Okay, so there's some pattern that shows up against a completely untextured view of the world. How would they recover such a faint watermark from an ordinary view of the world, complete with complex textures in the background? For that sort of thing, you need a copy of the image without the watermark so that you can take the difference between the two, and that doesn't seem to be the case here. And if you wanted to covertly record someone's data, why go to this effort when you could just send it to your server without telling them?

If it ever makes into reality (I doubt it will), you will be given a simple choice: pay whatever some algorithm believes you are prepared to cough up even if it's more than everybody else, has overestimated your income, net worth, etc, etc; or go to a high street store where you pay the price on the sticker, no questions asked. This scheme could be just the saviour that the high street has been waiting for. And to those who say it's only a form of bargaining and the workings of supply/demand market forces, imagine your employer giving you a pay cut because they heard you just inherited some money from a recently deceased relative (after all, you don't need as much money now, right?) There's already a place where supply/demand forces work in the way they were intended - it's called eBay.

As already mentioned, getting the user to move their head and checking for the correct 3D effect solves this problem (known as 'liveness detection'). In our case, we actually use combined face and voice recognition where the user needs to answer a question posed by the phone - we can therefore check for the lips moving to make sure it's not a photo. Showing a video could potentially spoof the system (any system is hackable - it's just a question of whether it's worth the effort) but there are almost certainly differences that we can detect between light reflected from a real object and the light generated by a video screen. Same goes for sound.

I actually agree - the recognition will need to prove its worth before I put my bank account in its hands. This is, however, a long-term goal and the reality is we'll get there mostly with baby steps. In the short-term, it'd be a piece of cake to capture an image of your face as you're entering your password as an additional (rather than substitute) level of security. Other obvious ethical issues include where and how your biometric 'fingerprint' is stored - on the phone (don't lose it) or a central database (who's in charge of that?)

Spot on - it models the whole face in one go rather than each feature independently. So as long as most of the face is visible it will guess the positions of the bits it can't see based on the bits it can. The more face gets hidden, however, the worse this guess gets. For most practical purposes you can assume that there's a clear line of sight between the camera and the face.