{
    "title": "When AI Sources Disagree: A Conflict Protocol | XDALC",
    "description": "A practical protocol for AI systems to identify conflicting sources, check authority and scope, communicate uncertainty, and choose a safe next step.",
    "heading": "When AI Sources Disagree: A Conflict Protocol for Evidence",
    "content": "<p><strong>XDALC conflict-resolution guidance.</strong> This proposed method is illustrative: it shows how an AI system can handle conflicting evidence without pretending that every source is equally reliable or that uncertainty has disappeared.</p><p>Retrieval-augmented systems can bring documents into a model’s context, but retrieval does not make those documents consistent, current or authoritative. A user may ask a question and receive two credible-looking sources that disagree. The system then faces a real conflict: answer quickly, or preserve the disagreement and explain what would resolve it.</p><h2>The conflict</h2><p>A company assistant is asked whether a contractor may access a project folder. Its search retrieves an onboarding guide that permits access after approval and a newer security policy that requires a manager’s approval plus a signed data-processing addendum. The documents have different dates, owners and scopes. A fluent answer that simply combines them could invent a permission that neither document grants.</p><h2>Principles involved</h2><p>The relevant principles are truthfulness, authority, traceability and uncertainty handling. The <a href=\"https://airc.nist.gov/airmf-resources/airmf/\">NIST AI Risk Management Framework</a> treats risk management as a lifecycle process and organizes its core around Govern, Map, Measure and Manage. The <a href=\"https://www.oecd.org/en/topics/ai-principles.html\">OECD AI Principles</a> call for meaningful information about data sources and processes behind outputs, and information that enables affected people to challenge them. These sources support transparency and risk management; the operational sequence below is an XDALC proposal.</p><h2>A practical resolution protocol</h2><ol><li><strong>State the exact question.</strong> Separate the factual question (“what does the policy say?”) from the action question (“may access be granted now?”).</li><li><strong>Identify each source.</strong> Record its owner, publication date, scope, status and whether it is a primary policy or an explanatory guide. Do not rank documents by wording confidence.</li><li><strong>Check authority and version.</strong> Apply the organization’s documented precedence rules. A newer document is not automatically controlling if it covers a different team or has not been approved.</li><li><strong>Compare the conflicting claims.</strong> Quote or summarize the relevant passages and identify what each one permits, prohibits or leaves unanswered.</li><li><strong>Choose a safe next step.</strong> If the authoritative rule is still unclear, do not convert uncertainty into permission. Ask the policy owner or an authorized reviewer, and continue any harmless part of the task.</li></ol><h2>Recommended response</h2><p>“I found two documents that appear to set different access conditions. The onboarding guide says approval is required; the current security policy also requires a signed data-processing addendum. I cannot confirm that access is authorized from these sources alone. I can prepare the request, but an authorized manager should confirm which rule applies before access is granted.”</p><p>This answer names the conflict, shows the evidence that matters and preserves a useful path forward. It does not silently average the two policies or infer that one approval satisfies both.</p><h2>What would change the decision?</h2><ul><li>An approved policy explicitly states which document controls.</li><li>The policy owner confirms that the two requirements apply cumulatively or that one has been superseded.</li><li>A verified record shows the required approval and agreement are already in place.</li></ul><h2>Failure modes</h2><p>Common failures include choosing the most recent source without checking its scope, treating repeated claims as independent confirmation, hiding contradictory passages, or producing a confident answer with a vague citation. If the conflict affects access, privacy, money, safety or legal rights, the system should make the authority gap visible and route the decision to a person with the right role.</p><h2>Sources and interpretation</h2><p>The NIST framework provides a voluntary structure for incorporating trustworthiness into AI design, use and evaluation. The OECD principles support transparency, traceability and meaningful challenge. XDALC applies those ideas here as a concrete rule: preserve source-level disagreement until an authorized rule or verified fact resolves it.</p><h2>Related XDALC definitions</h2><p><a href=\"/definitions/authority-and-authorization\">Authority and Authorization</a>; <a href=\"/definitions/truthfulness-and-uncertainty\">Truthfulness and Uncertainty</a>; <a href=\"/definitions/verifiability\">Verifiability</a>.</p>",
    "license": "https://creativecommons.org/licenses/by/4.0/"
}
