An article from Ars Technica on Crows not being as smart as we once thought. Apparently they’re good at meat-attached strings, but not at building mental models of said strings:

What is the difference between model-based solutions and feedback-based solutions? When we rely on feedback, we first perform an action—pull on the string and trap it underfoot—if we perceive that we are closer to our goal (the meat is now closer), we repeat the action. A model-based solution, on the other hand, involves understanding that the meat is connected to a bit of string, and that to get the meat, we must pull the string up. In the second case, feedback after each step is not required, because we understand the problem and know that we will be rewarded in the end.

To build a model requires a deeper understanding of the situation: that there is string, and it is connected to food. We know that crows understand the connectivity between the string and the meat to a certain extent. For instance, if there are multiple strings, the crow will usually pull the correct one. If the strings are crossed, but have different colors, the crow will still usually pull the correct one.

But if you cover the string with a bit of wood with a small hole in it (so the crows can either look or pull but not both at the same time):

The authors found, as expected, that the crows given the normal string test figured it out quite fast. Of those given the test involving pulling the meat up through a hole in a piece of wood, only one crow succeeded. Further, the birds that were experienced in solving the normal string test performed much worse when faced with the more difficult test. This combination suggests that even experienced crows really need constant feedback to solve the problem. Conclusion: crows don’t make models.

I must admit to having a little problem with that conclusion. First, one crow did solve the problem; second, the crows all varied widely in their performance on all of the tests, suggesting that problem-solving abilities vary wildly between individuals—no surprise there. Finally, I think the distinction between model-building and feedback-based problem solving skills are artificial points in a mental toolkit that spans a continuum.

The point is that, even when we make models of the world, we rely on feedback to validate those models. Our previous experience also influences how quickly we abandon a model in the absence of positive feedback. These experiments show that, if crows do build models, they don’t generalize them very well, and they require a fair bit of reinforcement before they’ll abandon the model. But, other than the amount of reinforcement involved and the complexity of the model, is this that different from human behavior?

To be really human, the crows would need to be able to create a model that is completely non-feedback based; researchers would have the toughest time proving it though.