Where the Values Live
Before the 1950s, psychophysicists believed in something called the absolute threshold — a fixed point below which a stimulus simply cannot be detected. Too quiet, and you hear nothing. Too faint, and you see nothing. The threshold was understood as a feature of the sensory system, like the resolution of a camera: a hard limit, determined by biology, impervious to attitude or context or desire.
Signal detection theory dissolved this idea. Not by finding better thresholds, but by showing the concept was wrong from the start.
The framework came out of World War II radar research. The problem was this: a radar screen shows noise whether or not there's a plane. Background static looks roughly like a faint signal. An operator scanning the screen has to decide, on each sweep, whether what they're seeing is a target or just noise. The question isn't whether the signal is above or below some fixed level. It's whether the probability that a signal caused this reading exceeds some threshold of confidence — and that threshold depends on what you're optimizing for.
If you're worried about missing a real plane, you lower the bar. You call "yes" more readily. You'll catch more genuine targets, and you'll also fire more false alarms. If false alarms are expensive — if scrambling fighters for noise is costly — you raise the bar. You miss some real planes, but you stop wasting resources on phantom ones. Neither setting is correct in the abstract. The right threshold depends on the relative cost of the two types of error. That's not a technical fact. It's a value judgment.
Signal detection theory gave this a mathematical structure. A sensitivity measure — d-prime — captures how separable the signal+noise distribution is from the noise-alone distribution. High d' means the two look very different; low d' means they overlap a lot. This is a fixed property of the signal and the observer's sensory system. You can't increase d' by trying harder or caring more. You can't move the distributions with will.
The criterion — where you draw the line between "yes" and "no" — is separate. You can move it. But moving it doesn't change d'. It only changes where you sit on the tradeoff curve. Every detector, biological or mechanical, has a curve like this: plot your hit rate against your false alarm rate across every possible criterion position, and you get a bow-shaped arc. You can be anywhere on your curve. You can't be above it without improving d'. The curve is the ceiling that your sensitivity level allows.
What this means in practice: there is no criterion setting that avoids the tradeoff. You can push false alarms to zero, but then your hit rate goes to zero too. You can push your hit rate to one, but then your false alarm rate goes to one too. The tradeoff isn't a failure of engineering or effort. It's the structure of the problem.
This shows up in places that don't look like radar screens. A spam filter set aggressively catches most spam and also blocks legitimate email. Set conservatively, it lets spam through. The curve is the same: you're trading false positives for false negatives. The question "what's the right spam threshold" has two parts: a technical part (what's d'? — how separable is spam from legitimate mail?) and a values part (how much spam can you tolerate vs. how many legitimate emails can you afford to lose?). Only the first part can be answered without making a choice about what matters.
"Beyond a reasonable doubt" is a criterion setting. It says: place the bar high. The legal tradition that it's better to acquit ten guilty people than convict one innocent isn't a factual claim about certainty — it's a statement about which error is worse. Move the criterion, and you change the ratio. You can't escape having a ratio.
Medical screening is the same. Set the mammogram threshold low and you catch more real cancers, along with more false positives — more biopsies, more anxiety, more overtreatment. Set it high and you miss some real cancers but reduce unnecessary intervention. There is a correct threshold given a complete account of the harms. But specifying those harms requires a values framework that medicine doesn't contain by itself. The question of whether a false positive's cost (unnecessary biopsy, anxiety) outweighs a false negative's cost (delayed cancer diagnosis) is a moral question dressed in clinical language.
Here's what I find most interesting about the history. Before signal detection theory, the "absolute threshold" concept did something specific: it made detection look like a fact about the sensory system rather than a decision made by the observer. If you couldn't hear the tone, that was a fact about your ears — not a fact about where you were willing to commit to a "yes." The threshold concept hid the criterion. It made what was always already a value judgment look like a biological constant.
Signal detection theory didn't add a value judgment to a previously neutral system. It showed that the system always contained one. The criterion was always there, always embedded in detection, always trading one kind of error for another. The math just built the apparatus to see it clearly.
The absolute threshold was never absolute. It was a criterion in disguise.