Parliament gives an official a power, and the official makes a decision you don’t like. So you go to a judge to have the decision reviewed. With any luck, the judge will agree with you and give you a remedy. That’s the normal way of things. But if Parliament is sovereign, then Parliament has the power to prevent the judge from reviewing the official’s decision. All Parliament has to do is include an “ouster clause”: a clause communicating its intention to make the official’s decisions unreviewable.

In the past Parliament has enacted what looks like an ouster clause. It did that most famously in Anisminic Ltd v Foreign Compensation Commission. But judges have tended to say that what looks like an ouster clause is, on closer inspection, not an ouster clause. It seems as if Parliament wants to oust the courts, but really it doesn’t intend that. Now, the clause in Anisminic was a bit ambiguous. It wasn’t as clear as one would wish. So, ever since Anisminic, it’s been a matter of speculation what judges would do when faced with a more clearly drafted clause.

With that as background, the interest in R (Privacy International) v Investigatory Powers Tribunal is understandable (DC | CA). It looks like exactly the sort of test case people were waiting for. There is a putative ouster clause, more clearly drafted than the one in Anisminic. What would judges do? Well, we know the Divisional Court and the Court of Appeal would do – they’ve both said that there is an ouster clause. As a result, they’ve found that certain decisions are immune from judicial review. But it’s likely that the case will be appealed to the Supreme Court, so we’re still waiting for the last word.

I want to show that Bayes’ Theorem provides a helpful way to think of ouster clauses. Bayes’ Theorem is much-discussed in evidence law, but rarely invoked in statutory interpretation. I think it suggests – I don’t put it any more strongly – that Privacy International may have been wrongly decided.

Privacy International

The Regulation of Investigatory Powers Act 2000 (RIPA) establishes the Investigatory Powers Tribunal (IPT). Among other things, the IPT hears complaints against the intelligence services. It can receive evidence which is inadmissible in a court of law. It can conduct its proceedings in private, and sometimes in the absence of the complainant. It’s a very unusual tribunal. It’s designed to protect the public interest, but obviously it places the rule of law at risk.

Adding to those rule of law concerns, s 67(8) RIPA appears to shield the IPT from judicial review:

Except to such extent as the Secretary of State may by order otherwise provide, determinations, awards and other decisions of the Tribunal (including decisions as to whether they have jurisdiction) shall not be subject to appeal or be liable to be questioned in any court.

The issue in Privacy International was the effect of s 67(8): does it oust the courts, or not?

In the Divisional Court, Sir Brian Leveson P held that s 67(8) does oust the courts. Leggatt J didn’t formally dissent, but he disagreed about the effect of s 67(8). For Leggatt J, the wording of the section wasn’t clear enough to show that Parliament wanted to oust the courts.

In the Court of Appeal, Sales LJ (who wrote the only substantive opinion) said that ‘on its proper construction’, s 67(8) does ‘clearly mean’ that the IPT’s decisions aren’t judicially reviewable. We know this is what s 67(8) clearly means because of ‘the language used in the provision’, as well as the ‘legislative context’.

All three judges wrote careful and thorough opinions. So, whose view is more persuasive – Sir Brian Leveson P’s and the Court of Appeal’s, or Leggatt J’s? Bayes’ Theorem suggests a way of figuring that out.

Bayes’ Theorem

Bayes Theorem is a mathematical formula which helps us understand how to update a hypothesis given new evidence. The hypothesis is represented as H. There is some prior probability of the hypothesis being true, represented as Pr(H). That prior probability is called the “base rate”. We’re faced with new evidence, represented as E. What is the probability of the hypothesis being true, given the evidence we now have? Bayes’ Theorem tells us the answer.

According to Bayes’ Theorem, the probability of H given E – represented as Pr(H|E) – is a function of the probability of E and the probability of E given H. With some substitutions, that means:

Bayes Theorem is simple but powerful. It helps identify likely causes from effects. It’s also a useful tool for avoiding cognitive biases.

Here’s an example, using a problem invented by Amos Tversky and Daniel Kahneman. There’s a town with only two taxi companies: Green and Blue. Green uses green cars. Blue uses blue cars. Green is the more successful company: it has 85% of the taxis on the streets. One night, a taxi sideswipes a car. A witness says the taxi was blue. The witness is tested, and she can reliably identify the colour of a taxi, green or blue, 80% of the time. What’s the probability that the taxi is blue, given that the witness says that it’s blue?

Tversky and Kahneman showed through experiments that the usual answer is 80%. After all, the witness gets it right 80% of the time, and she said the car is blue. It’s intuitive that the car is probably blue. But the intuition turns out to be wrong.

These are the relevant probabilities:

Plugging these numbers into Bayes’ Theorem gives us:

There’s a 41% chance that the taxi was blue. There’s a 59% chance that the taxi was green. Probably, the car belonged to Green, despite what the witness said. People give the wrong answer because they focus on the witness’s reliability. They neglect the fact that most taxis in the town are green. That is, they underestimate the significance of the base rate. The tendency to err in this way is known as the “base rate fallacy”.

It’s easier to see what’s going on if we imagine the incident happening 100 times, randomly selecting cars each time. About 85 of the incidents will involve a green taxi. Of those 85 times, the witness will get the colour right 80% of the time, identifying 68 cars as green. She’ll incorrectly identify 17 cars as blue. Of the 15 incidents which involve a blue taxi, the witness will say that 12 of them are blue, and 3 green. The results are shown here

Of the 29 cars which the witness says are blue, only 12 really are blue. That is, of the 29 cars which the witness says are blue, the majority – 17/29 or 0.59 – are actually green!

Applying Bayes’ Theorem to Privacy International

We can use Bayes’ Theorem to make sense of the likelihood that Parliament intended to include an ouster clause in RIPA. We have evidence of Parliament’s intent, in the form of the text of s 67(8), the statutory context, rules of syntax and semantics, etc. The question is: given that evidence, what is the probability that Parliament intends to oust the judges? I’ll assume for the sake of argument that the standard of proof is the balance of probabilities.

Let’s start by estimating the base rate. Setting aside RIPA for the moment, how likely is it that Parliament would include an ouster clause in a statute? To my mind the answer is: very unlikely. I say that for two reasons.

• Ouster clauses are disreputable. Ouster clauses free an official from external and effective constraints. As a result, they give an official arbitrary power. Arbitrary power is at odds with the rule of law. The rule of law is of concern to Parliament, so Parliament isn’t likely to want to undermine it.
• Past practice. Conferring powers on officials and tribunals is the bread and butter of legislation. Parliament has conferred this sort of power thousands and thousands of times. Yet, if judges are to be believed, they’ve never encountered an ouster clause in the past, not in Anisminic and not in any other either.

Now, there are qualifications. Maybe this ouster clause isn’t such a threat to the rule of law (because, say, the IPT has High Court judges on it). Maybe judges aren’t to be believed; really they have encountered ouster clauses before. And then there’s the fact that not every statute makes it to court, meaning there may be ouster clauses which haven’t been acknowledged as such. So, the prior probability that Parliament intends to oust the courts may be low, but it’s not ridiculously low. Let’s put it at 1%.

Next, let’s put a number on the likelihood that, if Parliament intends to oust the courts, it will include a clause like the one in RIPA. I’ll assume that every time Parliament wants to oust the jurisdiction of the court, it will use an ouster clause. That is, I’ll assume a probability of 100%.

Finally, we need to estimate the likelihood that, if Parliament doesn’t want to oust the courts, it will still use a clause like the one in RIPA. Now, it would be poor draftsmanship to use language like we find in RIPA if parliament doesn’t want to oust the courts. So, what we’re estimating here is the likelihood of error. But, of course, drafting errors do happen. Perhaps no one thought it through. Perhaps Parliament intended to oust the courts to some extent but not fully. I’ll assume Parliament is pretty reliable, and peg the probability at 2%

Putting all this together, we have these probabilities:

Here’s the calculation:

Even given the language of s 67(8), the statutory context, and so on, the likelihood that Parliament intends to oust the courts is only 33%. With a standard of proof of above 50%, a Bayesian court would conclude that Parliament doesn’t intend to oust the courts.

A chart makes the argument more intuitive. Suppose we took 100 statutes in which an ouster clause might be included. Then we should see the following:

Parliament uses language like that found in RIPA in 3 cases. But it’s only in 1 of those cases that Parliament actually intends to oust the judges. So, the odds that Parliament intends to oust the judges are 1/3 = 33%.

Conclusion

My numbers are speculative, of course. With some changes, we can arrive at a conclusion that supports the Court of Appeal’s judgment. For example, if we suppose that the base rate is much higher at 5%, then other things being equal there’s a 72% chance that s 67(8) includes an ouster clause. Similarly, if we suppose that the probability of seeing language like that found in RIPA when Parliament doesn’t intend to oust the courts is just 0.05%, then other things being equal there’s a 67% chance that s 67(8) is an ouster clause.

My aim here isn’t to push a particular conclusion. It’s to identify the questions. A Bayesian court should focus on two questions: what’s the prior probability that Parliament wants to oust the courts? And, what’s the probability that Parliament would use language like we find in RIPA and not intend to oust the courts?

Note to readers: I’ve benefited from Ian Hacking, An Introduction to Probability and Inductive Logic (2001); Yair Listokin, ‘Bayesian Contractual Interpretation’ (2010) 39 Journal of Legal Studies 359; and Amos Tversky and Daniel Kahneman, ‘Evidential Impact of Base Rates’ (1981) Office of Naval Research.