🔬 This guide covers the multi-layer verification workflow used by professional journalists and OSINT investigators. No single tool — including Scascan — provides a definitive verdict. The goal is building a weight of evidence.
In 2026, a viral image or video can spread to millions of people within minutes of being posted. Generative AI has made it trivially easy to create synthetic media that appears credible at a glance. Yet the tools that claim to "detect AI" are increasingly unreliable — modern models are explicitly trained to evade them. This guide explains the professional verification workflow used by journalists, fact-checkers, and open-source intelligence researchers.
Why AI Detection Alone Is Not Enough in 2026
The fundamental problem with AI detection tools in 2026 is that the adversarial arms race has shifted decisively in favour of the generators. Models like Flux, Midjourney v7, Stable Diffusion 3.5, and Sora 2 are trained with explicit anti-detection objectives. Major AI labs report that their own detection APIs fall to near-noise accuracy on fresh outputs from current models.
Models evade detectors by design
Modern diffusion models are trained to minimise statistical artifacts that detection tools rely on — making detection increasingly unreliable.
Recompression destroys signals
Images shared on WhatsApp, X, or Telegram are recompressed multiple times, stripping most forensic signals before investigators see them.
False positives undermine trust
Legitimate photos edited in Photoshop, HDR-fused exposures, or artistic photography regularly score high on AI detectors.
The gap is widening
Detection tools improve incrementally. Generation models improve exponentially with each funding round. The technical gap is not closing.
The 5-Step Professional Verification Workflow
Professional fact-checkers at organisations like Bellingcat, the BBC Verification Unit, AFP Fact Check, and Reuters Graphics use a structured multi-step process. No single step is conclusive on its own. The goal is building a weight of evidence.
Technical Signal Analysis (Starting Point, Not Verdict)
Use a browser-based media signal analyser to inspect: EXIF metadata fingerprints (what software created this?), compression history (how many times re-encoded?), frequency domain patterns (GAN signatures?), texture fractal continuity (over-smoothed like diffusion?), and Error Level maps (where are the editing boundaries?). The Scascan Media Signal Analyser runs all seven layers entirely in your browser — no upload needed.
Run Signal Analysis →Reverse Image Search (Finding Earlier Appearances)
Search for earlier appearances of the same image or near-duplicates. Use Google Lens (right-click → Search image), TinEye (oldest-first dating), Yandex Images (strongest for faces and artworks), and Bing Visual Search. If the image has appeared before the event it allegedly depicts, it is misattributed. If it appears in a completely different context, it is repurposed.
Geolocation Verification (Proving the Location)
Cross-reference background elements with satellite imagery and local knowledge: building architecture, terrain, street furniture, vegetation, signage, vehicle types. Use Google Earth for 3D terrain matching, Sentinel Hub for satellite imagery, and Mapillary for street-level comparison. Shadow analysis tools like SunCalc verify whether shadow lengths and angles match the claimed time, date, and location.
Source Triangulation (Following the Trail)
Find the earliest known post. Search X (Twitter) with image search filters for the date range. Search Telegram channels. Check Facebook and TikTok. The earliest post often reveals the true origin — and whether the claimed attribution matches. Look at the account that first posted: when was it created? What else have they posted? Is there a pattern of posting unverified viral content?
Cross-Reference with Wire Agencies (The Authority Check)
AP, Reuters, AFP, Getty, and EPA all timestamp and geolocate their images. If a major news event occurred as claimed, there should be independently verified wire photos from the same location and time. If the viral image has visual qualities or content completely absent from wire coverage, that is significant. Eye-witness video from bystander phones, dashcams, or CCTV should also appear.
How to Read a Signal Analysis Report
If you use the Scascan Media Signal Analyser, here is how to interpret the results responsibly:
| Signal Layer | What It Tells You | Limitation |
|---|---|---|
| Metadata Fingerprint | Software that created or saved the image; GPS if present | Stripped by all social platforms on upload |
| Patch Variance | Whether texture is unnaturally smooth over large regions | Can vary with image subject matter; false positives on soft subjects |
| Frequency Signature | GAN upsampling periodic artifacts in frequency domain | Limited signal on modern diffusion models |
| Texture Intelligence (ML) | Statistical features of pixel distribution across channels | Most reliable cross-format signal; still not definitive |
| Compression History | File-size ratios, DCT artifacts, AI-common dimensions | Recompression degrades signal; not conclusive alone |
| Error Level Map (ELA) | Where editing boundaries and uniform generation exist | More robust post-compression; strongest on JPEG |
| Temporal Coherence (video) | Whether frames are statistically too consistent for real video | Low σ is a signal, not a verdict; check with OSINT |
Key Visual Signals to Check Manually
Trained human eyes remain a crucial verification layer. Look for these signals before running technical analysis:
Hands and anatomy
AI generators frequently produce hands with incorrect finger counts, merged digits, or unnatural joint geometry. In crowd scenes, faces may have subtle bilateral symmetry or ears that merge into hair.
Text and insignia
Military content, signage, and document text are common AI failure points. Unit insignia may be invented or garbled. Vehicle plate numbers, flag details, and labelling on equipment rarely survive generation intact.
Shadow and light consistency
Check whether shadow directions from multiple objects in the scene agree. AI images sometimes produce scenes with two implied light sources — physically impossible without studio setup.
Background over-detail
Diffusion models frequently generate backgrounds with an artificially high level of convincing detail while subjects in the foreground show over-smoothed skin or blurred edges at object boundaries.
Video stillness
AI video generators often produce footage where backgrounds are suspiciously static — no wind movement in vegetation, no camera shake, no dust variation. Authentic handheld footage always has organic motion.
Quick Reference: Verification Toolkit
Signal Analysis
- ›Scascan Media Signal Analyser — 7-layer browser-based
- ›EXIF.tools — metadata inspection
- ›FotoForensics — ELA online
Reverse Image Search
- ›Google Lens — broadest index
- ›TinEye — oldest-first dating
- ›Yandex Images — face & artwork matching
Geolocation
- ›Google Earth — 3D terrain matching
- ›Sentinel Hub — satellite imagery
- ›SunCalc — shadow & sun position
Source Tracing
- ›X Advanced Search — date-filtered
- ›Telegram search — early posts
- ›InVID / WeVerify browser extension
Run a Signal Analysis Now
Use the free Scascan Media Signal Analyser as your first forensic step. No upload. No account. Files never leave your device.
Related Resources
📝 Editorial Note
This guide is for educational purposes. Verification methodology described here reflects standard practice in professional journalism and open-source intelligence. No specific geopolitical event or actor is referenced. Signal analysis results are technical probability estimates — not forensic evidence. Always cross-reference with multiple independent sources before drawing conclusions or publishing.