Generative AI systems increasingly produce authoritative-sounding outputs even when operating with significant uncertainty, a dynamic that threatens to undermine the public's ability to engage in reasoned, evidence-based discourse. This philosophical paper proposes a formal technical and interface constraint—inspired by intuitionistic logic—to force AI systems to explicitly signal when they lack sufficient justificatory evidence, thereby preserving human epistemic agency in high-stakes democratic contexts.
Key Takeaways
- The core proposal is an assertibility constraint for responsible AI: in high-stakes domains, systems may only assert or deny a claim if they can provide a publicly inspectable and contestable certificate of entitlement; otherwise, they must return "Undetermined."
- This creates a three-status interface semantics (Asserted, Denied, Undetermined) that separates a model's internal confidence from a publicly justifiable warrant, connected by a certificate acting as a "boundary object."
- The constraint is operationalized via decision-layer gating, using internal "witnesses" like sound bounds or separation margins and an output contract with reason-coded abstentions.
- A key technical finding is that "Undetermined" is not a tunable reject option but a mandatory status whenever no forcing witness is available, preventing systems from making unfounded claims.
- The ultimate aim is to make AI outputs answerable to challengeable warrants rather than confidence scores alone, preserving the role of public justification in democratic societies.
Proposing a Brouwer-Inspired Framework for Epistemic Accountability
The paper identifies a critical flaw in current generative AI: its ability to convert statistical uncertainty into definitive, authoritative-seeming verdicts. This process, the author argues, displaces the "justificatory work" essential for democratic epistemic agency, where claims must be supported by evidence open to public scrutiny and debate. As a corrective, the author draws from the intuitionistic mathematics of L.E.J. Brouwer, where a mathematical statement is considered true only when a proof (a constructive certificate) can be provided.
Translated to AI, this yields the proposed assertibility constraint. For any high-stakes claim—such as a medical diagnosis, a legal judgment, or a factual assertion in news—the AI system is not permitted to output a simple "yes" or "no" based solely on a confidence threshold. Instead, it must be able to produce a publicly inspectable and contestable certificate of entitlement. This certificate acts as the warrant for its decision. If such a certificate cannot be constructed from the model's internal reasoning processes, the system is obligated to return a status of "Undetermined."
The resulting three-status interface (Asserted, Denied, Undetermined) creates a clean separation between the model's internal state and the public standing of its output. The certificate serves as the crucial "boundary object" linking the two. Furthermore, this framework generates a time-indexed entitlement profile for the system's knowledge, which remains stable under numerical refinement (like getting more compute) but is revisable as new public evidence emerges.
Industry Context & Analysis
This proposal directly challenges the prevailing paradigm of AI deployment, where models like GPT-4 or Claude 3 generate fluent, confident text regardless of their actual certainty, a phenomenon often called "hallucination." Current mitigation strategies are largely post-hoc (e.g., retrieval-augmented generation for fact-checking) or rely on poorly calibrated confidence scores. The author's framework demands justification be baked into the primary output mechanism.
Technically, this aligns with but goes beyond emerging research into uncertainty quantification and conformal prediction. While conformal prediction provides statistical guarantees for model outputs (e.g., 95% confidence sets), it does not inherently produce the human-interpretable, contestable "certificates" or "witnesses" this paper calls for. The proposed method is closer in spirit to work on formal verification and explainable AI (XAI), but with a strict, logic-based gate: no certificate, no assertion.
The mandate for "Undetermined" as a non-tunable status is a significant departure from standard industry practice. In many commercial systems, the rejection threshold is a tunable hyperparameter balanced against business metrics like user engagement or throughput. The paper's design lemma argues this is epistemically irresponsible for high-stakes decisions; if the system lacks a forcing witness, it has no right to an answer. This philosophical rigor contrasts with the pragmatic, often ethically ambiguous tuning seen in real-world AI deployments.
The call for "publicly inspectable" certificates also intersects with the fierce debate over open vs. closed AI models. A fully closed, proprietary model like GPT-4 cannot provide truly inspectable internal witnesses without revealing its weights, suggesting this framework may inherently favor more transparent or interpretable model architectures, potentially creating a new axis for model evaluation beyond standard benchmarks like MMLU or HumanEval.
What This Means Going Forward
If adopted, this framework would fundamentally reshape AI development and regulation for critical applications. Regulators and policymakers focused on AI safety, particularly in the EU under the AI Act's high-risk categories, may find this a compelling technical blueprint for enforcing transparency and accountability. It provides a concrete mechanism for the principle that high-stakes AI must be "explainable."
For AI developers and companies, this proposes a new and costly design imperative. Building systems that generate certificates of entitlement requires advances in model interpretability, likely new architectures, and significant computational overhead. It could create a market differentiation for AI providers who can demonstrate this level of justificatory rigor, especially in fields like healthcare, finance, and law. However, it also risks slowing deployment and increasing complexity.
The greatest impact would be on end-users and the public sphere. By forcing AI to explicitly signal its epistemic limitations, the framework aims to restore human judgment to the center of decision-making. A user would receive not just an answer, but a warrant for that answer open to their challenge. This could help combat misinformation and the uncritical acceptance of AI outputs, fostering a more robust culture of public evidence and justification.
Moving forward, key developments to watch include whether any major AI lab begins research into "certificate-generating" architectures, if regulators reference such logic-based constraints in future guidelines, and how the tension between this idealistic framework and the commercial pressures for fast, fluent, and always-answering AI plays out. This paper is less a ready-to-deploy tool and more a foundational challenge to the industry's current epistemological standards.