Recaptcha & Automated Driving
Recently during late nights, I have been participating in United Airline's "Billion mile giveaway." The idea is to spend money on your credit card to earn chances to win, but they also have a no-purchase-required way of entering for the sweepstakes online.
I want to win free flights and I hoped to automate the entire process, but the problem is that the Google recaptcha seems near-impossible to bypass. You have to choose from 9 initial images, but sometimes a new image pops up. They come in over the network as a stream of text and not as images, and the image quality is low and from varying angles. And there are different types of questions e.g. "Select the pictures with pancakes."
I've automated the opening up of the browser, the form-filling, and clicking the checkbox stating that I am not a robot, but the captchas, submit button, and closing of the browser are done manually. Because of recaptcha, I am reduced to only being able to do 8 submissions a minute.
Around this time last month, there were three frequent recaptcha questions:


I would estimate those appeared at a rate of 40%-30%-30%, respectively. The other 10% were pictures of food, mountains, rivers, lakes, or trees. Almost all pictures appeared to have been taken from a "Street View" car.
I took a two-week break, and I noticed two very big differences.
- About 90% of the questions were "Select all squares with street signs."
- The second part of that question was "If there are none, click Skip."
There's no doubt in my mind that this is related to automated driving. I'm not quite sure if this is what's happening, but I think Google is using humans to verify their test set.

Some other types of pictures I noticed that were interesting:
- Billboards
- Banners
- Housing "For Sale" signs
- Various different different angles, including behind the sign
- Signs of different countries
- blur (signs, license plates, people)
- Only part of a sign was in the photo
- A coconut tree that at first glance looked like a street sign