In theory, you could make the letters overlap so much that it is difficult to distinguish them in a single frame. This would give humans a distinct advantage over robots. However as it stands, a single frame is barely obscured at all, and I'm sure it could be easily broken.

I agree. Their assumption must be that OCR fails when a letter has covered regions, and to make it more readable to humans they provide the missing information in other time frames. But the letters have very nice borders and shadows, they are not deformed and have rigid motions which is trivial to track. Even in a single frame they provide more information than a traditional captcha. But the concept is promising and will definitely work with few modifications (e.g. no stroke color, time dependent deformations, not showing all letters in all frames etc.)

What makes you think the individual frames look like the letters? They certainly don't have to. At the price of a little flickering you could, say, show the top half of the letters in one frame, the bottom half the next, the right half the next and so on. Or randomly leave out one letter in every frame. Or even throw in different replacement letters for one frame each.

Or even better: Iterate over the frames until you find one which is very easily solved.

It can look at all frames and average predictions; this should boost the accuracy as errors will likely cancel out.

Accuracy is not always the only factor in play. The computational power required to process 20 frames is 20x the normal image. And the overhead of actually comparing the frames 20^2 if each frame is compared to every other frame. Since this is done automatically by bots, speed most certainly does affect the final performance of a bot.

Comparison of frames? I was only thinking about averaging predictions for each frame which is just 20x slower. Still I think that the process is way more constrained by the time to fetch new CAPTCHAs than by OCR, so this doesn't really matter.

If this becomes a trend, then surely the first frame won't always be the most obvious frame to choose.

If you were coding this I'm sure you'd jumble the start point anyway.

It shouldn't be necessary to have separated letters in the first frame.

Sure, but then it gains none of the advantage given to the human.

Why not? If it can grab all 20 frames, it has 20 times more data than it would normally have to solve the captcha.

Even if the captcha recognition's success rate is only 5%, its overall recognition rate is going to go through the roof.

You also get to do something you can't do with static captchas, compare the results. If 70% of the frames recognized the first letter as A, it's likely to be A, and so on.

As the letters move, they briefly do the really hard task of segmentation for you.

If there is any incentive to crack this captcha, it will be cracked very soon if not already, it looks like a very easy one to write.