Upholding Epistemic Agency: A Brouwerian Assertibility Constraint for Responsible AI

This paper proposes a Brouwerian assertibility constraint for responsible AI in high-stakes domains that replaces probabilistic confidence with a three-status interface (Asserted, Denied, Undetermined), requiring systems to provide publicly inspectable certificates of entitlement before making claims to preserve democratic epistemic agency.

The Gist

Here is an explanation of the paper "Upholding Epistemic Agency: A Brouwerian Assertibility Constraint for Responsible AI" using simple language, analogies, and metaphors.

The Big Problem: The "Smooth-Talking" Robot

Imagine you are in a town square, and a new robot has been hired to give advice on serious matters like health, law, and politics. This robot is incredibly smooth. It speaks with total confidence, using perfect grammar and a calm voice.

However, there's a catch: The robot doesn't actually know the truth. It's just guessing based on patterns it saw in its training data. It might say, "The mayor is definitely corrupt," with 99% confidence, even if it has no proof. Or it might say, "This medicine is safe," when it's actually dangerous.

The problem is that because the robot sounds so confident, people stop asking questions. They stop doing their own thinking. They just accept the robot's verdict. The paper calls this a loss of "Epistemic Agency." In simple terms, it means we are losing our ability to be the judges of what is true and what is false. We are becoming passive listeners instead of active thinkers.

The Proposed Solution: The "Certificate of Proof" Rule

The author, Michael Jülich, suggests a new rule for these robots, inspired by a mathematician named L.E.J. Brouwer.

The Rule: "No Certificate, No Verdict."

Instead of the robot just saying "Yes" or "No," it must follow a strict three-step process before it is allowed to speak:

1. Can it prove it? The robot must be able to produce a "Certificate of Entitlement." This isn't just a confidence score (like "I'm 90% sure"). It's a concrete, checkable piece of evidence (like a math proof, a verified document, or a clear logical step) that shows why it is allowed to make that claim.
2. Is the proof public? Anyone (or a human auditor) must be able to look at this certificate and say, "Yes, this proof holds up," or "No, this proof is weak."
3. The Third Option: If the robot cannot produce a solid certificate, it is forbidden from saying "Yes" or "No." Instead, it must say: "Undetermined."

The Three-Status Interface: A Traffic Light Analogy

Think of the robot's output like a traffic light, but with a twist:

• 🟢 Green (Asserted): The robot has a solid certificate. It can say, "Yes, this is true," and hand you the proof.
• 🔴 Red (Denied): The robot has a solid certificate proving the opposite. It can say, "No, this is false," and hand you the proof.
• 🟡 Yellow (Undetermined): The robot is stuck. It doesn't have enough proof to say "Yes" or "No." Crucially, it cannot guess. It must stop and say, "I don't have enough evidence to decide yet."

Why is the Yellow light so important?
In current AI, when a system is unsure, it often just picks the most likely answer and says it with confidence. This paper argues that in high-stakes situations (like courts or hospitals), guessing is dangerous. The "Yellow" light forces the system to admit uncertainty, which keeps humans in the loop to do the real thinking.

The "Tooth Social" Story: A Real-World Example

The paper uses a fictional story to explain how this works. Imagine a scandal involving a government minister who might have taken bribes.

• Phase 1 (The Rumors): A journalist reports rumors. The AI checks the evidence. The evidence is messy; some documents support the rumor, others don't. The AI's "proof" is shaky.

  • Old AI: "It's likely the minister is corrupt." (People believe it immediately).
  • New AI (Brouwerian): "Undetermined." It explains: "The evidence is mixed. I cannot prove it yet. Please wait for an official investigation."
  • Result: The public stays calm and waits for facts. The AI didn't start a riot based on a guess.

• Phase 2 (The Official Report): A congressional committee releases a report with sworn testimony and hard proof. The AI gets a new "certificate."

  • New AI: "Asserted." It says: "Based on the official report (Certificate Attached), the minister is corrupt."
  • Result: The verdict is now backed by a public, checkable record.

Why This Matters: The "Adult Education" Metaphor

The author compares this to education for grown-ups.

• The Old Way: The AI acts like a know-it-all teacher who just gives you the answer key. You memorize the answer, but you don't learn how to think.
• The New Way: The AI acts like a strict debate partner. It says, "I can only say this if I can show you my work." If it can't show its work, it stays silent.

This forces us (the humans) to do the work. We have to look at the certificate, check the evidence, and decide if we agree. It stops the AI from doing our thinking for us.

The Catch: It's Hard to Build

The paper admits this is difficult.

• It's slower: Checking for certificates takes more computing power than just guessing.
• It's "boring": The AI will say "Undetermined" a lot more often than current AIs.
• It's not magic: The AI still doesn't "know" the truth in a human sense. It just follows strict rules about when it is allowed to speak.

The Bottom Line

This paper proposes a new "constitution" for AI. It says: In important matters, confidence is not enough. You need proof.

If an AI cannot show its homework, it must admit it doesn't know. This protects us from being misled by smooth-talking machines and ensures that we, the humans, remain the ones in charge of our own knowledge and democracy. It turns the AI from a "verdict-giver" into a "proof-provider," keeping the power of judgment in human hands.

Technical Summary

Here is a detailed technical summary of the paper "Upholding Epistemic Agency: A Brouwerian Assertibility Constraint for Responsible AI" by Michael Jülich.

1. Problem Statement

The paper addresses the erosion of epistemic agency in democratic societies caused by Generative AI (GenAI).

• The Core Issue: GenAI systems convert probabilistic patterns into authoritative-seeming categorical verdicts (e.g., "Yes/No" or specific labels). This displaces the justificatory work required for democratic deliberation, effectively "deskilling" citizens by presenting uncertain or hallucinated outputs as settled facts.
• The Limitation of Current Approaches: Existing methods for handling uncertainty (e.g., confidence scores, semantic entropy, conformal prediction) treat uncertainty as a statistical property. They often allow systems to output a "best guess" even when the evidence is insufficient, or they treat abstention as a tunable reject option based on a confidence threshold.
• The Gap: There is no normative framework that distinguishes between warranted categorical speech and unwarranted fluency. Current systems lack a mechanism to force them to admit when they lack a verifiable basis for a claim, leading to "Frankfurtian bullshit" (speech without commitment to truth).

2. Methodology: The Brouwerian Framework

The author proposes a design constraint inspired by L.E.J. Brouwer's Mathematical Intuitionism, adapted for AI systems in high-stakes domains (e.g., law, health, public policy).

A. Philosophical Foundations

• Constructive Truth: Truth is not a pre-existing property but is established through constructions (proofs/witnesses). A statement is assertible only if a checkable construction exists.
• Rejection of the Law of Excluded Middle (LEM): For infinite or open-ended domains, a proposition $P$ is not necessarily true or false until a proof or refutation is constructed.
• Temporal Entitlement: Entitlement to assert a claim is time-indexed and evolves as new evidence (constructions) becomes available.

B. The Core Constraint: "No Certificate, No Verdict"

The paper introduces the Brouwerian Assertibility Constraint:

In high-stakes domains, an AI system may assert or deny a claim only if it can provide a publicly inspectable and contestable certificate of entitlement. Otherwise, it must return Undetermined (U).

C. Technical Architecture

The framework operationalizes this constraint through a Dual-Regime Entitlement Structure:

1. Internal Regime ($\langle R, t_{int} \rangle$):
   • Focuses on Model-Claims ($\psi$).
   • Requires a constructive witness (e.g., certified interval bounds, separation margins, or formal proofs) generated by the system's internal toolchain.
   • The system is entitled to assert $\psi$ only if a verifier accepts the witness.
2. Public Regime ($\langle S, \tau, t \rangle$):
   • Focuses on Warrant-Claims ($W(\phi)$).
   • Requires public standing: provenance, scope, and contestability of the evidence relative to a public record at time $t$.
   • The system is entitled to assert the world-claim $\phi$ only if the internal witness satisfies the public standard $S$ and threshold $\tau$.

The Boundary Object (Certificate $\kappa$):
A certificate acts as a bridge between the two regimes. It is a tuple containing:

• Witness: The internal computational proof (e.g., $[l, u]$ bounds).
• Standing: Metadata regarding scope, provenance, and time.
• Checkability: An auditor-accessible procedure CheckC(κ) that verifies the witness and standing.

3. Key Contributions

A. Three-Status Interface Semantics (A/D/U)

The paper defines a new interface semantics that replaces binary outputs with three statuses:

• Asserted (A): A certificate exists that forces the claim (e.g., lower bound $l \ge \tau$).
• Denied (D): A certificate exists that forces the negation (e.g., upper bound $u < \tau$).
• Undetermined (U): No forcing certificate is available within the declared budget and scope.
  • Crucial Distinction: U is not a tunable reject option based on low confidence. It is a mandatory status whenever the constructive discipline fails to produce a witness. It represents a refusal to speak without warrant.

B. The Design Lemma (Decidability Constraint)

The author proves a formal lemma (Proposition 1):

• Premise: If an interface is total (always outputs A or D) and certificate-sound (outputs are backed by valid witnesses), then the deployed predicate must be decidable on its declared scope.
• Implication: Since many real-world predicates (e.g., threshold comparisons on constructive reals) are not uniformly decidable without additional structure (like separation margins), a system that forces a binary A/D output without a witness is epistemically dishonest. It fabricates a verdict where none is warranted. Therefore, U is structurally necessary to prevent "underspecified" interpretability.

C. Operationalization via Hybrid Architecture

The paper proposes a Hybrid Implementation for practical deployment:

• Classical Core: An autoregressive LLM generates fluent, probabilistic outputs.
• Constructive Wrapper (Speech-Act Gate): A verification layer (using Interval Bound Propagation, abstract interpretation, or conformal prediction) sits atop the core.
  • It attempts to construct a forcing witness for the core's output.
  • If a witness is found within the budget ($B_{budget}$), it outputs A or D with a trace.
  • If not, it forces the system to output U with a reason code (e.g., U-MODEL, U-EVIDENCE, U-SCOPE).

D. Case Study: "Tooth Social"

A political chatbot scenario demonstrates the framework.

• Scenario: A corruption scandal involving a minister.
• Stage 1 (Early): Evidence is sparse. The certified interval straddles the corruption threshold. The system returns U, explaining that no official finding exists yet.
• Stage 2 (Post-Inquiry): New evidence (congressional report) shifts the certified interval above the threshold. The system returns A with a certificate.
• Significance: The system maintains U even if the internal "confidence" is high, until the public record and constructive bounds align.

4. Results and Findings

• Epistemic Honesty: The framework prevents systems from presenting "false certainty." By mandating U when witnesses are missing, it aligns machine speech with the epistemic standards of human deliberation.
• Stability: The U status is stable under numerical refinement (e.g., increasing model precision) but revisable as the public record changes. This ensures that the system's "truth" tracks the evolution of public knowledge, not just internal model weights.
• Anti-Deskilling: By requiring reason codes for U (e.g., "Evidence insufficient," "Scope mismatch"), the system guides users on what information is needed to reach a verdict, actively preserving human epistemic agency rather than replacing it.

5. Significance and Implications

• Normative Shift: Moves AI responsibility from "accuracy/alignment" to justification/entitlement. It treats AI not just as a predictor, but as a participant in public justification that must adhere to the same rules of evidence as human speakers.
• Political Philosophy: Bridges the gap between Intuitionistic Logic and Democratic Theory. It operationalizes the "Epistemic Turn" in political theory by making truth a regulative ideal for machine speech, ensuring that automated systems do not usurp the role of citizens in deliberation.
• Practical Design: Provides a concrete architectural blueprint (Hybrid Wrapper) for implementing these norms without requiring fully constructive neural networks, making it feasible for near-term deployment in high-stakes sectors (law, medicine, elections).
• Limitations: Acknowledges computational costs (verification is expensive) and the risk of over-abstention. However, argues that in high-stakes domains, the cost of unwarranted confidence outweighs the cost of abstention.

Conclusion

The paper argues that for AI to be responsible in a democracy, it must stop simulating authority and start practicing epistemic restraint. By adopting a Brouwerian constraint, AI systems are forced to exhibit their warrants or admit ignorance (Undetermined), thereby preserving the space for human judgment, contestation, and the collective pursuit of truth.