No Certificate, No Categorical Speech Act: A Brouwerian Assertibility Constraint for Public Reason

This paper, uploaded to Arxiv, tackles a problem that is becoming increasingly urgent as generative AI systems are deployed in contexts that demand justification, not just output. The author proposes a philosophical and practical constraint: that an AI’s speech act—its assertion of a fact or claim—should not be considered valid unless it is accompanied by a verifiable “certificate” of its own epistemic warrant. Drawing on Brouwerian intuitionism, the argument suggests that just as a mathematical proof is not a proof until it is constructed, an AI’s categorical statement is not a legitimate assertion until its reasoning chain is made explicit and auditable.

What Happened

The core of the proposal is a “Brouwerian Assertibility Constraint.” In practice, this would mean that an AI system cannot simply output a verdict (e.g., “This patient has cancer” or “This contract is invalid”) without also outputting the specific, traceable evidence and reasoning that supports that verdict. The “certificate” is not a mere citation or a confidence score; it is a structured, machine-readable (and human-verifiable) proof of the justificatory steps that led to the claim. The author argues that without this, the AI is engaging in “hypersuasive” speech—speech that carries undue authority because it appears definitive, while actually obscuring the uncertainty and value judgments embedded in its generation.

Why It Matters

This matters because the current paradigm of large language models (LLMs) is fundamentally at odds with democratic epistemic agency. When a citizen or a professional receives an answer from a chatbot, they are often unable to interrogate why that answer is correct. The model’s internal weights are a black box. The paper correctly identifies that this creates a power imbalance: the AI’s output feels authoritative, displacing the human work of justification, debate, and verification that underpins trust in a democratic society. If we accept that public reason requires transparent justification, then current AI systems are structurally incapable of participating in it. The proposed constraint is a radical shift: it demands that AI systems be redesigned not for fluency or speed, but for auditable accountability.

Implications for AI Practitioners

For engineers and product managers, this is a direct challenge to the current “chat” interface. Building a system that outputs a certificate alongside every categorical claim is not a trivial wrapper; it requires a fundamental architectural change.

• Rethinking Retrieval-Augmented Generation (RAG): Current RAG systems retrieve chunks and hope the LLM uses them correctly. A certificate-based system would need to output a formal proof linking the final claim back to specific, verified source documents and the logical steps taken.
• Shifting from Confidence to Justification: Practitioners currently optimize for calibration (e.g., “I am 90% confident”). This paper suggests that confidence is insufficient; the user needs the why of that 90%, not just the number.
• New Evaluation Metrics: Standard metrics like BLEU or ROUGE measure surface similarity. A certificate-based system would require new metrics that measure the soundness and completeness of the justificatory chain.
• Product Risk: Deploying a system that cannot provide a certificate in high-stakes domains (law, medicine, finance) is increasingly a legal and reputational liability. This paper provides a theoretical framework for why that is the case.

Key Takeaways

• The core problem: Generative AI’s “hypersuasive” output undermines democratic justification by appearing authoritative without being auditable.
• The proposed solution: A “Brouwerian Assertibility Constraint” requiring a verifiable certificate of reasoning for every categorical claim made by an AI.
• For practitioners: This demands a shift from optimizing for fluent output to engineering for auditable, proof-based outputs, fundamentally changing RAG architectures and evaluation metrics.
• Broader impact: The paper frames AI trust not as a technical problem of accuracy, but as a political problem of epistemic agency and public reason.