A global wire service processes over 3,000 media assets per day — photographs, video clips, audio recordings, and written reports from correspondents and agencies around the world. In 2025, the verification desk began requiring AI detection screening on all incoming multimedia assets before publication. Here is how that workflow operates.
The verification workflow
Incoming media assets enter a staging queue. Each asset is automatically routed to the appropriate TruthScan endpoint based on file type. Results are returned in under 3 seconds for images and text, under 8 seconds for audio, and under 15 seconds for video. Assets with an AI probability below 20% are automatically cleared for editorial review. Assets above 70% are automatically flagged and routed to a senior verification editor. Assets between 20% and 70% are placed in a manual review queue.
The API integration is straightforward: TruthScan's REST API accepts file uploads and returns a standardized JSON verdict. The verification desk built a thin integration layer in Python that connects the staging queue to the TruthScan API and routes results to the appropriate editorial workflow.
What the verification desk found
In the first six months of operation, the verification desk screened approximately 550,000 assets. Of these, 2.3% were flagged for potential AI generation. Of the flagged assets, senior editors confirmed AI generation in approximately 60% of cases — a false positive rate consistent with TruthScan's published accuracy figures. The confirmed AI-generated assets included fabricated photographs of public figures, synthetic audio recordings presented as leaked conversations, and AI-written agency reports that had been submitted under human bylines.
Key integration lessons
Several lessons emerged from the implementation. First, threshold calibration matters: the 70% flagging threshold was initially set at 50%, which produced too many manual review cases for the team to handle. Raising the threshold to 70% reduced manual reviews by 65% while catching all the confirmed AI cases. Second, the per-sentence text analysis proved more valuable than the document-level verdict for wire reports, where individual paragraphs are frequently added or edited after initial filing. Third, the TruthScan certificate report format was directly usable as documentation for editorial decisions, reducing the time required to justify rejection of a submission to a contributing agency.