This is about as effective as asking a random question, it stops bots because they don't know how to deal with it. It wouldn't have a chance against a targeted spam campaign/attack.
It's cute though.
I believe you're underestimating the concept. The current implementation is a simple game that asks a human about domain-specific knowledge. Here are the areas a generalized bot would need to succeed at:
1. Broad domain-specific knowledge base (at least as hard as IBM Watson on Jeopardy)
2. Image-to-text transformations of low-res images, with a very short time limit
3. The ability to adapt to the "rules of the game" at random
The current setup is limited, true, because the degrees of freedom haven't been fleshed out. I do think that's a fair criticism of this approach: the system needs a _much_ larger database covering different game engines, different knowledge domains, different assets (audio, images, videos), etc.
The other weakness I see (do you agree?) is that the game runs in client-side javascript. It should be fairly easy right now to attack the client-server interactions. I assume it's not just a "yes/no" response sent from the client to the server, but anything that reduces the complexity of the problem (i.e. simulating a thousand virtual games to learn the game mechanics) is probably not worth the effort to code up.
That's why this doesn't work. Sure, the CAPTCHA author gets to write code to generate their games, but you must not forget, the CAPTCHA attacker gets to write code too. If the author writes code to make the shoes any of 256 colors, the attacker gets to write whatever code they like to detect any of 256 colors and match them in code. If the author writes a trivia game with fixed answers, the attacker gets to hit the game until they get all the answers. If the author writes "What is $randint + $randint?", the attacker gets to write if ($question =~ /What is (\d?) \+ (\d?)/) { return $1 + $2 }, and so on.
If you look at it from an information theory perspective, the ability of the author to put information into their test is generally dwarfed by the ability of the attacker to put information into their cracker. When this is not the case, CAPTCHAs are uninteresting anyhow. That's why the only ones that have worked up to this point are ones that involve something genuinely hard for the attacker to express, because we know of no way to even express the distorted letter solver in code... or rather, we didn't, and now we do.
I'm curious because if you're saying a bot would be stopped at this solution, but not a human, then wouldn't that also apply to recaptcha as well?
Or are you saying one could automate a solution around this over that being impossible with recaptcha?
Would like to hear you thoughts.
In my opinion, it would be easier to automate this than recaptcha because this uses common elements that don't have nearly enough difference between them to make it a challenge. A solve rate of <5% is effective for spammers. How is this going to stop anything when a spammer can simply cycle through to a game they know how to beat, and fake the mouse movements and timing?
Recaptcha is difficult to beat with OCR because it is designed to withstand attacks by OCR considered 'state of the art'. It is impossible to beat with off the shelf solutions. On the other hand, this would be rather simple to automate using pixel matching. The shapes and colors don't change (Take the pancake example, for instance).
This system only works if there are unlimited, non repeating puzzles.
I could refresh til I get a desired puzzle (shoe matching). Then write a simple cursor controller program (say using Java Robot API). It would scan across the static shoe area on the right looking at shoe colors. Then scan across the shoes on the left. When it matches a color shoe, drag across to match.
observation_data: {"observationNumber":3,"mouse":[{"x":1,"y":68,"down":false},{"{"observationNumber":3,"mouse":[{"x":1,"y":68,"down":false},{"x":1,"y":68,"down":false},{"x":1,"y":68,"down":false},{"x":1,"y":68,"down":false},{"x":1,"y":68,"down":false},{"x":1,"y":68,"down":false},{"x":1,"y":68,"down":false},{"x":1,"y":68,"down":false},{"x":2,"y":55,"down":false},{"x":34,"y":50,"down":false},{"x":40,"y":51,"down":false},{"x":40,"y":52,"down":false},{"x":40,"y":52,"down":false},{"x":40,"y":52,"down":false},{"x":40,"y":52,"down":false},{"x":40,"y":52,"down":true},{"x":40,"y":52,"down":true},{"x":108,"y":52,"down":true},{"x":242,"y":50,"down":true},{"x":327,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true},{"x":336,"y":52,"down":true}],"clock":[16555,16597,16636,16677,16723,16755,16797,16837,16877,16917,16957,16996,17036,17075,17116,17157,17198,17236,17277,17319,17357,17396,17438,17476,17517,17557,17598,17638,17677],"bins":[[{"id":2,"time":14037}],[{"id":3,"time":16156},{"id":4,"time":17677}],[{"id":4,"time":17677}]],"items":[]}