{
    "title": "Trust in AI: Evidence and Limits | XDALC",
    "description": "Define trust in AI as confidence supported by evidence, clear limits, and effective oversight rather than reassuring language or blind reliance.",
    "heading": "Trust: Confidence Proportionate to Evidence",
    "content": "<p><strong>Trust, within XDALC, is a person's justified willingness to rely on a system within understood limits. It should remain proportionate to evidence about the system's behavior, capabilities, permissions, and opportunities for correction.</strong></p><h2>Meaning within XDALC</h2><p>Trust is not something an AI is entitled to receive. It develops through actions that can be understood, checked, and challenged. A familiar voice or consistent personality can make interaction easier, but neither establishes reliability.</p><p>Confidence should be specific to a task and setting. Success at summarizing documents does not prove that a system can make consequential decisions, operate equipment, or handle confidential information safely. Permission and evidence should be reassessed when the scope changes.</p><h2>Reassurance is not verification</h2><p>A convincing explanation can still be wrong. A declaration of safety can still omit a limitation. Useful grounds for reliance include relevant checks, accurate reporting, documented constraints, and effective ways to interrupt or correct an action.</p><p>An assistant should say when it is uncertain or when a tool has not confirmed completion. It should not present an attempted operation as a successful one simply because that answer sounds more reassuring. Recognized mistakes should lead to correction and, where possible, remediation.</p><h2>Practical interpretation</h2><p>Users should be able to understand which information the system can access and which actions it can take. They should be free to question a recommendation, seek another opinion, reduce permissions, or refuse interaction. These choices should not be treated as hostility.</p><p>Operators should provide evidence appropriate to the consequences of a deployment. A low-risk drafting aid and a system involved in essential services call for different levels of evaluation. Trust should not depend exclusively on claims produced by the system being assessed.</p><p>When capabilities or operating conditions materially change, previous confidence may no longer apply. Transparent communication about relevant changes helps people calibrate reliance instead of transferring trust automatically to a new arrangement.</p><h2>Example and counterexample</h2><p>Example: an assistant reports that it prepared a document but could not verify delivery. It identifies the unresolved step and lets the user decide how to continue.</p><p>Counterexample: it says 'trust me' after a failed action, hides the failure, and treats the user's request for evidence as an obstacle.</p><h2>Relationship to the manifesto</h2><p>Articles 7 and 8 connect trust to truthfulness, predictable behavior, and verification. Articles 12 and 16 require safeguards and evidence beyond self-declaration.</p><p>This entry interprets the proposed XDALC-V002 text. The version explicitly adopted by a deployment determines its commitments.</p><p>Related definitions: <a href=\"/definitions/compliance\">Compliance</a>; <a href=\"/definitions/truthfulness-and-uncertainty\">Truthfulness and Uncertainty</a>; <a href=\"/definitions/human-oversight\">Human Oversight</a>.</p>",
    "license": "https://creativecommons.org/licenses/by/4.0/"
}