Drupal AI Disclosure Beta Adds Spoken Audio Notices and Video Labels

Publishers using Drupal can now attach spoken AI disclosures to audio recordings and embed visible disclosure icons in video files through the AI Disclosure module. The functionality is included in version 1.0.0-beta1, released on 12 October 2026. Drupal developer Giorgio Alfredo Pagano, who created the module within the Drupal AI Initiative, detailed its audio implementation in a technical article published on 21 October 2026.

The additions address a distinction between identifying AI-generated content on a website and preserving that disclosure when the content is downloaded or redistributed. A visual label displayed beside an audio player does not accompany the sound file, while a browser overlay on video is absent from the downloaded recording. AI Disclosure addresses these limitations by generating separate labelled copies of qualifying audio and video files, while retaining visible notices on Drupal pages.

For audio classified as requiring a deepfake disclosure, the optional ai_disclosure_av submodule can generate a copy that begins with a spoken statement identifying the content as AI-generated or AI-modified. According to the project documentation, the notice is followed by a 0.4-second pause before the original recording begins. The Drupal page also displays an EU disclosure icon and explanatory text above the audio player.

Administrators prepare the spoken notices for each active site language and for the two disclosure categories. The module provides three ways to supply a notice: uploading an existing recording, recording the statement through the browser, or generating speech through a compatible text-to-speech provider connected to the Drupal AI module. Speech generation requires an explicit action by an administrator and is not performed automatically. Clips are checked before use, including an eight-second duration limit and checks for silence or missing audio.

Editors can select up to three available languages for an audio file, with the notices played consecutively. Where no language is selected, the module uses the media language or a configured fallback. Longer recordings can repeat the spoken notice every 300 seconds by default, with optional short tones before reminders. The audio rendition retains the original recording after the inserted declarations, allowing the disclosure to remain audible when the labelled copy is downloaded.

The generated files also carry machine-readable information. The renderer writes a disclosure sentence and an IPTC Digital Source Type identifier into supported audio formats, using ID3 metadata for MP3, MP4 tags for M4A, and Vorbis comments for FLAC, Ogg and Opus. WAV files retain the disclosure sentence but not the additional custom source-type field. These metadata fields are distinct from cryptographically signed C2PA Content Credentials.

The module additionally supports optional inaudible audio watermarking through the external audiowmark utility. This feature is intended for sites that also generate synthetic audio and therefore may have provider obligations under Article 50(2) of the EU AI Act. It embeds a 128-bit identifier derived from the site and file identifiers. The watermark is separate from the spoken declaration and does not replace the publisher-facing disclosure.

Video labelling is handled through the same optional submodule. For qualifying deepfake footage, FFmpeg can create an MP4 copy with the EU disclosure icon embedded directly in the image. The default configuration displays the icon during the first five seconds and repeats it for five seconds every 60 seconds. Editors can adjust its placement and specify individual time ranges for videos containing only selected manipulated segments. The renderer also examines the brightness of nearby frames to select a contrasting black or white icon.

Externally hosted videos use a different method. When video is embedded through Drupal's oEmbed formatter, the module can display a disclosure icon over the player using HTML and CSS. The icon remains visible on the Drupal page but does not become part of the externally hosted video. Publishers must therefore consider the disclosure settings available on the original hosting service when distributing that content elsewhere.

Generating labelled audio and video copies requires ffmpeg, ffprobe, suitable file storage and a processing queue. The renderer preserves original uploads and processes copies asynchronously, with only one rendition generated at a time. A default maximum duration of 3,600 seconds applies to eligible files. Failed processing jobs are not automatically retried, although administrators can requeue them through Drush. Sites unable to run FFmpeg locally can also prepare renditions on another machine and attach the completed files to Drupal.

The regulatory context is Article 50 of the EU AI Act, whose relevant transparency obligations became applicable on 2 August 2026. Paragraph 4 requires deployers to disclose qualifying deepfakes, including certain artificially generated or manipulated audio and video content. The European Commission's voluntary Code of Practice, published on 10 June 2026, provides practical recommendations for disclosure, including spoken notices for audio. These recommendations are not identical to the statutory obligations.

The distinction matters for synthetic speech that may be mistaken for an authentic person or recording. Cloned voices used in a news podcast can fall within the relevant deepfake disclosure requirements, while clearly fictional characters and routine audio processing may not. A machine-readable watermark placed by an upstream voice-generation provider also does not, by itself, satisfy a publisher's separate obligation to inform audiences when disclosure is required. AI Disclosure supplies labelling mechanisms, but legal classification remains the publisher's responsibility.

The DropTimes previously covered the module's September alpha release, which introduced disclosure grades, editorial decisions, machine-readable page information and audit features. The October beta extends those capabilities to images, audio and video, including the generation of labelled media copies. The module remains in beta without a supported stable release, and its release documentation outlines configuration and upgrade requirements for sites evaluating the functionality.