Real-World Practice:
Judging AI Credibility from Classrooms to Workplaces and Scams

Taiwan FactCheck Center researchers used: ① Comparing voice features (prosody patterns, voice prints) with large amounts of known authentic Lai Ching-te speech recordings; ② Source tracing of the claimed "speaking occasion" — found no journalist attendance or live broadcast records; ③ AI voice forensic analysis showed the voice's prosody during emotional passages had significantly below-normal variation, showing typical AI voice synthesis characteristics; ④ The candidate's campaign office denied making this statement.

Determined to be AI voice clone fabrication, constituting election interference. The case was referred for investigation; with deepfake audio released just weeks before the election, there was extremely limited time for verification and clarification during the electoral period.

• Source verification: Any "candidate statement" without corresponding news coverage, journalists present at livestream, or official release is highly likely to be fabricated
• Consistency check: Is this statement consistent with the candidate's known positions? Drastic position reversals without mainstream media coverage should be highly suspected

① Reverse image search: Screenshot key frames for reverse search; found the same footage appearing as early as 2017 in a medical education video about "vasovagal syncope" (brief fainting after injection due to pain/anxiety — not life-threatening); ② Medical context analysis: Personnel's response after the woman fainted was consistent with standard vasovagal syncope procedures, not death events; ③ Clinic equipment and medicine bottles in the video were confirmed through geolocation to be from a different country, inconsistent with the claimed location.

This case perfectly demonstrates the typical "false context" technique: a genuine medical education video (vasovagal syncope) paired with a false "vaccine death" description, creating completely different "meaning." Reverse image search is the fastest tool to expose this technique.

• Real professionals have official records: Real lawyers are registered with bar associations; real doctors have license numbers that can be checked; real companies have registration numbers. Verify credentials first before deciding to engage.
• GAN face detection: Do reverse image search on suspicious "professional" photos. GAN-generated faces usually have no historical appearance records; tested in Which Face Is Real or Hive Moderation, they typically get high suspicion scores.
• Video call verification: Require "live video" (not just voice call) and ask them to make unnatural movements (quick head turn, cover face with hand then remove) — deepfakes and AI-generated personas break down most easily in these situations.

Bellingcat investigators confirmed video authenticity through: ① Extracting buildings from the video and finding perfectly matching landmarks in Google Street View (even to specific wall graffiti); ② Confirming military identification numbers on vehicles in the video matched Russian 41st Combined Arms Army armored unit records; ③ Using SunCalc to confirm shadow directions in the video perfectly matched the claimed date and location; ④ AI forensic analysis of the video found no deepfake synthesis signs; ELA analysis showed normal JPEG compression patterns.

This case perfectly demonstrates the practical application of the "Liar's Dividend": Russian officials knew the public feared deepfakes and tried to use "deepfake accusations" to muddy the waters. The correct response isn't to "believe everything" or "doubt everything," but to do systematic authentication work. Geolocation matching, equipment identification, shadow analysis — these methods can both expose fake content and protect authentic visual evidence.