Standards of proof

Why can the same disputed act count as proven in a civil court and unproven in a criminal one?

A man is tried for a killing and acquitted. The family sues him over the same death and wins. Nothing about the world changed between the two courtrooms; the same witnesses could give the same evidence.

The natural reaction is that one court got it wrong. Consider instead that both did exactly what they were built to do — and that the pair of verdicts is no contradiction, because they are not answers to the same question.

Ask what a verdict actually is. Not a report of what happened; nobody in the room knows that. It is a decision taken under uncertainty, on evidence leaving the fact-finder short of certain. Every such decision needs a rule for how short of certain is close enough. That rule is the standard of proof.

Now notice that the rule cannot be derived from the evidence. Evidence gives you a degree of confidence; it does not tell you what to do with it. To convert confidence into a decision you need one more input: what the two possible mistakes cost.

Work it through with the civil case first. A dispute over money between two private parties. If the court finds for the plaintiff wrongly, a defendant pays a sum he did not owe; if for the defendant wrongly, a plaintiff goes without a sum she was owed. Same kind of loss, borne by parties with equal standing. If the two errors cost about the same, the rule that minimises total error is simply: decide for whichever account is more likely. Which is what the civil standard says — balance of probabilities, or preponderance of the evidence. The familiar "more likely than not" is a derivation, not a slogan: it is where you set the threshold when you are indifferent between the two directions of mistake.

Now change one input. In a criminal case the state prosecutes an individual, and the errors are no longer alike. Convicting the innocent inflicts punishment, on a person, by the community, in its own name; acquitting the guilty leaves a wrong unpunished. Systems treating the first as much worse cannot use the midpoint — they must set the threshold high, so a case must be strong before conviction follows. Blackstone's formulation, that it is better ten guilty escape than one innocent suffer, is best read as an explicit price ratio, and other jurists have named other ratios. The particular number is a value judgement nobody derives; the structure — threshold set by the ratio of error costs — is what does the work.

So the two opening verdicts are consistent. The civil jury found it more likely than not; the criminal jury did not find it beyond reasonable doubt. Both can be true of a single body of evidence, and there is a whole band of confidence in which they must both be true.

A boundary worth marking. Whose job it is to move the fact-finder is the burden of proof — upstream of this. Why a system tolerates freeing the guilty at all is the presumption of innocence, which prices the errors. This question is the middle term: given errors priced unequally, how far must the evidence go, and what happens when the same facts reach two courts with different prices.

What would falsify this? If standards are thresholds set by error costs, intermediate standards should appear precisely where the stakes sit between the poles — and they do. Several systems apply a middle standard, clear and convincing evidence, to matters such as terminating parental rights: civil in form, with consequences too grave for the midpoint. The refuting observation would be a system that varied its standard with the difficulty of the evidence rather than the severity of the outcome — demanding more proof for hard-to-prove claims. That would mean the standard tracks epistemics rather than stakes, and the account here would be wrong.

That gap deserves the next paragraph, because the threshold picture is cleaner than the practice. Courts in several common-law systems have repeatedly refused to express "beyond reasonable doubt" as a percentage, and appellate courts have reversed convictions where trial judges tried to quantify it. The stated reasons are that a number invites jurors to do arithmetic on things that cannot be counted, and that it makes explicit a trade-off the system prefers not to state as a rate of wrongful convictions it will accept. Critics answer that the trade-off exists whether or not it is spoken, and that leaving it unspoken means twelve people are each using a private threshold. That dispute is unresolved. The honest position is that the standard is a threshold in theory and an instruction in practice, and the fit between them is loose.

A second refinement. It is tempting to think grave allegations need more proof even within a single standard. The United Kingdom's highest court has rejected this squarely: there is one civil standard, and it does not slide. What does vary is the evidence needed to reach it, because an inherently improbable event starts further from the threshold. The bar stays put; the distance to it depends on how unlikely the claim was to begin with.

And the price. A high criminal standard protects the innocent by design, and by the same design leaves conduct unaddressed — so pressure escapes into channels with lower thresholds. Regulatory proceedings, disciplinary bodies, civil suits, and licensing decisions can attach serious consequences on less than the criminal standard. Whether that is a sensible division of labour or an erosion of the protection is contested, and plausibly differs by field. What is not contested is the direction: raising a threshold in one forum does not remove the cases, it relocates them.

• Whether jury instructions could be calibrated empirically, by testing what verdicts different wordings produce.
• How thresholds set the same way outside courts — screening tests, safety certification, content moderation — fail in recognisably similar ways.

Every term the collection defines is gathered in the glossary.