How is it any different than having a conversation with someone that's in the car?
There are, I would assume, at least two significant ones.
One is that you're putting some mental effort ("cognitive load" in the jargon of these studies) into figuring out what to say to Siri and how to interpret its responses and reply accordingly. That is, you're making a mental "map" of the interaction, a little like a phone tree, which takes some attention away from the road. A human in the car is much more likely to correctly interpret what you're asking for and adjust or repeat instructions as needed.
The other is the same problem you have with a cell conversation, even handsfree--the other human in the car is also aware of the situation around you, and will adjust (or even assist) accordingly. For example, if there's a hazard coming up, they might break off the conversation temporarily, or actually point it out to you if you didn't see it.
Even dealing with a car navigation system has some of this; with each instruction (unless it's really obvious, like an offramp on a small freeway or a turn on a country road), you're trying to figure out exactly how to interpret the "turn left ahead" instruction as it applies to the intersection(s) in front of you and surrounding traffic. With a human navigator, they point and say, "turn right at that sign up there" or "Follow that yellow truck." I know that Garmin will even sell you a navigation add-on that claims to give real-world landmark directions when telling you where to turn to simulate this.
This study honestly doesn't surprise me at all--I wouldn't really want to do more than the most rudimentary task with Siri on the road, and definitely not in heavy traffic.
That said, on a recent road trip, it was tremendously useful to be able to just say "Give me directions to the nearest Dairy Queen" while on the freeway somewhere in Nebraska and have Siri tell me what exit to take. I would not otherwise have been able to get that information without getting off of the freeway and doing a manual search.