False accusations as an epistemic risk
The rate of false accusations is a latent quantity. Most criminal complaints cannot be compared with an independently known factual record, and the legal outcome is not a truth label. A conviction cannot be counted as a verified true accusation; an acquittal or discontinued investigation cannot be counted as a verified false accusation.
This makes common prevalence arguments circular. Studies of police classifications estimate the proportion of reports that investigators eventually classify as demonstrably false. They do not estimate every false report that remained unrecognized, every sincere but mistaken accusation, or every true report that was wrongly rejected. An exoneration registry similarly observes only the rare errors that the correction system succeeded in recognizing.
A corroborating witness does not create ground truth either. Witness reports are not ground truth distinguishes independence between people from independence between error processes. Witnesses can sincerely reproduce the same mistake or cease to be independent through discussion, shared media, repeated interviewing, and investigator feedback.
Several mechanisms are hidden by one label
A critical analysis should distinguish:
- deliberate fabrication
- sincere memory error or source confusion
- suggestion or contamination during interviewing or treatment
- mistaken identification
- an honestly described encounter with disputed legal meaning
- an allegation whose time, place, or mechanism changes enough to defeat counterproof
- a true core allegation accompanied by inaccurate peripheral details
- an investigator or prosecutor converting ambiguous material into a more definite narrative than the source supports
These mechanisms create different evidence patterns and require different safeguards. Treating them all as either “victims lie” or “false accusations are rare” prevents serious analysis.
Risk without a prevalence estimate
Uncertain frequency does not imply negligible risk. The risk can be evaluated through exposure, failure mechanisms, detectability, and severity:
| Question | Critical measure |
|---|---|
| How can the error enter? | Suggestion, contamination, identification, inference, or fabrication |
| What amplifies it? | Tunnel vision, vague charges, weak disclosure, deference, or political salience |
| Can the defence test it? | Preserved interviews, exact dates, complete records, and funded expertise |
| Can later review detect it? | Independent investigation, access to unused material, and a realistic reopening threshold |
| What is irreversible? | Detention, imprisonment, expulsion, family loss, stigma, and exposed private data |
Felaktigt dömda is unusually valuable here because it starts with independently established correction outcomes and reconstructs the mechanisms in the full files. It still cannot tell us the frequency of uncorrected cases. Moa Lidén’s Confirmation bias in criminal cases goes further: its system-level estimates imply that many wrongfully convicted people who appeal or seek a new trial will not be acquitted under a wide range of assumptions.
The appropriate conclusion is not a guessed number. It is that claims about a low false-accusation or wrongful-conviction rate are not empirically secured when the institution generating the labels is the institution whose reliability is in question.
The popular 98-percent gloss on the criminal proof threshold does not solve this denominator problem. As Probabilistic interpretations of beyond reasonable doubt explains, it would imply an expected innocent person in fifty only if the case assessments were real and calibrated probabilities. Repeating the number supplies neither calibration nor ground truth.