What Do Laboratory False Confession Paradigms Measure?

What happens in these experiments

For three decades, researchers have studied false confessions by accusing innocent people of minor transgressions in the laboratory. Four paradigms dominate the literature. In each, the participant truly did not do the thing — the question is whether they will sign a statement saying they did.

The headline number — and its enormous spread

Pooling all 99 conditions, the meta-analytic false confession rate is 41%. But the more revealing number is the spread. A new experiment drawn from this literature could credibly produce a rate anywhere from 3% to 93% — near zero to near universal — depending on task details.

Each dot is one experimental condition (dot size = number of participants; hover for details). The horizontal line marks the pooled 41% estimate; the shaded band is the 95% prediction interval.

If these tasks measured a stable psychological construct, different implementations should produce reasonably consistent results. Instead, rates scatter across nearly the entire 0–100% range. That pattern looks less like a stable trait being measured and more like an outcome that is highly sensitive to how the task is built.

37% confess before interrogation even starts

In 30 control conditions, the experimenter simply stated the accusation — no false evidence, no minimization, no pressure tactics of any kind. 37% of innocent participants signed anyway. The most coercive tactic in the literature, the false evidence ploy, averages 62% — which means the no-tactic baseline already accounts for 60% of it.

Unweighted mean false confession rate by interrogation method, ordered from least to most coercive (computed directly from the 99 conditions; k = number of conditions). Dashed line: control baseline.

Tactics do matter at the margin — rates rise about 0.23 logit units per step up the coerciveness scale (Spearman ρ = .33). But in a formal dose-response model, coerciveness explains less than 1% of the variation across conditions. Most of the measured behavior is already there when an authority figure merely states an accusation in a low-stakes setting. That is the signature of compliance, not of interrogation-induced false confession.

Small details move the needle more than tactics do

If tactics explain almost none of the variation, what does? Task structure and context. Explore the three views below: the same “false confession rate” swings dramatically with the paradigm used, minor procedural details like which key the participant is accused of pressing, and the country where the study was run.

Two ways to read this literature

The validity question can be framed as a contest between two interpretations, each making observable predictions. The interrogation mechanism account says lab paradigms tap the processes that produce real false confessions. The compliance account says they mostly measure willingness to go along with an authority figure’s accusation when nothing is at stake.

These accounts are not mutually exclusive — laboratory tasks may capture mechanisms genuinely relevant to interrogation while still being poorly calibrated to the real-world phenomenon. The question is which account better describes the absolute rates and patterns of variation the literature actually produces.

Could these rates come from the real world?

Here is the paper’s central exercise. Suppose the true real-world false confession rate were some value — say 10%. If we then ran this exact literature (99 conditions, the same sample sizes) on that world, how often would it produce the observed pooled mean of 41%? Drag the slider.

Probability that a synthetic literature (99 conditions, the actual sample sizes, binomial sampling) yields a mean at or above the observed 0.41, by hypothesized true rate — the normal approximation to the paper’s 10,000-draw simulation. Amber band: plausible real-world range (5–15%).

At true rates of 5%, 10%, or 15%, the probability is essentially zero — not small, zero to hundreds of decimal places. The most forgiving version of this test (allowing the corpus’s own heterogeneity into the hypothetical real world) lowers the edge of compatibility from about 40% to about 32% overall, and to 23% for control conditions — still double the field’s upper-bound survey estimates. The laboratory numbers are only compatible with a world in which false confession is routine.

If lab rates were real-world rates

Another way to see the calibration gap: take the laboratory rate literally and see what it implies. The FBI reports roughly 7.5 million arrests per year in the United States. Choose what share you think involves custodial interrogation of an innocent person.

To be clear about the logic: this is a reductio, not an accusation. We do not claim researchers believe the field produces hundreds of thousands of false confessions a year. The point is the converse — because that implication is not reasonable, a laboratory rate that produces it cannot be measuring the same thing as real-world false confession. Even the control conditions alone (37%, no tactics) imply roughly 278,000 per year.

Even generous assumptions leave an enormous gap

Effect sizes are the standard language for how big a difference is. Express the gap between the pooled laboratory rate (41%) and conservative real-world rates (5–15%) in those terms, and it dwarfs the conventional benchmarks for a “large” effect — at every baseline.

Odds ratio comparing the pooled laboratory rate (41%) with each assumed real-world rate. Dashed line: the conventional “large effect” benchmark (OR = 2). Hover a bar for the other metrics.

Large-effect benchmarks: RR = 1.5, OR = 2.0, h = 0.5, d = 0.8, r = .30. Every value exceeds its benchmark except Cohen’s d at the 15% baseline (0.75, just below the 0.8 cutoff).

At a 5% baseline, the gap corresponds to an odds ratio of 13.1 and a Cohen’s d of 1.42. Even granting 15% — above the field’s own upper-bound survey estimates — nearly every metric still clears its large-effect threshold, and the one exception (d = 0.75) misses by a hair. Gaps of this size are rare for complex social behavior: associations this strong usually involve direct mechanistic relationships, and behavioral effects routinely shrink once confounding and measurement error are addressed. These numbers are best read as the size of the calibration gap itself — not as the size of any real-world interrogation effect.

High baselines are a feature — and the problem

Why would paradigms evolve toward tasks where a third of innocent people comply immediately? Statistical power. Detecting tactic effects at realistic base rates requires enormous samples; at laboratory base rates, small samples suffice.

Participants needed per condition for 80% power (two-sided α = .05) to detect the labeled increase, versus the average condition actually run in this literature (n = 37).

A four-arm study at realistic base rates would need nearly 5,900 participants; the average condition in this corpus enrolled 37. A paradigm that produces a 30–40% baseline makes tactic comparisons feasible at ordinary lab sizes — that is precisely why these tasks are useful experimental tools. But the same feature is the validity problem: a behavior common enough to appear in over a third of innocent control participants is unlikely to be the same quantity as the rare, high-cost event of false confession in a criminal case.

What this means in practice