The immediate and obvious problems stem from self-satisfying actors. "Oh, 10,000:1 on blowing up Country Y's embassy in Country Z this week! Lets go make some money boys."
If a terrorist plans to blow up an embassy and places bets on this action, he's directly transmitting his plans to the public. The price of "embassy bombing" skyrockets, security agencies are notified, and his odds of success go way down. (His odds of surviving the attack and collecting his money also aren't that high anyway.)
Now consider the possibility that the terrorist leaks information to his supporters/criminal underworld/etc - e.g. the gangsters selling him semtex. These guys would never betray him, but on the other hand they see some free money sitting on the table. They place bets on an attack, the price goes up, and again agencies are notified. That's a good thing - a truly anonymous way for criminals to probabilistically inform on each other.
When smart people like Robin Hanson (who was in part behind this proposal) and various CIA types get behind a proposal, you should always think twice about the "immediate and obvious problem". The folks proposing this aren't stupid, so they probably thought about it too.
Imagine a big corporation studying the feasibility of a large capital outlay for some factory in a semi-stable country. It would be very valuable for them to know the probability of a government coup, the odds of a natural disaster, and various other hard to estimate metrics before they spend their X billion dollars. If there was a highly liquid and active global prediction market, they could look up the current odds of any event that interests them, and factor it into their feasibility, without ever having to gamble.
Additionally organizations can better protect their businesses by using these markets to hedge against events that aren't covered in the few positions available on financial markets.
I could go on and on with valuable use-cases that go far above the gambling.
I'm the lead UI engineer at Augur, a decentralized blockchain-based prediction market, and we're attempting to solve this. I really hope we do, the world would be a better place.
The idea itself is a bit older: the Iowa Electronic Market had been taking election bets since the late '80s, but they exploded in popularity in the mid 2000s. HP had an internal prediction market that allegedly estimated DRAM pricing very accurately. Pfizer had one where employees could "bet" on new product ideas and there were tons of startups trying to build them for movies, sports, etc.
I haven't heard much about this since then, or perhaps the idea of crowdsourcing has just become pretty standard.
