If your tool has a 50% failure rate, you can flip a coin and get the same results.
That's not how it works at all. If you got four people in a line-up there's a 1/4 = 25% chance you'll find the right person by coin toss. If I can make a tool that's right half the time I've doubled that. Along the way we're redefined the question from how hard the underlying problem is to how successful we are at it.
If I could make a machine that'd give me the lottery jackpot numbers half the time I'd be ecstatic. Sure, I'd only win every other week but given that it's one set of winning numbers and hundreds of millions non-winning numbers each week I'd gladly take the 50% failure rate.
If you take a database of ~10000 celebrities, your coin toss would have 0.01% accuracy by luck. Modern facial recognition algorithms get about 99.8% so like ten thousand times better. The problem is they want to find your face in a database of 640 million photos. One needle and 639.999.999 straws.