Would need to know how the sample was chosen for that test. If you sample women of reproductive age at random, there's decidedly not a 50/50 chance that she's actually pregnant: by rough estimation it should be about (1.75 [fertility rate] * 9 months [duration of pregnancy] / 30 years [reproductive lifespan] = 4-5%). If you give all of those woman a wheat/barley test and 70% of the positive tests were truly pregnant (implying about 6.25% test positives) that's actually pretty good, roughly 8x better than a coin flip. If you take a sample that's known to be half pregnant and half not pregnant and only 70% of the pregnant women are identified, it's decidedly less good.

So, 100% is 50% better than random? Weird math.

Depends on the true distribution of the classes. Although the naive strategy gets even higher accuracy if the classes are unbalanced.

Sure it's better than flipping a coin, but how does it compare with just making an educated guess?

For some demographics you might get a 70% success rate with something as stupid as "Have you been trying to get pregnant for a few months? If so assume you are."

Might invest in a bushel basket.