Catfishing in the AI Era: Detecting Synthetics

The Death of the 'Stolen Photo'

Until recently, identifying a catfish was relatively simple: Download their profile picture, run it through Google Reverse Image Search, and find the real person whose photos they stole (often an influencer or military personnel).

In 2025, that methodology is dead. Scammers now use Generative Adversarial Networks (GANs) and models like Midjourney to summon an infinite array of non-existent humans. Because the image is entirely new, a reverse image search will return zero results.

1. Detecting the Synthetic Face

While Generative AI is powerful, it still leaves digital fingerprints. Analysts use deep-learning models to hunt for these artifacts:

• Pupil Asymmetry: AI often renders pupils that are slightly irregular in shape or pointing in micro-divergent directions.
• Background Melting: The AI focuses its rendering power on the face. If you look at the background—the texture of a brick wall, the lines of a bookshelf, the text on an abstract sign—it often breaks down into nonsensical 'melting' shapes.
• Ear Cartilage Topology: Human ears are incredibly complex. Synthetics frequently generate ears that lack depth or biological logic.

2. The Deepfake Video Call

The ultimate proof of life used to be a live video call. Today, real-time deepfake filters can map an AI face onto a scammer's body during a Zoom or WhatsApp call.

• The Profile Test: An investigator will ask the target to turn their head 90 degrees to the side. Most real-time consumer deepfakes struggle to render side profiles seamlessly, resulting in visual "tearing" or sudden blurriness.
• The Hand-over-Face Test: Asking the target to pass their hand quickly in front of their nose breaks the facial mapping mesh, momentarily revealing the true user underneath.

3. Syntactic and Temporal Analysis

If the visual data is inconclusive, we deploy behavioral OSINT.

• LLM Scripting: Scammers are using ChatGPT to generate highly romantic, persuasive dialogue. We analyze text frequency and syntactic structure. If a target's writing style is wildly inconsistent, or perfectly mimics the output of a specific LLM model, it is flagged.
• Timezone Inconsistencies: Use our IP Address Geolocation protocols to determine if the user claiming to be in London is exclusively logging in and messaging during standard business hours in Lagos or Manila.