Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.18223207 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866901645469679616 |
|---|---|
| author | GA, BIRI |
| author_facet | GA, BIRI |
| contents | <h1><strong>When Should an AI Be Allowed to Answer?</strong></h1> <p>Most large language models are evaluated by how well they <em>respond</em>. This work focuses on a quieter but more consequential question: <span><strong>when should an AI respond at all?</strong></span></p> <p>We present a <span><strong>verifiable governance protocol</strong></span> that grants an AI system the <em>Right to Answer</em> only when a <span><strong>minimal evidence chain</strong></span> satisfies explicit policy constraints. Model outputs are treated as untrusted proposals; answer delivery is a privileged operation enforced by a minimal <span><strong>governance kernel</strong></span>.</p> <p>At the core of the protocol is <span><strong>Reverse Evidence Chain Inference (RECI)</strong></span>, which frames governance not as a matter of confidence scores, but as a <span><strong>feasibility problem</strong></span>: determining whether there exists a smallest possible set of evidence that meets coverage thresholds, consistency constraints, and risk budgets. If such a set exists, delivery is authorized; if not, the system is required to ask or abstain.</p> <p>The protocol further specifies an <span><strong>audit interface</strong></span>, including environment fingerprints, structured reason codes, and deterministic replay commands, enabling post hoc verification of every delivery decision.</p> <p>No new learning algorithm is proposed; the contribution lies at the <span><strong>protocol and system boundary</strong></span>, where accountability is enforced not by trust, but by replayable constraints.</p> <p>This record targets high-stakes AI deployments—such as medicine, law, and public services—where answering the wrong question confidently is often worse than not answering at all.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18223207 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | When Can an AI Answer? GA, BIRI <h1><strong>When Should an AI Be Allowed to Answer?</strong></h1> <p>Most large language models are evaluated by how well they <em>respond</em>. This work focuses on a quieter but more consequential question: <span><strong>when should an AI respond at all?</strong></span></p> <p>We present a <span><strong>verifiable governance protocol</strong></span> that grants an AI system the <em>Right to Answer</em> only when a <span><strong>minimal evidence chain</strong></span> satisfies explicit policy constraints. Model outputs are treated as untrusted proposals; answer delivery is a privileged operation enforced by a minimal <span><strong>governance kernel</strong></span>.</p> <p>At the core of the protocol is <span><strong>Reverse Evidence Chain Inference (RECI)</strong></span>, which frames governance not as a matter of confidence scores, but as a <span><strong>feasibility problem</strong></span>: determining whether there exists a smallest possible set of evidence that meets coverage thresholds, consistency constraints, and risk budgets. If such a set exists, delivery is authorized; if not, the system is required to ask or abstain.</p> <p>The protocol further specifies an <span><strong>audit interface</strong></span>, including environment fingerprints, structured reason codes, and deterministic replay commands, enabling post hoc verification of every delivery decision.</p> <p>No new learning algorithm is proposed; the contribution lies at the <span><strong>protocol and system boundary</strong></span>, where accountability is enforced not by trust, but by replayable constraints.</p> <p>This record targets high-stakes AI deployments—such as medicine, law, and public services—where answering the wrong question confidently is often worse than not answering at all.</p> |
| title | When Can an AI Answer? |
| url | https://doi.org/10.5281/zenodo.18223207 |