Skip to content

T2 – First-line triage for a suspected AI-generated video

TL;DR

Use this tree for short videos, livestream clips, political endorsements, alleged protest or violence footage, and screen recordings carrying cloned audio. The tree separates URL preservation, keyframe-driven reverse search, and source behaviour from synthetic-video detection, which sits at the end as a weak signal.

When to use this tree

Short-form video is where most of the synthetic-content workload lands in regional newsrooms, and it is also where the forensic surface is weakest. TikTok, Reels, and Shorts re-encode aggressively: real compression artefacts mimic synthetic ones, and cloned audio survives the same compression that erases visual detector signal. The tree leads with URL and keyframe preservation because evidence is fragile. TikTok takedowns, Facebook share-edits, and LINE re-encoding all destroy what a forensic step would otherwise read. When the voice carries the central claim and the picture does not, the workflow leaves T2 for T3 and returns only after audio is resolved.

The tree

%%{init: {"flowchart": {"nodeSpacing": 25, "rankSpacing": 40, "curve": "basis"}} }%% flowchart TD A["T2.1 / 2.3 Preservation + claim isolation
URL · platform · visual / voice / caption split"]:::decision B["T2.4 / 2.7 Trace + provenance
keyframes · reverse search · metadata · visible anomalies"]:::action C["T2.8 / 2.10 Voice handoff + transcript + detector"]:::decision D["T2.11 / 2.13 Plausibility + behaviour + cross-platform"]:::decision E["T2.14 / 2.17 Outcome classes
synthetic-likely · false-context · inconclusive · no evidence"]:::terminal A --> B B --> C C --> D D --> E click A "#t2-1" "T2.1 in Node detail" click B "#t2-4" "T2.4 in Node detail" click C "#t2-8" "T2.8 in Node detail" click D "#t2-11" "T2.11 in Node detail" click E "#t2-14" "T2.14 in Node detail" classDef decision fill:#fff4cc,stroke-width:0,color:#000; classDef action fill:#e6f0ff,stroke-width:0,color:#000; classDef terminal fill:#c8f0c8,stroke-width:0,color:#000;

The diagram is a macro view of the main video-triage chain. Click any block to jump to its Node detail row.

Side exits, kept out of the diagram for clarity:

How to read this tree

Video carries two evidentiary surfaces – picture and voice – and either can be authentic while the other is fabricated. That is why T2 runs longer than T1 or T3. Work T2.1 through T2.6 first; if the voice ends up carrying the claim, jump to T3 at T2.8 and come back only when the audio-side answer is in. Keep three ledgers open while the case moves: claim authenticity, voice authenticity, and caption truth. Each gets resolved on its own evidence. A real video with a fake voice and a misleading caption is the typical Indonesia, Malaysia, and Philippines short-form pattern; collapsing the three early is what produces both false positives and the liar's-dividend abuse on the other side.

The four classes of first-line outcome are:

  • old or authentic video used in false context (T2.15 → response or T5);
  • likely synthetic or manipulated video, requiring professional verification (T2.14 → T5);
  • caption or transcript mismatch, ready for fact-check correction (→ T7);
  • no first-line manipulation evidence (T2.17 → ordinary verification continues).

Plus two hand-offs: T2.8 sends voice-central cases into T3, and T2.13 sends multi-account synchronised reuse into the coordinated-operation analysis branch.

Node detail

Node Question or action Time Tools
T2.1 Save URL, uploader name, platform, upload time, caption, and pinned comments. Request original if safe. 1 to 3 min
T2.2 Platform preservation: TikTok, Facebook, WhatsApp / LINE, Telegram, YouTube. 2 to 5 min Auto Archiver, Citizen Evidence Lab YDV for YouTube.
T2.3 Write the claim as one sentence. Identify whether the visual scene or the voice is doing the evidentiary work. 1 to 2 min
T2.4 Extract five to ten representative keyframes. 3 to 6 min InVID-WeVerify, Citizen Evidence Lab YDV
T2.5 Reverse-search full keyframes and cropped landmarks, faces, signs. Search several frames, not only the first. 5 to 10 min InVID-WeVerify
T2.6 Read provenance / metadata only if the original file exists. Treat absent metadata as normal on TikTok / WhatsApp / LINE. 3 to 6 min InVID-WeVerify, MetaDataKit, ExifTool, Content Credentials Verify
T2.7 Watch at normal then half speed. Lip-sync, face edges, lighting, reflections, hand interactions, sudden cuts. 2 to 5 min
T2.8 Voice central? Separate the audio question from the visual; route into T3 while continuing visual checks. 1 to 3 min
T2.9 Transcript matches caption? Use ASR plus local-language reviewer. 5 to 15 min OpenAI Whisper, Google Cloud Translation
T2.10 Two first-line video detectors at most. Record file quality, score, upload safety. 5 to 15 min InVID-WeVerify deepfake tab, Deepware Scanner, Hive AI, TrueMedia / Georgetown
T2.11 Event plausible? Schedules, prior posts, weather, landmarks, uniforms, signage. 5 to 15 min GeoSpy for leads, manual OSINT
T2.12 Account age, post history, sudden topic shift, repost watermarks, off-platform funnel. 5 to 10 min
T2.13 Cross-platform identical reuse. Search exact caption, key quote, distinctive frame across TikTok, Facebook, Telegram, local fact-check or tipline systems. 10 to 20 min Information Tracer, Media Cloud
T2.14 Two independent classes of synthetic / manipulated evidence. Build the packet for T5; do not publish "confirmed deepfake" without provenance, expert, or source confirmation. 5 to 10 min
T2.15 Authentic or older video used in false context. Save earlier source, current misleading caption, date and place mismatch. 5 to 10 min
T2.16 Inconclusive but high impact. Bundle file or URL, keyframes, transcript, platform notes, tool outputs, unanswered questions. Hand off to T5. 5 to 10 min
T2.17 No first-line manipulation evidence. Continue ordinary claim verification. 1 to 2 min

Regional and platform routing

The country branches mirror T1.15 and T7.20: see T7 – tipline routing for the full per-country deployment table.

Platform branch:

  • -tiktok / -fb: capture URL before takedown; keep raw file alongside the URL because Facebook share-edits propagate.
  • -wa / -line: tipline submissions only; do not expose sender identity.
  • -telegram: public channel data plus message ID for public posts; private channels under S5.
  • -youtube: prefer Citizen Evidence Lab YDV for upload time and thumbnail provenance.

Apply S2 for state-linked or legally sensitive video in Laos, Sri Lanka, Thailand, Philippines red-tagging, Indonesia EIT Law, Malaysia CMA contexts before any contact with subjects, uploaders, or authorities.

Cross-references

Sources

  • WITNESS Media Lab and Reuters Institute. Thinking About Deepfakes: A Verification Framework for Journalists. WITNESS, April 2024. witness.org. (Video triage methodology; keyframe and motion-physics checks at T2.3–T2.5.)
  • Lyu, S. et al. Deepfake-Eval-2024: A Real-World Benchmark for Deepfake Detection. 2025. arxiv.org/abs/2503.02857. (Best commercial video detector 0.78 accuracy / 0.79 AUC; category ceiling underlying T2.10 detector-conflict handling.)
  • InVID-WeVerify Consortium. InVID-WeVerify Plugin Documentation. WeVerify Project, 2023. weverify.eu. (Keyframe extraction, reverse-image, and deepfake-tab workflows at T2.3–T2.9.)
  • Architectural Anchors — Anchors 1, 2, and 3 as operationalised in this tree.