The Flood Changed the Job
There is a myth that AI made open-source intelligence easier. The opposite is true. Generative models flooded every channel an analyst draws from — synthetic accounts, fabricated images, plausible-sounding documents, entire manufactured personas with consistent backstories. The collection problem got cheaper. The verification problem got vastly more expensive. The analysts who are still good at this in 2026 are the ones who treated AI as a contamination of their sources rather than a new source.
The First Casualty: "I Saw It Online"
Screenshots, photos, and documents now have no evidentiary weight on their own. None. A convincing image proves nothing. A coherent document proves nothing. A network of accounts agreeing with each other proves nothing — that is the cheapest thing in the world to manufacture now.
So you go back to fundamentals: a claim is only as strong as the independent, physical-world evidence you can correlate to it. The four-layer model that governs internet cartography governs verification too.
- Geographic layer. Does the claimed location survive geolocation? Sun angle, shadows, architecture, signage, vegetation, license-plate formats. The physical world is stubbornly hard to fake consistently.
- Physical/logical network layers. Where does the content actually originate? Hosting, ASN, registration timing. A "grassroots" campaign that all resolves to one prefix registered last Tuesday is a campaign with an owner.
- Cyber-persona layer. Account age, posting cadence, the metadata that survives, the cross-platform fingerprint. Personas are cheap; consistent histories across independent systems are expensive.
When a claim holds across all four, you have something. When it lives only inside the content itself, you have a guess.
Verification Workflow That Still Works
1. Capture, do not trust. Archive the artifact (timestamp, hash, source URL).
2. Strip and read metadata. EXIF where it survives; absence is itself a signal.
3. Reverse-search broadly. Multiple engines — fakes seed one and miss others.
4. Geolocate physically. Match against satellite/street imagery, terrain, sun.
5. Chronolocate. Weather records, shadow length, known events to fix time.
6. Correlate the network. Hosting, ASN, registration, cross-platform identity.
7. Seek the independent second source. One physical-world confirmation > 100 reposts.
Steps 4 and 5 are where AI fakes still fall apart most reliably. A model can render a convincing street; it struggles to make the shadows agree with the claimed latitude, date, and time and match the real building that exists at those coordinates. Physics is a harsh fact-checker.
Use AI Without Being Used by It
This is not a Luddite argument. AI is a superb force-multiplier inside a verification discipline:
- Translation and transcription at scale, so language is not a collection barrier.
- Triage of large dumps to surface what a human should examine first.
- Pattern surfacing across datasets too large to read.
What it must never be is the arbiter of truth. The model summarizes; the analyst verifies. The instant you let a model's confident paragraph stand in for physical-world correlation, you have outsourced your judgment to the exact technology your adversary is using to deceive you.
The Noise Is the Point
Understand the adversary's actual goal. Most influence operations in 2026 are not trying to make you believe a specific lie. They are trying to make you believe nothing is knowable — to drown every real signal in enough plausible fakes that you give up on verification entirely. Exhaustion is the objective. The counter is not cynicism; it is method. A disciplined analyst with a verification checklist and the patience to run it is the thing these campaigns cannot defeat at scale.
The Standard Holds
The tools changed. The standard did not: extraordinary claims require independent, physical-world corroboration, correlated across layers that are expensive to fake simultaneously. Hold that line and the flood is just weather. Drop it and you will drown in content that was generated specifically to drown you.