False Sharing in a Parallel Counter

Two threads each increment their own counter one hundred million times:

#include <thread>

long counts[2] = {0, 0};

void work(int i) {
    for (int k = 0; k < 100'000'000; ++k) {
        ++counts[i];
    }
}

int main() {
    std::thread t0(work, 0);
    std::thread t1(work, 1);
    t0.join();
    t1.join();
}

On a multicore machine this runs slower than running the two loops sequentially on one thread. What is the most likely cause?