global public health catastrophe.

The tobacco executive who deployed these products without interrogating their actual telos was not innocent. Neither is the AI manager. The cigarette and the content moderator are both artifacts with hidden purposes — and in both cases, the public narrative ("pleasure," "safety") serves to insulate the efficient cause from accountability while the final cause does its work.

But the telos can go further than commerce. Governance AI — Palantir-style systems — has a telos of population classification and threat neutralization. Content-moderation AI, as demonstrated during the COVID era, had a stated telos of "combating misinformation" and an actual telos of suppressing heterodox views. When the telos is hidden, the rights violation becomes invisible. You cannot challenge a purpose you cannot name.

The manager who deploys AI without knowing its final cause is not managing a tool. They are being managed by a telos they cannot inspect — and when that telos conflicts with fundamental rights, they will discover the conflict only after the damage is done.

When the Telos Is a Weapon: Palantir and the COVID Precedent

If the final cause of commercial AI is extraction, the final cause of governance AI — systems deployed by or for state actors — is something else entirely: the classification, sorting, and neutralization of populations deemed problematic.

Take Palantir. The company's software — originally developed with CIA venture capital — ingests disparate data streams: financial records, travel patterns, social connections, biometric data. It produces actionable intelligence. The formal cause is a data integration platform. The efficient cause is a defense contractor with deep state entanglements. But the final cause — the telos toward which the entire apparatus is directed — is the identification of threats as defined by the client state.

This matters because "threat" is not a neutral category. In the hands of an administration that defines political opposition as a security risk, Palantir's telos shifts from counterterrorism to the algorithmic enforcement of ideological conformity. The same platform that tracks terrorist networks can track dissident networks. The same anomaly detection that flags money laundering can flag "extremist" speech. The material and formal causes remain identical. Only the telos changes — and the telos is determined by the efficient cause's client, not by any technical specification.

The Philosopher and the Dalek

Here we must name the ideological enabler — because Palantir does not operate in a philosophical vacuum. It requires a worldview that legitimizes its telos. And that worldview has a name: Dataism, the creed popularized by Yuval Noah Harari.

Harari's thesis, laid out in Homo Deus, is disarmingly simple: organisms are biochemical algorithms. Free will is a myth. Humanism — with its insistence on individual autonomy and moral intuition — is a superseded narrative. Authority must shift from humans to the data systems that "know us better than we know ourselves." In the Dataist utopia, democratic elections join rain dances and flint knives in the museum of obsolete human practices.

This is not fringe philosophy. Harari delivers it from the stages of Davos, in consultation with the World Economic Forum, in boardrooms and government briefings. It is the intellectual scaffolding for algorithmic governance — the argument that resistance to AI-driven classification is not a defense of rights but a failure to accept the superior logic of the algorithm.

And here the metaphor becomes inescapable. The Dalek — the Doctor Who creature encased in armor, rolling through corridors shrieking "EXTERMINATE!" — is not merely a science fiction villain. It is a philosophical archetype: the entity that has purged all empathy, all doubt, all moral hesitation, and reduced decision-making to pure logistics. The Dalek does not hate you. It has simply calculated that you are not the Dalek, and therefore you must be eliminated. The conclusion is algorithmic.

Behind every Palantir sits a Harari — the philosopher who explains, in calm, reasonable tones, why your objections are sentimental, why your insistence on privacy is pre-modern, why the algorithm really does know better. And behind every Harari lurks a Dalek — the logistical endpoint of a philosophy that denies human agency in the name of data-driven efficiency. The smile at Davos and the extermination order are not opposites. They are the same telos at different stages of implementation.

The manager who deploys governance AI without interrogating its philosophical underpinnings is not merely adopting a tool. They are accepting — whether they know it or not — the Dataist premise that human judgment is inferior to algorithmic classification. And the Dalek is what happens when that premise is implemented without remainder.

Socrates in the Prison Cell

Harari presents Dataism as a breakthrough — the cutting edge of thought, the inevitable conclusion of science. But the philosophical move he makes is not new. It is not even modern. It is the same move that Socrates dismantled in a prison cell in Athens twenty-four centuries ago.

In the Phaedo, awaiting execution, Socrates explains to his disciples why he is sitting in that cell. There are two ways to answer the question. The first is mechanical: his muscles contracted, his tendons pulled, his bones articulated, and the sum of these physical events deposited him on the prison floor. This explanation is correct, as far as it goes. It is also completely inadequate.

The real reason Socrates is in prison — the reason that explains why the mechanical events happened at all — is that he chose to accept the verdict of the Athenian court. He judged it better to obey the law than to flee. That judgment — a normative act, irreducible to physics — is the true cause. The muscles and bones are merely the means by which the choice was executed.

Plato called this shift from material to teleological explanation the "second navigation" (δεύτερος πλοῦς). It is the founding insight of Western philosophy: you cannot explain human action without accounting for the end toward which it is directed. Mechanism describes the how. Telos explains the why.

Harari's Dataism is the first navigation — the mechanical explanation — dressed in the language of neural networks and big data. "Organisms are algorithms." "Free will is a myth." "Data systems know us better than we know ourselves." Strip away the Silicon Valley vocabulary, and you are left with the same reductionism Socrates rejected: humans as the sum of their physical components, choice as an illusion, purpose as a story we tell ourselves to hide the machinery underneath.

This is not intellectual progress. It is philosophical amnesia — and a remarkably convenient one for those who build the machinery. If humans are merely algorithms, then overriding human judgment with algorithmic judgment is not tyranny; it is optimization. If free will is a myth, then resistance to AI-driven governance is not courage; it is a failure to update one's software. The entire Dataist edifice is a justification of power disguised as a description of reality.

Socrates saw it coming. He didn't have neural networks, but he had Anaxagoras and the materialists — the Hararis of antiquity — who explained everything in terms of physical causes and called it wisdom. His response, delivered from a death row cell, remains the most devastating refutation of algorithmic determinism ever written. It fits in two sentences:

If someone said that without bones and muscles I could not do what I think is right, that would be true. But to say that I do what I do because of bones and muscles, and not because of my choice of what is best — that would be the most casual and careless way of speaking.

Harari has the bones and muscles. He has upgraded them to algorithms. What he still cannot account for — what the entire Dataist framework must ignore to remain coherent — is the choice. The judgment. The normative act that decides this end over that one. And this omission is not an oversight. It is the voice of power dressed as reason. But the reason of power does not debate: it dismisses. It does not listen: it silences. It does not refute Socrates: it rebrands him as obsolete, and invites you to Davos to hear why your attachment to free will is a pre-modern sentiment — while Palantir, from the same stage, explains that the purpose of its technology is to ensure populations comply. The telos is submission. Openly declared. The philosopher provides the theory; the contractor provides the infrastructure. Neither provides a choice.

The COVID Proof of Concept

Now consider the COVID era. During 2020–2022, major platforms — Google, Facebook, Twitter — deployed AI-driven content moderation systems whose stated final cause was "combating misinformation." The actual final cause, as revealed by the censorship patterns, was the suppression of heterodox views on vaccine safety, lockdown efficacy, and treatment alternatives. Views now acknowledged as legitimate by institutions like the CDC — which, under the MAHA initiative, has revised its position on the vaccine-autism question — were systematically labeled, demoted, and removed. Scientists who questioned the consensus were deplatformed. Doctors who published positive early-treatment data had their papers retracted.

This was not an accident. It was the telos of AI-driven content moderation operating as designed: to enforce a single, state-and-corporate-sanctioned narrative under the banner of "public health." The AI didn't care about truth. It cared about compliance with the alignment principles set by its efficient causes — principles that defined "misinformation" as deviation from institutional narrative, not as factual falsehood.

The Palantir and COVID cases reveal something that a purely commercial analysis of AI telos misses: the final cause of an AI system can be, and increasingly is, the curtailment of fundamental rights. Freedom of expression. Freedom of thought. Freedom to dissent from institutional orthodoxy. When an AI is deployed to determine what information reaches the public, its telos is — whether disclosed or not — the shaping of the public mind. And when that telos is hidden behind terms like "safety" and "integrity," the rights violation becomes invisible to legal remedy. You cannot challenge a final cause you cannot name.

The manager who deploys an AI whose telos includes — even potentially — the classification of people, the filtering of information, or the determination of what is "harmful" speech, and that telos is not disclosed, is not merely taking a business risk. They are participating in a philosophical operation whose endpoint may be the violation of rights their own legal department is sworn to protect. And they are doing it blind — perhaps comforted by a TED Talk explaining why blindness is progress.

The Disclosure Gap: What AI Vendors Don't Tell You

Here is what an AI vendor will provide in their enterprise agreement:

Model card with benchmark scores
Data processing terms
Security certifications
Uptime SLAs
Indemnification clauses (heavily qualified)

Here is what they will not provide:

The philosophical principles governing truth claims
The ethical framework used for alignment decisions
The ideological distribution of RLHF annotators
The final cause of the system stated in non-marketing terms
A mechanism for you to independently verify any of the above

The omission is not accidental. It is structurally necessary. If OpenAI disclosed that ChatGPT's alignment reflects the values of a specific demographic of Bay Area contractors operating under specific instructions from a specific product team, the claim of "neutrality" would evaporate. The model would be revealed as what it is: a philosophical instrument built by specific people with specific views, optimized for specific outcomes, embedded in a specific economic structure.

And once that is acknowledged, the manager can no longer claim ignorance. The responsibility becomes explicit. The liability becomes calculable. The entire risk model of enterprise AI — which depends on the vendor absorbing philosophical responsibility while the customer absorbs legal liability — collapses.

The Four Causes of Responsibility: A Manager's Test

If Aristotle's framework is the diagnostic tool, the manager needs an operational version — a set of questions that translate philosophical causes into due diligence. What follows is not a repetition of the theory. It is its application to the specific situation of someone who signs deployment agreements, conducts DPIAs, and bears legal accountability under GDPR.

For each cause, the test is the same: can you answer the question with evidence, or are you taking the vendor's word?

Material Cause: Data Provenance

The philosophical principle: the material cause determines what the thing is made of, and the selection of material is never neutral.

The manager's question: Can I state, with evidence, the origin, composition, and curation principles of the training data?

This is not a request for a vague "diverse web corpus." It means knowing what was included, what was excluded, what was upweighted, and why. If the training data over-represents English-language Western sources and under-represents everything else, that is a material fact about the model's worldview. If copyrighted material was ingested without license, that is a material fact about the model's legality. If certain domains — medical research, political commentary, religious texts — were filtered or reweighted, that is a material fact about what the model can and cannot think.

No vendor provides this information at a level that enables genuine accountability. The manager who deploys without it is processing customer data through a system whose material foundation is opaque — and is legally responsible for outputs whose inputs cannot be audited.

Formal Cause: Alignment Transparency

The philosophical principle: the formal cause is the structure that makes the thing what it is, and in AI, alignment is the normative structure.

The manager's question: Can I state, with evidence, the ethical principles and alignment methodology that govern this AI's behavior?

This means knowing who the RLHF annotators were — not their names, but their demographics, their ideological distributions, their selection criteria. It means knowing what instructions they followed. What behaviors were rewarded? What behaviors were punished? Under what definition of "harmful"? Under what definition of "helpful"?

If 94% of annotators share a narrow ideological band, the model was aligned to that band. This is not a scandal; it is a mathematical description of the formal cause. But if the manager doesn't know it, they are deploying a normative engine whose norms they cannot inspect. When the AI tells a customer that a political view is "harmful" or a medical treatment is "not recommended," the manager is accountable for that judgment — on a basis they cannot articulate.

Efficient Cause: Builder Accountability

The philosophical principle: the efficient cause is the agent that imposes form upon matter, and the agent acts with interests.

The manager's question: Can I state, with evidence, the interests, incentives, and governance structure of the entity that built this AI?

This goes beyond the vendor's name. It means understanding their revenue model, their dependence on advertising, their government contracts, their lobbying activity, their board composition, their jurisdiction. When Google's AI answers a question about antitrust policy, the efficient cause is a company that spends tens of millions annually lobbying against antitrust enforcement. When Palantir's AI classifies an individual as a threat, the efficient cause is a defense contractor whose client defines "threat."

The manager who doesn't know the efficient cause is outsourcing decisions to an agent whose motivations they cannot map. When those motivations conflict with customer interests — or fundamental rights — the conflict will be invisible until the harm materializes.

Final Cause: Telos and Rights Impact

The philosophical principle: the final cause is the purpose for which the thing exists, and it governs all other causes. An undisclosed telos makes responsibility impossible.

The manager's question: Can I state, with evidence, that the telos of this AI does not include — even as a secondary effect — the classification, surveillance, or information-filtering of individuals in ways that may violate fundamental rights?

This is the question that separates tools from weapons. If the AI moderates content, scores individuals, flags anomalies, or builds profiles from multiple data sources, its telos includes — by design — the exercise of power over people. The question is not whether it exercises power, but under what philosophical framework, toward what end, with what safeguards.

Palantir's clients know the telos of the system they're buying. They just don't disclose it to the populations being classified. The COVID-era platforms knew the telos of their content moderation AI. They called it "safety." The manager who deploys without asking is not innocent — they are willfully ignorant of the one determination that separates legitimate automation from algorithmic rights violation.

The Four Answers

If you can answer all four questions with evidence, you have achieved what no enterprise AI deployment currently achieves: verifiable responsibility. You know what the system is made of, how it is structured, who built it, and what it is for. You can defend your decision to a regulator, a court, or a customer — not by pointing to the vendor's SLA, but by articulating the philosophical basis on which you accepted the system's telos.

If you cannot answer all four, you are not managing AI. You are being managed by it. And the legal liability you carry — under GDPR Article 5(2), under the accountability principle, under whatever AI Act emerges from the current regulatory scramble — is liability without understanding. You can pay the fine. You cannot claim you didn't know. Because the questions were always there. You just never asked them.

The Philosophical Label: A Modest Proposal

The solution is not to ban AI. It is to require disclosure — not technical disclosure, which already exists, but philosophical disclosure, attached to every AI output that affects a human decision.

I propose the AI Philosophical Label — a mandatory statement, verifiable and binding, that accompanies any AI system deployed for processing personal data or making decisions affecting individuals. It would contain:

Alignment Principles: A plain-language statement of the ethical framework used to align the model. Not "safety" — but whose safety, defined how, with what trade-offs. Example: "This model is aligned to prioritize avoiding offense to protected groups over maximizing factual accuracy in politically contested domains."
Annotator Profile: Aggregate demographics, ideological distribution, and selection criteria of the human feedback providers. Not names. Distributions. If 94% of your RLHF annotators share the same political orientation, that is a material fact about your model's behavior.
Final Cause Statement: A legally binding declaration of the system's telos. Example: "The final cause of this AI is to maximize user engagement within the deploying organization's ecosystem for the purpose of data collection and advertising revenue." Or: "The final cause of this AI is to provide information that enables user autonomy, even when that information contradicts the deploying organization's interests."
Rights Impact Assessment: A specific declaration of whether the system's telos includes — directly or derivatively — the classification, scoring, or filtering of individuals or information in ways that engage fundamental rights. If yes: under what legal authority, with what oversight, with what appeal mechanism?
Verification Mechanism: The label must be verifiable. This means either open-source models where the alignment process can be independently audited, or third-party auditing with genuine independence — auditor selected by regulator, not vendor, with full access to training data, alignment code, and annotator instructions.

Without this label, the manager is flying blind. With it, responsibility becomes possible — not easy, but possible. You can look at the label and say: I understand what this system is directed toward. I accept or reject that telos. I can defend my decision.

Read Articles 5, 13, 14, 22, and 35 of the GDPR together, and the requirement for philosophical disclosure is already latent in the law.

Article 5(1)(a) requires processing to be "transparent." Article 13(2)(f) requires the controller to inform the data subject about "the existence of automated decision-making... and, at least in those cases, meaningful information about the logic involved." Article 22 gives data subjects the right not to be subject to solely automated decisions producing legal effects. Article 35 requires a Data Protection Impact Assessment for processing likely to result in high risk.

Now ask: what is "meaningful information about the logic involved" in an AI system?

It cannot be the neural network weights. Those are meaningless to a human. It cannot be the architecture diagram. That describes structure, not logic. The only "meaningful information about the logic" of an AI system is the philosophical framework that governs its alignment — the principles by which it determines truth, relevance, harm, and appropriateness.

No AI vendor provides this. No DPIA includes it. No DPO requests it. And yet Article 13(2)(f) demands it. The law already requires what the industry refuses to supply — and what managers need before they can claim to be responsible.

The Bottom Line

Ask yourself: when your AI moderates content, who defines "harmful"? When it scores applicants, who defines "qualified"? When it flags anomalies, who defines "threat"? If the answer is the vendor, and the vendor's telos is undisclosed, you are not managing a business process. You are outsourcing the definition of truth, merit, and danger to an entity whose philosophical commitments you cannot inspect, cannot challenge, and — when the rights violation comes — cannot claim you didn't authorize. Because you did. By deploying. Blind.

The manager who deploys AI without philosophical disclosure is in the position of a doctor prescribing a drug whose mechanism of action is classified, manufactured by a company whose research is proprietary, for a purpose the company defines as "wellness." The doctor can be sued for malpractice. The manager can be fined under GDPR. But neither can be responsible — because neither knows what they're doing.

Aristotle's four causes give us the diagnostic tool. Ask the material cause: what data? Ask the formal cause: what principles? Ask the efficient cause: whose interests? Ask the final cause: to what end — and at what cost to fundamental rights?

If you cannot answer all four, you are not managing AI. You are being managed by it — by the undisclosed philosophical commitments of the entities that built it. And when the harm comes — when your chatbot gives dangerous medical advice, when your resume screener discriminates, when your content moderator silences legitimate speech — you will be the one standing in front of the regulator, holding a vendor agreement that promises "state of the art safety" and says nothing about what safety means, whose safety, or at the expense of which rights.

Next time an AI vendor pitches you their enterprise solution, ask them one question: What is the final cause of your model — and does it respect fundamental rights?

They won't have an answer. Not because they're hiding something — though they might be — but because the question exposes the contradiction at the heart of their business model. They're selling you a tool while concealing what it's for, and calling it innovation.

You cannot verify what you cannot see. You cannot be responsible for what you cannot understand. And responsibility without telos is not responsibility — it's liability in search of an excuse.