The Book of Why by Judea Pearl and Dana Mackenzie

Perhaps, after hearing rumours that artificial intelligence (AI) will replace judges and juries, you have picked up a book called The Book of Why – The New Science of Cause and Effect by Judea Pearl and Dana Mackenzie (2018). This sounds like just the thing to bring you up to speed on that topic.

The book is not really about replacing judges and juries, although it does deal with how to get computers to find causes by analysing data. Conventionally, unless a randomised controlled experiment has been conducted, data produced in an experiment says nothing about causes, only about correlations, although observational data, such as (for example) may be gathered in surveys and analysed by statisticians, can generate conditional probabilities from which causal effects are inferred. The main focus of the book is on how to avoid the need for randomised controlled experiments, and yet be able to say something about causes, or at least how to identify when there is no alternative to conducting a randomised controlled experiment to determine causal effects.

What is of potential interest to lawyers, it seems to me, is the diagrammatic representation of causal relationships that can arise in given circumstances. Can the kinds of diagrams described in this book give lawyers tools for analysis of issues in trials, where the fundamental question will be whether the defendant caused the relevant harm?

Here I have borrowed from Chapter 4 some of Pearl and Mackenzie’s examples of the use of causal diagrams. What I borrowed is the diagrams, and I have tried to apply them to the facts of cases I have imagined. Any mistakes are therefore mine, not the authors’. Decide for yourself whether this kind of approach may be of assistance in your work. There are, as the authors acknowledge, other ways of addressing whether cause has been established, but here is the causal diagram method, inasmuch as I currently understand it.

The authors claim the method is such fun that it can be thought of as a game.

The letters represent events or states of affairs (or, as one could say, variables, in which we may have varying degrees of belief), not people. I illustrate this in Game 1 for X and A, just so you get the hang of it.

Game 1

                                       

Defendant pulls trigger (X = this action by defendant)), gun fires (A = what the gun was caused to do), cartridge casing ejects (B), bullet hits victim (Y).

Here there is nothing to block the causal chain from X to Y. B is irrelevant to whether the firing of the gun caused the bullet to hit the victim.

Game 2

                                          

The prosecutor charges the defendant with participating in the manufacture of an illegal drug on two occasions, C and Y. The prosecutor's case is that the defendant had an appetite (A) for the illegal drug and that this appetite grew (X). On each occasion the appetite for the drug caused the defendant to collect (B and E) quantities of a substance that could be used to make the illicit drug, and his possession of this substance caused him to be accepted as a participant in the manufacturing process. However, the defendant explains (D) the accumulation of the substance by saying that the appetite is for that substance and that there was no purpose in accumulating it except to consume it.

Blocking (constriction) of a causal chain can occur if there is more than one possible cause for an event or a state of affairs. Here, there is a blockage in the prosecutor's causal chains A to C, X to Y, and B to E. B to E is blocked by the defendant’s explanation D. If the blockage is strong enough to maintain the causal direction D to B then C will be irrelevant to Y and A will be confined to enhancing X, which again will be irrelevant to Y because of the blocking of E. X to Y is blocked if D to E is of sufficient strength (ie if it carries sufficient credibility to raise a reasonable doubt about X to Y).

Game 3

                                      

                                           

This diagram can be taken in stages to reflect the course of a trial. The prosecutor alleges the defendant hated (X = the defendant's emotion) the victim, and this caused the defendant to kill (Y = what happened to) the victim. Hence the first step is the arrow from X to Y.

The defendant claims that it was another person who both (B) caused the defendant to hate the victim and killed the victim. There is now an arrow from B to Y. This creates a back-door cause, B to Y, of the victim’s death by the hand of that other person, which weakens the causal connection X to Y.

The prosecutor replies by adopting the defendant's allegation that the other person caused the defendant to hate the victim. Hence the arrow from B to X. The prosecutor argues that the causal chain B, X, Y strengthens the link X to Y.

The defendant adduces evidence of an alibi, that he is identified by security camera footage as being somewhere else at the time the victim was killed. This is represented by the arrow from X to A. The defendant hopes this will reduce the strength of the alleged causal link X to Y.

The prosecutor replies by challenging the identification of the defendant in the footage and says it really showed the other person who hated the victim. This is represented by the arrow from B to A.  

If the arrow B to A were to be reversed, because the defendant's alibi made it more likely that the other person who hated the victim was the murderer, there would be a causal chain A, B, Y, which the defendant would hope would reduce further the alleged link X to Y. The prosecutor's attack on the defendant's alibi was an attempt to block the suggestion that the other person was the murderer.

Game 4

                                         

Prosecutor alleges the defendant has large accumulated debts (X) which have caused the commission of the present fraud (Y), relying on propensity evidence: when the money problems began (A) they led to the defendant committing fraud (B), and contributing to that earlier fraud was a propensity for dishonesty (C), which also caused the presently alleged fraud (Y).

The proposed causal chain generated by the propensity evidence is blocked by C to B. This reflects the legal position that propensity alone is insufficient to prove current offending. There must be some evidence of current offending before the propensity evidence has anything to corroborate. If there is some such evidence, the admissibility of the propensity evidence will depend on the extent to which the prosecutor can reverse the causal link from C to B, for example by showing that previous commission of fraud (B) strengthened the defendant’s propensity for dishonesty. If the link then went from B to C, there would be no blockage in the causal chain X, A, B, C, Y.

Game 5

                                         

This is said to be a more realistic version of Game 4, with two new causal links B to X and X to Y. Using facts like those in Game 4, X to Y is the evidence necessary before propensity evidence has anything on which to operate (that is, to corroborate). B to X is the allegation that previous fraud (B) has caused present indebtedness (X), for example by resulting in reparation orders of a magnitude beyond the present means of the defendant.

Now the alleged causal link X to Y is supported by A, X, Y, and by A, B, X, Y, and by C, B, X, Y.

So, how secure is the position of judges and juries in coming to verdicts? Computers can't perform experiments, they can't use what the authors call "do-operators" to change things in the experimental environment. Where there are causal chains, data may be available to be used to generate conditional probabilities which AI can handle. But where a blocked chain cannot be ignored as being irrelevant, an experiment will be necessary to provide more data. A computer could not, therefore, play Game 2 because of the blocking effect of D (which would require an investigation of the likelihood that the defendant did consume the substance), or Game 3 because of B to A (unless it was able to compare the video image with the defendant, although there may be legal objections to that, or with the other person), or Game 4 because of C to B. Judges and juries surmount these difficulties with something ineffable called judgement.

Update: for more on AI and probabilities, including reference to this book by Pearl and Mackenzie, see the review by Paul Taylor in London Review of Books, 2021 vol 43, p37.

Go back!

What should be done about the failure of the Psychoactive Substances Act 2013 to control unapproved psychoactive products?

Unapproved psychoactive products should be controlled drugs. As such, their classification should follow recommendations of the Expert Advisory Committee on Drugs, and pending any advice to the contrary they should be Class C controlled drugs.

An arguably premature revision of the mental elements for liability for drug offences was undertaken in Cameron v R [2017] NZSC 89, [2018] 1 NZLR 161 (not freely available online, strange to say). [Background: perhaps the delay in its availability was due to outstanding proceedings against an absconding defendant keeping suppression orders in force until his death, reported on 2 March 2019 - although there had been extensive media coverage in 2018 of the entire case up to sentencing, with everyone named. No, that's not the reason. There is no reason.] The new requirement of knowledge that the substance was a controlled drug, or of recklessness as to its being a controlled drug, is too narrow in that it pinpoints “controlled drug”. This means knowledge of, or recklessness as to, the substance being scheduled in the Misuse of Drugs Act 1975, including knowing it by a name that is a common term for a scheduled substance (see Cameron at [40], [93]).

The problem here is that under this requirement it would be a defence for a person to have thought that the substance was an unapproved psychoactive product (there are currently no approved products, and none are in the process of being tested for approval).

Under the pre-Cameron law, the mental elements for guilt included knowledge that the substance was a drug that had an effect on the mind. This requirement, settled for some 30 years after some interesting cases were digested, did not give rise to intrinsic difficulties of proof. If anyone asked what I thought, I would say, go back to that. Cameron is suited to an anticipated future in which there are approved psychoactive products and it is necessary to recognise the innocence of people who honestly think they have such a product when in fact they have a controlled drug. On this view, we have here a rare example of reality not having caught up with the law.

As far as controlled drug analogues are concerned, we should wonder whether the legislated criterion of "substantially similar" structure to a controlled drug is really on point. The purpose of the Act is to prevent the harm that is caused by the use of controlled drugs, and substantial structural similarity serves (whether accurately or not, we don't necessarily know) as a proxy for that. In any event, for liability, mens rea should address the effects of the substance when used. If requiring knowledge of, or recklessness as to, structural similarity seems somewhat absurd (it has never been necessary for a defendant to know that a substance was an isomer, a salt, an ester or an ether of a scheduled substance) courts should be more comfortable assessing whether a defendant knew of, or was reckless as to, the effect of the substance on the mind of a user. People who manufacture and distribute such substances should be well aware of their effects when used, and of the similarity of those effects to those caused by use of controlled drugs. So, for offences in relation to analogues, there is no reason to treat the mental elements for liability as being any different from those applicable to other controlled drug offences. Our current legislation follows international precedents, at least as to definition of the actus reus, but that should not stop the search for improvements in the case law definition of the elements of mens rea.

It's a matter for you ... but you may well think ...

The High Court of Australia has accepted that, as a general rule, judges presiding at jury trials should not comment on issues of fact: McKell v The Queen [2019] HCA 5 at [31], [46]. Comment should only be used to correct an error or restore balance: [53]-[55]. The power to comment is to be used to ensure trial fairness, and is not to be used to add force to one side so as to sway the jury: [3].

It is important, obviously, that perceptions of trial fairness be kept consistent internationally. One reason for this is so as not to impede extradition processes.

The Supreme Court of New Zealand may well be alert for an opportunity to revisit the approach to judicial comment set out in R v Keremete CA247/03, 23 October 2003, applied recently in B v R [2018] NZCA 80.

Keremete tolerates a contradiction that is dispensed with by the general rule in McKell. This is that the judge may express in strong terms a view on the facts, while at the same time telling the jury that it is the sole arbiter of the facts. Keremete accepts a position in which strong comment can be consistent with fair presentation of the issues while at the same time leaving the issues of fact to the jury. Muddled, to say the least.

If jury trial judges disobey McKell, appellate courts will have to grapple with summings-up like that considered in B, where the court had to accept that there was prima facie lack of balance, but decided that in context the combined significance of the errors was considerably reduced, to the point where the court was able to conclude there was no unfairness.

This cleansing-by-context process is unsatisfactory, not the least because it is mysterious. B was not a case where comment was aimed at correcting an error or restoring an imbalance. It can only have been speculation for the court to conclude that there was no real risk that jurors had been influenced by the improper judicial comments. Trial outcome and trial process are separate matters, and it would be wrong to think that, because a verdict seems to have been correct, it was arrived at fairly. An analysis analogous (and here I stretch a bit for an analogy) to the requirements for the defence of withdrawal may be appropriate in determining whether a judge has restored the balance after making an inappropriate comment: compare Ahsin v R [2014] NZSC 153 at [140], applied in De Soto v R [2018] NZCA 366. Outside of strict criteria of that sort, breach of the McKell rule should, of itself, amount to a miscarriage of justice.

In what one hopes was a flash of brilliant wit, Gageler J declined to comment on judicial comments.