For anyone initiated into the intricacies of the Fourth Amendment, it is axiomatic that probable cause "is incapable of precise definition or quantification into percentages." Maryland v. Pringle, 540 U.S. 366, 371 (2003). It is, instead, "a practical, nontechnical concept," Brinegar v. United States, 338 U.S. 160, 176 (1949), that refers to the amount of proof that would "warrant a belief by a man of reasonable caution that a crime is being [or has been] committed," id. at n.4. In the vast majority of cases, this conception of probable cause makes sense; trying to quantify the amount of proof with anything resembling numerical accuracy would be a frustrating exercise in futility. Consider the conversations one might overhear among a squad of officers on the side of the road, a group of prosecutors in an office, or a judge and her clerks in chambers as they try to calculate how different pieces of evidence affect the amount of certainty in a given case:
"The officer's observations establish 70% certainty."
"That's reduced by the 3% penalty for each twenty-foot increment between the officer and the subject, and the penalty is multiplied by two because it was dark out."
"But there is no penalty for the first fifty feet, and it was twilight, not fully dark, so the distance penalty only gets multiplied by one-and-a-half."
That snippet of hypothetical conversation may evoke the lively exchanges heard during tabletop role-playing game sessions, but it would be an absurd approach to assessing probable cause. Most of the time.
In some cases, though, evidence is capable of more precise evaluation, if not to the level of exactitude suggested above. Police canines are one example. When a canine handler is preparing to walk a drug-detecting dog around a vehicle, the dog's work history is a known quality. There are, or should be, records that document the dog's performance in controlled circumstances as well as in the field. These records may be reviewed to determine the relative weight that should be put on a drug-dog's "alert" (the signal that a police canine has detected one of the substances it is trained to recognize). An alert by a police canine with, say, a 95% accuracy rate could be considered quite strong evidence, while an alert by a dog who performed at or worse than chance would carry very little weight, if any. If this were true, canine handlers--who use their dogs specifically to establish probable cause that allows them to search a car or get a warrant for a house or package--would have a strong incentive to make sure that their dogs were well trained and accurate.
Such a regime may be appealing on a number of levels, but it is not the approach taken by the Supreme Court. In a unanimous 2013 decision, the Court held that an alert by a drug-detecting dog that had achieved "satisfactory performance in a certification or training program" was presumptively reliable. Florida v. Harris, 11 No. 11-817. (U.S. Feb 19, 2013). A criminal defendant could attempt to rebut that presumption, calling into question a dog's training or performance in the specific case at bar, but even if the defendant did so, the dog's historical performance in the field was "of relatively little import":
[I]f the dog alerts to a car in which the officer finds no narcotics, the dog may not have made a mistake at all. The dog may have detected substances that were too well hidden or present in quantities too small for the officer to locate. Or the dog may have smelled the residual odor of drugs previously in the vehicle or on the driver's person. Field data thus may markedly overstate a dog's real false positives.
Florida v. Harris, 10 No. 11-817. (U.S. Feb 19, 2013). Although the Court acknowledged that a dog's false positive rate in the field "may sometimes be relevant," id., it viewed performance in a controlled setting as a more appropriate measure of a dog's accuracy.
As a result, courts have approved alerts by police canines with troublingly low accuracy rates in real world situations. The Fourth Circuit, for example, has held that an alert by a police canine with an accuracy rate of only 43% (in the field) was sufficient to establish probable cause. United States v. Green, 19 No. 12-4879 (4th Cir. Jan 17, 2014). And that was after including vehicles in which "officers found direct evidence that drugs or drug users had recently been in the vehicle," even though no actual drugs were found (otherwise, the dog was accurate only 25.88% of the time). Id. Nor does such a low accuracy rate appear particularly unusual. According to the Chicago Tribune, police canines in Chicago had an overall accuracy rate of 44% over three years. An experiment involving 18 police canines at the University of California, Davis found an accuracy rate of only 15%. Ironically, this may be because of how police canines are trained; dogs may become conditioned to expect contraband far more consistently then they will encounter it in the field.
Acknowledging that the accuracy of a drug-dog's alert raises "[n]agging questions," the Seventh Circuit recently upheld a district court's similar conclusion about Lex, a dog who alerted to 93% of the vehicles that he sniffed, but had an accuracy rate of only 59.5%. United States v. Bentley, No. 13-2995 (7th Cir. Jul 28, 2015). In other words, he alerted almost every time he was called to sniff a vehicle, yet was only marginally more accurate than a coin flip would have been in determining whether there were drugs in those vehicles.
This is worth repeating: Lex alerted to almost every car that he was called out to sniff. Assuming, as the government argued, that the officers typically called for a canine to sniff vehicles that they already suspected contained drugs, we might expect a relatively high alert rate. A high alert rate and a high success rate would indicate that the officers were good at identifying vehicles that had drugs in them. A low alert rate with a high success rate would indicate that the officers weren't very good at identifying such vehicles, but the dog was. Lex, however, had a high alert rate and a low accuracy rate. That suggests that neither the officer nor the canine were very good at identifying vehicles that actually had drugs in them. Worse, it suggests that Lex may have been reacting to the officer's behavior, rather than the vehicles he was smelling.
So what explains all the false alerts? One possibility, certainly, is that the drugs in some vehicles were "too well hidden or present in quantities too small for the officer to locate." Florida v. Harris, 10 No. 11-817. (U.S. Feb 19, 2013). Inevitably, this is the preferred explanation of canine trainers and handlers; it is not uncommon to see some reluctance or outright refusal to acknowledge false positives. The transcript of a 2004 New York case hints that "unverified alerts" is the preferred phrasing :
Q: Do you keep records as to the effectiveness of your dog?
A: Yes, sir, I do.
Q: Do you know how often your dog gives false positives?
A: He doesn't give any false positives. We're just unable to verify the alerts at the time.
Another possibility is that Lex was responding to positive reinforcement training rather than the presence of narcotics: the handler rewarded the dog after every alert in the field, regardless of whether drugs were eventually discovered or not. Thus, it was Lex's action (the alert itself), rather than the successful detection of contraband, that Lex was rewarded for. And that, of course, means that Lex was trained--inadvertently, one hopes--to alert regardless of whether he actually detected drugs. The Seventh Circuit rightly criticized the practice, calling it "a terrible way to promote accurate detection on the part of a service animal." United States v. Bentley, 10 No. 13-2995 (7th Cir. Jul 28, 2015). Nevertheless, the court held, Lex's history was "good enough to support a finding of his reliability and thus to allow his alert to constitute a significant piece of evidence supporting the ultimate conclusion of probable cause." Id. That history certainly was not enough to overcome the deference owed to the district court's factual findings.
Bentley perpetuates a rule that establishes perverse incentives. Police canines will never be perfect. In fact, in a 2006 George Mason Law Review article, Richard E. Myers, II, concludes that even a well-trained detection dog will have more false positives than valid alerts. But they can be trained to detect contraband in a way that reduces the chances that innocents will be subjected to invasive searches. Allowing officers to rely on dogs with a high-alert, low-accuracy performance history, however, changes them into rubber stamps; if a dog alerts to almost all the vehicles that he is called to search, he is little more than a furry license for officers to search almost any car they might want to. This is particularly problematic given the racial dynamics in modern policing. In its analysis, the Chicago Tribune found that police canines had an overall accuracy rate of 44% , but an accuracy rate of just 27% with regard to Latino drivers.
In Bently, the Seventh Circuit said, "We hope and trust that the criminal justice establishment will work to improve the quality of training and the reliability of the animals they use." And certainly it could do so. Dogs could be trained not to alert on weak odors, so as to avoid hitting on contraband "present in quantities too small for the officer to locate." Both canines and handlers could be trained in a way that minimizes the possibility that the dogs will alert to handler cues, rather than actual scents, thus providing a check on officers' implicit and conscious biases. States could establish minimum standards for such training (as of 2012, only two states had done so). More rigorous scientific testing and a national reporting requirement could provide the information policy-makers and trainers need to identify best practices, performance averages, and outliers.
All of these things are possible, but right now, the Fourth Amendment rules governing police canines not only do not encourage such changes, they effectively discourage them.