SMAT uses configured OpenAI services for text generation, assistant functions and selected image work, and configured Google services for selected image generation, embeddings and visual analysis. The precise route depends on the feature. Anthropic is not an active fallback. Generated material may be inaccurate, incomplete or unsuitable and should be checked for context, rights and platform rules.

When a person interacts directly with the SMAT assistant, the interface discloses at the first interaction that it is an automated AI assistant unless the context already makes that obvious. This control is separate from marking generated media and from disclosure of public-interest text.

The current policy distinguishes ordinary generated media, materially generated or manipulated image, audio or video that can appear authentic, and standard corrections of user-provided media. Deepfake-like image or video receives a clear on-asset disclosure. Standard corrections do not receive an AI label solely because automated tools assisted them. Ordinary generated media is subject to the applicable machine-readable marking control, but SMAT does not claim production proof of that control until the separate marking gate is evidenced.

AI-generated or manipulated text published to inform the public on matters of public interest receives a clear disclosure at the beginning of the text unless a lawful exception is established for that exact use. SMAT does not rely on a blanket human-review exception and does not promise that every output is manually reviewed.

Where perceivable disclosure is required, the current fail-safe is an on-asset label for image or video and a leading disclosure for public-interest text. SMAT does not claim that a channel-native label is available or preserved without production proof for that channel. User uploads and ordinary corrections are not relabelled unless their nature or use independently triggers a disclosure duty.

Owner decision: grace_50_2: true. The grace is limited to the existing text and image generation or editing systems and their unchanged intended purpose documented before 2 August 2026. It does not cover a new system, a new modality, a materially changed purpose or any use after 2026-12-02.

During this bounded period, the current Article 50(1) direct-interaction disclosure and Article 50(4) perceivable disclosure controls continue to apply now. The grace is not a waiver of those controls and is not evidence that a system qualifies. New or changed capabilities require a fresh assessment before release.

The public disclosure states implemented controls and conservative fallbacks only. It does not claim verified C2PA preservation, SynthID detection, channel-native marking, provider no-training terms or a production machine-marking receipt. Those claims require dated evidence before publication. Questions about a specific output or disclosure can be sent to [email protected] with the account and content reference.