The Great AI Language Collapse: Why Marketing Is Killing Accountability

Most AI titles and terms being used right now are dead wrong. That should scare us more than the technology itself.

What passes for authority today is often confidence without structure. A dangerous flattening is happening in plain sight. Operational requirements turn into marketing slogans, and accountability quietly disappears with the language.

Clarity of language is a prerequisite for accountability. Without precise terms, there is no way to assign responsibility, measure compliance, or hold anyone to account when a system fails. The AI industry has abandoned that precision. The result is ambiguity with consequences. Three distinct disciplines, each with different functions, different standards, and different consequences for failure, are being mashed into a single rhetorical paste. The people doing the mashing call it thought leadership. What it actually produces is a world where no one can tell the difference between a moral framework, an engineering practice, and an enforceable system of human authority.

If the goal is to move past abdication disguised as certainty, these three pillars cannot be treated as interchangeable.

Ethical AI is not a ban. It is a socio-technical framework for understanding power. It identifies harm vectors, protects human agency, and draws the line where AI should never stand alone. Ethics operates upstream as a human decision discipline, not as a system property. It does not require consensus to be valid. It requires the courage to hold a line when the crowd is confused.

Responsible AI is not trusting the system. It is the engineering of constraint. It means rigorous testing, transparent documentation, and the technical capacity to intervene the moment a system fails. It produces the traceability and accountability artifacts that prove a system behaves as intended. It is not sentiment. It is a practice of friction.

AI Governance is not a model checking itself. Governance only exists when a named human holds binding authority at specific checkpoints. If accountability cannot survive an audit or lead back to a specific desk, it is not governance. It is branding.

These three tiers form a stack, not a menu. Ethics provides the normative foundation. Responsible AI translates those norms into engineering discipline. Governance enforces human authority over both. Strip away any one, and the other two lose their structural integrity. Collapse all three into a single term, and what remains is a press release with no operational spine.

The Test Case: Constitutional AI

No single term in the current AI landscape demonstrates the language collapse more clearly than “Constitutional AI.” It is a real technical method. It is also a masterclass in branding that obscures where ethics ends, where engineering begins, and where governance never arrives.

Constitutional AI was introduced by Anthropic in a 2022 paper, “Constitutional AI: Harmlessness from AI Feedback,” as a training technique for large language models. The core idea is to replace much of the human feedback loop in model alignment with a written set of principles, a “constitution,” that the model uses to critique and revise its own outputs. Instead of thousands of human raters labeling every response as acceptable or harmful, the model asks itself: “Does this answer violate the principles?” It then generates a revised response that better conforms to those rules.

The method operates through two phases. In the first, the model receives adversarial prompts designed to provoke harmful outputs. It generates an initial response, critiques that response against a randomly selected constitutional principle, and produces a revised answer. Those revised answers become training data. In the second phase, the model (or a copy of it) compares pairs of responses to the same prompt and decides, using the constitution, which answer better follows the principles. That preference data trains a reward model, which then drives reinforcement learning. The result is a model that has internalized behavioral patterns aligned to the written rules.

This is genuine engineering work. It is rigorous. It is documented. It produces measurable improvements in the tradeoff between helpfulness and harmfulness. By every reasonable standard, Constitutional AI is a serious contribution to the practice of building safer AI systems.

It is Responsible AI. Tier 2. Full stop.

The Ethical Inheritance

The constitutional principles themselves draw from Tier 1. Anthropic’s published constitution references the UN Universal Declaration of Human Rights, internal policy commitments, and normative standards about harm, honesty, and human dignity. These are ethical inputs. They identify harm vectors. They encode judgments about where AI should and should not operate autonomously.

But the constitution is authored by a small internal team. It lacks democratic legitimacy. It has no institutional separation of powers. It does not require external ratification or stakeholder consent. The Digital Constitutionalist project has argued that Anthropic’s use of the word “constitution” borrows the weight of democratic governance without fulfilling any of its structural requirements. The principles are curated policy preferences, not a constitutional instrument in any meaningful political or legal sense.

This matters because the ethical layer of any AI system gains its legitimacy not from the quality of its principles alone, but from the process by which those principles are selected, contested, and revised. A constitution written by twelve engineers in San Francisco, however thoughtful, is not the same as a framework that has survived public deliberation.

Constitutional AI takes ethical inputs seriously. It does not constitute an ethical framework. It draws from Tier 1 without satisfying its requirements.

The Governance Gap

Here is where the category error becomes dangerous.

The word “constitutional” carries centuries of political meaning. Constitutions create binding authority. They establish separation of powers. They name who holds decision rights and under what constraints. They produce accountability that survives an audit and traces back to a specific desk.

Constitutional AI does none of this.

No named human holds binding per-decision authority at runtime checkpoints in the CAI pipeline. The model critiques itself. The model generates its own preference data. The model trains on its own feedback. Humans designed the system, chose the principles, and tuned the parameters. But at the operational stages where behavioral norms are established and enforced, human authority is precisely what has been reduced. That is the method’s stated advantage: scalability through the replacement of granular human judgment with AI self-evaluation. The operational locus of decision making moves into the model, even as the design authority remains with the team that wrote the constitution.

This is not a criticism of the engineering. It is a description of what the engineering does. And what it does is the opposite of governance.

An organization could, in theory, wrap Constitutional AI inside a real governance architecture, with named approvers, documented checkpoints, and auditable decisions about whether and how to deploy constitutionally trained models. But that governance would sit on top of CAI. It would not be CAI itself. The method does not preclude governance. It simply does not provide it. And the branding implies that it does.

Governance requires that accountability can survive contact with an external auditor. It requires that a specific person can be identified as the authority who approved a specific decision at a specific checkpoint. In the HAIA-RECCLIN framework, this is Identity Binding. The audit trail must connect an authenticated human identity to the decision record. Without that binding, there is no chain of custody for judgment. Constitutional AI disperses that authority into a training pipeline. The principles are inspectable, which is valuable. But inspectability is a property of Responsible AI, not governance. A transparent process without a named authority is still a process without governance.

Under regimes like the EU AI Act and prEN 18286, “governance” is not a metaphor. It is a set of documented, auditable responsibilities that can be assigned, challenged, and enforced. When the term is used to describe a training method that reduces human decision authority, the regulatory meaning is emptied.

Anthropic’s own researchers have acknowledged this structural reality. The 2022 paper positions CAI as an alignment technique, not a governance framework. But the branding, the word “constitutional,” does governance’s rhetorical work without doing governance’s operational work. Every time a vendor, consultant, or policy analyst uses “Constitutional AI” to suggest that a model is self-governing, the three-tier distinction collapses. Ethics becomes a word list. Responsible AI becomes a training loop. And governance disappears entirely, replaced by a metaphor borrowed from political theory and stripped of its structural meaning.

Human-in-the-Loop: When Feedback Replaces Authority

Constitutional AI borrows governance language from political theory. “Human-in-the-loop” borrows it from oversight itself. The term sounds like exactly what governance requires: a human, present and active, inside the decision process. In practice, it rarely delivers that.

At its most rigorous, human-in-the-loop means that humans hold binding authority at defined checkpoints, with clear specification of which decisions require approval, audit trails documenting who approved what and when, and override protocols when the human disagrees with the AI. That version of HITL would satisfy Tier 3 requirements. It would be governance.

That version almost never exists.

What “human-in-the-loop” usually means is that a human can review AI outputs. Not must review. Can. There is no specification of when review is mandatory versus optional. There is no named individual who holds binding authority. There is no audit trail documenting what was approved, by whom, and on what basis. The human is technically present but operationally decorative. Approval becomes rubber-stamping, accelerated by the automation bias that the EU AI Act explicitly names in Article 14(4)(b) as a risk requiring mitigation.

The International Association of Privacy Professionals has stated directly that requiring human oversight for AI systems is not a panacea, that it can lead to a false sense of security, and that at worst it can compound the risks. Solving AI risks with HITL, the IAPP argues, requires clearly defining the loop, specifying principles, and having metrics to assess results. Most implementations have none of this.

The problem is structural, not just operational. The word “loop” implies a cycle of collaboration, a continuous feedback relationship between human and machine. That metaphor describes partnership. Governance is not partnership. Governance is authority. The human who governs an AI system does not sit inside the workflow as a collaborator. That human sits outside and above it, controlling the gates through which AI work must pass before it reaches a decision with consequences. The distinction between “humans can improve AI outputs” and “humans control AI decisions” is the distinction between Tier 2 and Tier 3. The word “loop” erases it.

The EU AI Act does not ask for a feedback loop. Article 14(4)(d) requires that humans be able to “decide, in any particular situation, not to use the AI system or otherwise disregard, override or reverse the output.” That is override authority. It is the power to stop a system, not the opportunity to refine its outputs. When organizations claim HITL and deliver feedback without override authority, without named checkpoint owners, and without audit trails, they have answered a Tier 2 question while regulations are asking a Tier 3 question.

The Checkpoint-Based Governance model in the HAIA-RECCLIN framework addresses this directly by positioning human authority at defined gates (before, during, and after AI operational work) with named arbiters and documented decision records. The full architecture is detailed at BasilPuglisi.com. The point here is not to prescribe a specific implementation but to make visible what “human-in-the-loop” obscures: the difference between a human who participates and a human who decides.

Human-Centric AI: When Design Philosophy Claims Governance

If “Constitutional AI” borrows from political theory and “human-in-the-loop” borrows from oversight language, “human-centric AI” borrows from something even harder to pin down: good intentions.

Human-centric AI, as defined across industry and academic literature, means prioritizing human needs in design, building transparent and explainable systems, augmenting human capabilities rather than replacing them, and embedding ethical values in development. These are legitimate commitments. They reflect a genuine concern for the people who interact with AI systems. They belong in the conversation about how AI should be built.

They are Tier 1 ethical commitments and Tier 2 design practices. They are not governance.

The governance erosion follows the same pattern. “Human-centric” sounds like it answers the governance question. It sounds like the human is at the center of the decision architecture, holding authority over what the system does and does not do. In practice, human-centric AI specifies no checkpoint authority. It names no individual who holds binding decision rights. It produces no audit trail. It describes how a system should be designed, not who controls it after deployment.

When a vendor claims “human-centric AI” and a procurement officer reads that as “human oversight is built in,” the same damage occurs. A design philosophy has been mistaken for a governance structure. Tier 1 and Tier 2 language has claimed Tier 3 legitimacy.

The problem compounds when human-centricity slides into anthropomorphization, the practice of attributing human traits to AI systems. When interfaces describe AI as “understanding,” “thinking,” or “deciding,” they create a subtle inversion. Instead of centering AI on human authority, they center human trust on AI behavior. Users over-rely on outputs because the system sounds authoritative. Accountability diffuses because the framing suggests the AI itself bears some form of responsibility. The question “which human approved this?” never gets asked because the language implies the AI is the responsible party. This is not a design philosophy. It is authority laundering disguised as user experience.

Human-centric AI can inform Tier 1 values and improve Tier 2 engineering. It cannot substitute for Tier 3. A system designed around human needs is not the same as a system governed by human authority.

What Gets Lost

The consequences of this vocabulary drift are not abstract.

When organizations evaluate AI systems for deployment in high-stakes environments, healthcare, criminal justice, financial services, hiring, they need to know which tier of accountability applies. A system trained with Constitutional AI has engineering safeguards (Tier 2). A system marketed as human-in-the-loop may have feedback mechanisms (Tier 2) but no binding checkpoint authority (Tier 3). A system described as human-centric may reflect thoughtful design values (Tier 1) without any governance architecture at all.

If any of these systems produces a harmful outcome, the question is the same: does the audit trail lead to a named human who held binding authority at the critical checkpoint? For Constitutional AI, it leads to a training pipeline. For most HITL implementations, it leads to a review process no one was required to follow. For human-centric AI, it leads to a design philosophy with no enforcement mechanism.

The category error makes all three gaps invisible. When a procurement officer reads “Constitutional AI,” “human-in-the-loop,” or “human-centric AI” and concludes that governance is handled, each term has done its damage. The officer has been given Tier 1 and Tier 2 answers dressed in Tier 3 clothing.

This pattern extends beyond these three terms. “Responsible AI” has become a department name at most major technology companies, a team that publishes guidelines and reviews launches but holds no binding authority to halt a deployment. The engineering practice has been replaced by an organizational label. “AI Ethics Boards” proliferate across the industry, but most operate in an advisory capacity with no power to override a business decision. They offer counsel, not constraint. The distinction between an ethics board that advises and a governance body that binds is the distinction between Tier 1 and Tier 3, and most organizations have erased it. “Trustworthy AI” has become the emptiest vessel of all, broad enough to mean whatever the speaker needs it to mean, specific enough to sound credible in a slide deck, and accountable to no standard, no audit, and no named authority.

The collapse is systemic. Each term borrows legitimacy from governance language. Each delivers something valuable but insufficient: alignment, feedback, design, guidelines. Each creates false confidence that governance is handled. And each obscures where accountability actually sits.

Rebuilding the Distinctions

The fix is not to condemn any of these contributions as engineering or design. The fix is to enforce the vocabulary, and to operationalize it through separate, auditable artifacts.

The three-tier taxonomy proposed here (Ethical AI, Responsible AI, AI Governance) is operationalized in the HAIA-RECCLIN framework, developed through multi-platform review and documented operational practice at BasilPuglisi.com. These are not theoretical categories. They emerge from the operational reality that collapsing them produces systems no one can audit and no one can hold to account.

Every AI system deployed in a high-stakes environment should be evaluated against three distinct questions:

• Tier 1 (Ethical AI): What values does this system reflect, and who had standing to contest those values?
• Tier 2 (Responsible AI): What engineering practices ensure the system behaves as intended, and how are failures detected and corrected?
• Tier 3 (AI Governance): Which named human holds binding authority at which checkpoints, and can that authority survive an audit?

These questions cannot share a single answer. When they do, one or more tiers has been collapsed.

Constitutional AI lives in Tier 2 with Tier 1 inputs and no Tier 3 presence. Human-in-the-loop can reach Tier 3 if it specifies named authorities, mandatory checkpoints, audit trails, and override protocols, but in most implementations it remains Tier 2 feedback dressed in governance language. Human-centric AI occupies Tier 1 and Tier 2 as design philosophy and engineering practice, with no structural claim on Tier 3. Each is valuable. None is governance unless governance is built on top.

Organizations serious about accountability can enforce this by requiring three separate artifacts for any high-risk AI system: an Ethical AI Framework documenting harm vectors, agency protections, and the process by which values were selected and contested, signed by executive leadership and community stakeholders; a Responsible AI Engineering Log maintaining testing protocols, bias mitigation results, and version-controlled alignment documentation managed by technical teams; and an AI Governance Authority Matrix naming the specific humans who hold binding authority at each checkpoint, with override procedures enforceable by internal audit. If any of these three documents is missing, the system is not governed. If all three are collapsed into a single document, the distinctions have been erased.

The Bottom Line

Right now, the loudest voices in AI collapse these distinctions into a vague moral posture and call it leadership. Real ethics has no barrier to entry and no need for applause. It holds whether it is popular or not.

Most people are not arguing about governance yet. They are arguing about language. And the language is losing.

Constitutional AI is good engineering. Human-in-the-loop is a valuable feedback mechanism. Human-centric AI is a worthy design commitment. None of them is governance. Calling a training method “constitutional,” a feedback loop “oversight,” or a design philosophy “human-centered” does not make any of them a system of binding human authority, any more than calling a thermostat a democracy makes it one.

If your AI position depends on everyone agreeing with you, it is not leadership. It is marketing.

Governance begins where agreement ends. And right now, it has not even begun.

References

Bai, Y., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073. https://arxiv.org/abs/2212.08073

Anthropic. (2023). Claude’s Constitution. https://www.anthropic.com/news/claudes-constitution

European Parliament and Council. (2024). Regulation (EU) 2024/1689 (AI Act), Article 14: Human Oversight. Article 17: Quality Management System.

CEN-CENELEC. (2025). prEN 18286: Artificial Intelligence, Quality Management System for AI.

International Association of Privacy Professionals. (2022). Human-in-the-Loop in AI Risk Management: Not a Cure-All Approach. https://iapp.org/news/a/-human-in-the-loop-in-ai-risk-management-not-a-cure-all-approach

Digital Constitutionalist. (2024). On Constitutional AI. https://digi-con.org/on-constitutional-ai/

Fink, M. (2025). Human Oversight under Article 14 of the EU AI Act. SSRN.