If we didn't fly it, we'll tell you.
Air Pixels comes from film and television. We fly aerials for commercials, music videos and creative projects, and we capture data and progress imagery on construction sites. Generative video models can now produce aerial shots that pass for real. This statement covers where we think that's heading, when we use AI, how we label it, and how to check footage yourself.
- FLOWN
Recorded in camera on a logged flight.
- ASSISTED
Flown, then changed or added to by AI or VFX.
- GENERATED
Made by a model. No flight behind it.
Outlook
Where aerial content is heading.
Our read on the next few years, from the set and from the job site.
The generic aerial gets cheap.
Video models learned from a great deal of stock aerial footage, and a slow push over a coastline or skyline is one of the things they do best. Shots like that will cost less and carry less weight with audiences.
Specific footage gains value.
A real building on a known date, a real event, a location the client can stand in. Footage tied to a place, a time and a licensed pilot becomes worth more as generic shots get cheaper.
Proof moves to the moment of capture.
Labels added after the fact are easy to strip. Proof will come from cameras that sign files as they record, and from the flight logs drones already keep.
Disclosure is becoming a requirement.
YouTube, TikTok and Meta ask creators to label realistic synthetic video. The EU AI Act requires deepfakes to be disclosed, with lighter disclosure for clearly artistic work. We would rather work to a clear standard now than retrofit one later.
Our commitments
Five things we hold ourselves to.
We never present generated imagery as flown footage.
Any shot without a flight behind it carries the Generated label in the deliverable, the delivery notes and the file's Content Credentials.
We disclose AI changes shot by shot.
Removing a crane, replacing a sky or upscaling to 8K changes what the camera recorded. Those shots are labeled Assisted, and the change is listed.
Measurement data stays captured data.
Orthomosaics, point clouds, progress reports and inspection imagery are never generated or filled in by a model. AI can analyze that data. It never adds to it.
Every flown shot has a flight record.
We keep the aircraft, pilot certificate, airspace authorization and flight log behind each shot, and show them to clients on request.
Client footage doesn't train models.
We don't use client footage or site data to train or fine-tune AI models unless the client opts in, in writing.
Labels
Every deliverable carries one of three labels.
The label travels with the file in its Content Credentials and in the delivery notes. Where the client agrees, it also appears in the credits or on screen.
Recorded in camera on a logged flight.
- Color grading and LUTs
- Exposure, noise reduction and lens correction
- Stabilization, cropping and speed changes
Delivery note
- Shot
- AP-2611-014 · 00:12
- Flight
- Oct 14, 2026 · 16:42 UTC
- Airspace
- LAANC · 200 ft AGL
- Pilot
- Part 107 · on file
- Edits
- Grade, stabilize
Flown, then changed or added to by AI or VFX.
- Object, crowd and rig removal
- Sky replacement and relighting
- AI upscaling and frame interpolation
- Frame and set extension
Delivery note
- Shot
- AP-2611-022 · 00:08
- Flight
- Oct 14, 2026 · 17:05 UTC
- Changes
- Sky replaced · crane removed
- Frames
- 112–260
- Tools
- Named in notes
Made by a model. No flight behind it.
- Previs and animatics
- Pitch and mood films
- Concept frames
- Impossible shots for music videos and creative work
Delivery note
- Shot
- AP-2611-031 · 00:06
- Flight
- None
- Source
- trainedAlgorithmicMedia
- Model
- Named in notes
- Use
- Previs
When a shot sits between two labels, it gets the stronger one.
Where AI fits
When AI-assisted content makes sense.
One question decides it: could someone be misled about a real place, event or condition? The more a viewer, client or court would rely on the footage being real, the less AI we allow.
Commercials
Cleanup, sky continuity across shoot days, previs and animatics, and stylized shots the client approves as generated.
Generated shots of a real location, product or result that the ad presents as real.
Music videos & creative projects
Impossible camera moves, surreal worlds, and shots that start flown and end generated. Generative tools add the most here.
Claiming a generated shot was a real stunt or location in press or behind-the-scenes material.
Documentary & news
Stabilization and upscaling, labeled. Maps and graphics that read clearly as graphics.
Any generated or reconstructed event shown as footage.
Real estate
Exposure blending and removing our own equipment. Sky replacement only when the listing discloses it.
Changing the property, the view, the neighbors or the condition of anything.
Construction & survey
Progress tracking, stockpile volumes and change detection, run on captured data and labeled as model output.
Filling gaps, sharpening or generating any pixel in a measurement deliverable.
Inspection & insurance
Flagging cracks, corrosion and roof damage for an inspector to confirm.
Enhancing or altering imagery that may become evidence.
How we build software
Flight OS uses AI to help pilots. It never invents a record.
Air Pixels Flight OS handles airspace, weather, shot lists, compliance and flight logs for commercial drone pilots. We use AI to build it and inside it, under four rules.
Records are captured, never generated.
Flight logs, LAANC authorizations and checklists come from the aircraft, FAA systems and the pilot. A model can summarize a record. It can't write one.
AI drafts. The pilot in command signs.
Briefings, risk notes and shot lists written with AI stay marked as drafts until a certificated pilot reviews and signs them. Under 14 CFR 107.19 the remote pilot in command is responsible for the flight, and the software keeps it that way.
Provenance starts at the sortie.
Flight OS links each media file to the flight that produced it: aircraft, authorization, track and timestamps. That link becomes the flight assertion in the file's Content Credentials.
People review every change we ship.
Our engineers use AI coding assistants. Every change is read by a person, tested, and traceable to whoever approved it.
Delivery drones get the same treatment in Proof of Flight, our framework for recording every AI decision an aircraft makes.
Transparency & standards
Helping set the standard for aerial provenance.
Content Credentials, built on the open C2PA standard, give a file a signed history of what made it and what changed. Aerial footage can carry one more piece of evidence that most media can't: the flight that captured it.
Content Credentials on every deliverable
We sign files at export with a manifest that records the label, the edits and the tools used. Credentials signed by the camera are kept wherever the hardware supports them.
An open flight assertion
We're drafting a C2PA assertion that binds a clip to its flight: aircraft serial, takeoff time, a hash of the flight track and the airspace authorization. We'll publish the schema for operators and tool makers to use and critique.
Labels anyone can use
Flown, Assisted and Generated are free to adopt. We version the definitions on this page so every change is visible.
Shot-level disclosure
Delivery notes list every Assisted and Generated shot with timecodes, so legal, clearance and ad standards teams don't have to ask.
Content Credentials
AP-2611-014.movAir Pixels Flight OSdigitalCapturec2pa.color_adjustments
c2pa.croppedFlowncom.airpixels.flight · draft
1581F…7Q2C2026-10-14T16:42:07Z9f2c…41aeLAANC · 200 ft AGLPart 107 · on fileAir Pixels, Inc.Illustrative manifest. Field names in the flight assertion are a working draft.
Ground Truth
How we spot generated aerial footage.
Ground Truth is the review we run on footage we're asked to verify, and on any stock or client-supplied aerial before it goes into our edits. It works through three layers, strongest first. No single tell proves a shot was generated, and a clean result doesn't prove it was flown.
Provenance
Is there a record?Content Credentials, watermarks, the original file and the flight log.
Physics
Could a drone have shot it?Sun, parallax, scale, motion, focus and airspace.
Pixels
Does it hold up frame by frame?Texture, object permanence, text, infrastructure, water and sensor noise.
01 · Provenance
Is there a record?Content Credentials
Open the file in a C2PA verifier such as contentcredentials.org/verify. Check who signed it and what it declares as the source.
AI watermark
Run it through model makers' detectors, such as Google's SynthID Detector. Many tools add no watermark, so a miss proves little.
Original file
Drone files carry a camera model, standard frame rates and often a telemetry sidecar. Clips of exactly 5, 8 or 10 seconds with no camera data match common model defaults.
Flight log
Ask whoever supplied the clip for the flight log. A pilot who flew it can export one in minutes.
02 · Physics
Could a drone have shot it?Sun & shadows
Shadows should all fall one way, at a length that fits the sun's height for the claimed place and time.
Parallax
Near objects should slide past far ones while buildings stay rigid. Watch rooflines and corners as the camera moves.
Scale
Cars, lane widths and people should stay the right size for the altitude through the whole move.
Camera motion
Drones accelerate, brake and drift in wind. A frictionless glide, or a path no aircraft would fit through, is a tell.
Depth of field
Small drone sensors focused far away keep nearly everything sharp. Shallow focus from altitude usually means a tilt-shift effect or a model.
Airspace
Check the location in Can I Fly Here. Over a stadium during a game, beside a runway or inside a US national park, a real flight needs approvals that are rarely granted.
03 · Pixels
Does it hold up frame by frame?Texture
Step through frame by frame. Leaves, gravel and windows that shimmer or redraw between frames are a model artifact.
Object permanence
Things that leave frame should come back unchanged. Cars shouldn't merge, vanish or pass through each other.
Text & markings
Signs, road markings, roof numbers and logos should be legible and stay the same from frame to frame.
Infrastructure
Roads should connect, power lines should run pole to pole with an even sag, and bridges need supports.
Water
Wakes trail boats, reflections match what's above them, and waves move with the wind.
Sensor signature
Real cameras show noise that tracks exposure, rolling-shutter skew on fast moves and compression blocking in foliage. Waxy, noiseless surfaces are a tell.
The strongest test
Track match
Structure-from-motion software can solve where the camera was for every frame of a clip. For a flown shot, that solved path lands on the flight log within a few meters. A generated clip has no flight log to land on, and forging one that agrees with the footage, an aircraft serial number and an airspace authorization is far harder than generating the clip.
Flight logSolved from flown footageSolved from generated footageIllustrative
Ground Truth review
Unsure about a shot?
Send us the clip and anything you know about where it came from. We'll run Ground Truth and tell you what we found and how confident we are. Operators and production companies who want to adopt the labels can write to the same address.