Airline barcodes aren't QR codes. They are Aztec codes or PDF417 codes, and software for reading them is generally less robust and less smart because less investment has gone into clever ways to align the codes quickly. The codes themselves also have less good alignment markers, especially for dealing with lens distortion.

Scanning an on-screen QR code is usually done with barcode scanners, not with phones. Typically, barcode scanners are doing the decoding on-device, and passing out keyboard data (their drivers usually identify them as USB HID). Unfortunately, most scanners don't have loads of onboard processing power, though ability to deal with damaged linear barcodes has been quite good for a long time. Damaged 2D barcodes are harder to deal with, and while ML may be able to solve that, it would require a paradigm shift in how barcode scanners operate -- and essentially bifurcate the market, as some scanners would simply no longer be compatible with embedded systems, and would require specialised software to function (which isn't really ideal; most users of barcode scanners want them to be plug-and-play).

(My first post-university job was doing tech support for a company which sold point-of-sale equipment. I'd say about 50% of my job was helping people with barcode scanners.)

Phone screens are awfully flat though - which means glare from a light is typically a small area. QR codes are robust to small areas being missing. In fact, a typical QR code is readable with about 20% of the code missing.