[maxerickson]

There's no margin of error for the test they did, just potential sampling bias.

Given the 5000 questions they asked, the system will provide the wrong answer 2.6% of the time. Every time, until they improve it. There's a chance that they managed to ask the only 150 questions that it doesn't know the answer to, but not a very big one.

[johnhenry]

I agree -- given the sample space, there is not effective way to calculate the margin of error. Even so, I don't think that there are many examples of non-deterministic mechanisms that produce the correct answer 97% of the time.