After reading Jacob's piece from last week about the Supreme Court ruling that a drug dog's sniff isn't a search, I got to thinking about the problem of error rates. Jacob cites Souter's dissent, which references cases involving dogs in the field giving false positives 8 percent of the time or more and studies where in "artificial testing situations" they gave false positives between 12 and a mindboggling 60 percent of the time.
A few fairly obvious problems occur to me belatedly. The first is that to the extent that law enforcement officers now feel increased license to do indiscriminate sweeps, the conditions under which prior accuracy rates were ascertained in the field no longer apply. Presumably, dogs are typically brought in when there's already some independent reason to think the person thus searched is holding drugs. If police begin to feel they don't need that reason, we can expect the proportion of false alerts to rise.
The possibility of a shift to indiscriminate sweeps presents a related problem from another direction as well. Let's grant that the dog is 95 percent accurate. Now, you might think that sounds pretty good—95 percent certainty would surely count as probable cause, right? The problem is making the error—and I wonder whether maybe the justices did this—of inferring from a 95 percent accuracy rate that you're only going to end up physically searching one innocent person for every 19 who really do have drugs. But if searches are indiscriminate, that's wrong, because the vast majority of motorists won't have drugs.
Assume, and I suspect this is a big overestimate, that one in 100 motorists are driving around with drugs. If you're sweeping them randomly, a 95 percent accurate dog is going to "alert" for five innocent people for every one it catches with actual drugs. And those five people are going to get their cars torn apart. If the dogs are less accurate, it'll be many more than that. Anyone think that passes Fourth Amendment muster?
Addendum: Commenter Shannon Love turns up a handy-dandy java applet illustrating the problem mentioned above. Just substitute "drug possession" for "infection" and you can test the likely results for different levels of dog accuracy and motorist drug possession. On the extremely generous assumptions above (1 percent of motorists have drugs; dogs are 95 percent accurate), the chances that someone IDed in a random sweep (and therefore subject to an extensive and intrusive vehicle search) actually has drugs are about 16 percent. Relax those assumptions and the picture quickly gets far worse.