Don’t do a poll. What people say they like doesn’t always correspond to what they do. If you’re going to take the blank-button approach, put users in a situation where they need to click OK (and another situation where they need to click Cancel) and force them to guess which one to click.
It’s probably better to have a more realistic UI which has the labels. However, I wouldn’t rely on measuring error rates alone. The truth is users very rarely click the wrong button in a dialog of a desktop app no matter which button is left and right. That’s one reason debate on this topic has yet to be resolved –one way doesn’t blatantly out-perform the other. In addition to measuring error rates, also measure response time to the millisecond. If users are the least bit confused by the button positions, it will show up as a slight hesitation. This saves you from having to unrealistically rush your users (but it’s okay to provide realistic pressure). Also, efficient user performance is its own usability benefit apart from accuracy.
For the test, try to come up with a task where the choice between OK and Cancel is naturalistic but you can tell definitively when the user is right or wrong (just in case there are some accuracy issues). For example, you could have a spell-check feature that goes through a body of text and sequentially presents via a dialog box each word that may need correction. By design, some words need to be corrected (user should select OK) and some should not (user should select Cancel). Correct and incorrect spellings should be obvious to anyone with reasonably good education in English in order to minimize variance.
Try to recruit users with various levels of Apple OSX versus other platform experience, and record each user’s experience with Apple and others (how often and how recently), and use that information as covariates in your analysis.
Analyze your results with inferential statistics. If no one on the team is comfortable doing inferential statistics, then get someone who is. Because button position may not make much difference, select a difference in accuracy and response time that is worth caring about and determine the sample size you’ll need to reliably detect that difference by performing a statistical power analysis using the variance from a few pilot-test users. Be prepared to run a lot of users.
Publish your results, preferably in a peer-review journal. I think we’d all want to know the answer.