Putting aside any actually truthful captions, how do I know that "image of X" is actually an image of X?

Reading some of the Bellingcat investigations, and time spent, doesn't bode well.

I guess you could TinEye and index/hash the entire web's worth of rich media, then spot discrepancies (listed as X here, but Y there), but that seems horrendous in compute/bandwidth/storage terms.

No, it's not. This is done because stories with images perform better, and obtaining images (& licenses) for photos of every event is not always possible.