AI flight planning helped American Airlines reduce contrails across the Atlantic

American Airlines’ AI-assisted flight-planning trial recorded 11.6% less contrail length per flight kilometre across all 1,232 eligible treatment flights than in controls, adjusting for aircraft type. Google, Contrails.org and Flightkeys helped put forecasts into ordinary planning software so dispatchers could consider the clouds a route might produce.[2]

Google described the larger trial in March 2026, following earlier tests that required substantial manual coordination. The useful advance was making a potential environmental improvement available during routine flight planning.[1]

Why a small change in altitude can matter

Contrails form when aircraft pass through suitable cold, humid air. They can persist and spread into clouds. Those clouds can reflect incoming sunlight, but they also trap heat that would otherwise escape. At night, there is no sunlight to reflect, making the heat-trapping effect especially relevant.[3]

A flight can sometimes avoid the conditions that produce persistent contrails by using a different altitude. The aircraft, fuel and destination can remain the same; the intervention changes where it travels through the atmosphere. Its environmental value depends on which trails are avoided and any additional fuel the route requires.[3]

What the transatlantic trial measured

The January–May 2025 trial repeatedly randomised city pairs between treatment and control. Its 1,172 control flights provided the comparison. Satellite imagery supplied the outcome measurement.[2]

The whole-group result measures the benefit of offering the option, including flights that did not use it. Only 112 of 1,232 eligible flights—about 9%—flew avoidance plans as specified. The larger reduction in that selected subset cannot establish what every eligible flight would achieve. Operational uptake is part of the result, alongside forecast quality.[2]

The preprint remained under review in October 2026. Its public discussion distinguishes observed contrail reduction from its estimated climate benefit.[7]

AI helps before and after the flight

Google’s approach combines weather information, flight tracks and satellite imagery. A computer-vision model learns to recognise contrails from labelled images. Connecting those detections with aircraft movements creates evidence about the conditions under which a flight left a visible trail.[4]

Forecasting then turns that accumulated information into something a dispatcher can use before departure: where contrail formation is likely. Later satellite observations can check what happened. This links a prediction to an observable outcome, giving the team a way to improve both the forecast and the route-selection process.[4]

Google now offers forecast datasets through its Contrails API at no charge. Forecast grids and regions carry a Creative Commons attribution licence; its observation products have separate, more restrictive licences. That makes the forecast an input other planning systems can obtain, while keeping the terms for different datasets explicit.[5]

The next test covers an airspace

Operation Blue Skies will examine coordinated altitude changes across the Shanwick oceanic airspace. Cambridge describes a £5 million, 30-month programme, including £2.65 million from the UK Department for Transport. Operational testing is planned for the winters of 2026–27 and 2027–28, with roughly 20–40 test days each winter.[6]

NATS, the Met Office, Google, Contrails.org, Cambridge and Imperial College London are among the partners. Cambridge and Imperial will evaluate the outcomes and help verify climate impact. The programme moves the question from individual airline routes to coordination through established air-traffic-control channels.[6]

That progression is useful: a forecast can first help one flight choose a better path, then become part of how an airspace handles many flights. The coming trials will show how much of that opportunity survives everyday traffic and weather.[6]

Report an error or suggest a correction