HOW INTELLIGENCE WORKS · SIGNAL DETECTION REASONING · eduKateSG
How a Mind Separates Sensitivity to a Signal From Its Willingness to Say the Signal Is Present
Signal detection reasoning separates two jobs that raw accuracy can hide: how well a system distinguishes signal from noise, and where it places the decision criterion for saying the signal is present. A cautious observer may miss real signals; a liberal observer may detect more signals while also raising false alarms.
Signal/noise distributions → sensory or evidential sensitivity → decision criterion → hit / miss / false alarm / correct rejection → adjust criterion to stakes without confusing criterion with sensitivity.
This article belongs to the How Intelligence Works series. Discrimination owns learning which differences matter. The Attention Gate owns what enters active processing. Evidence Accumulation owns combining signals over time. Signal detection reasoning owns the decision geometry that separates sensitivity from criterion.
Accuracy Alone Cannot Tell You Why Decisions Changed
Two observers can have the same percent correct while using different criteria. One may produce many hits and many false alarms. Another may produce fewer false alarms but also miss more real signals.
The distinction matters whenever the costs of misses and false alarms differ.
Sensitivity describes separation. Criterion describes willingness to call the signal present.
1. Every Detection Problem Contains Signal and Noise
A detector receives evidence generated under at least two possible states: signal absent and signal present. Because the evidence distributions overlap, perfect classification may be impossible.
Detection therefore requires a decision under uncertainty, not merely perception.
2. Hits and False Alarms Must Be Read Together
| Reality | Say signal present | Say signal absent |
|---|---|---|
| Signal present | Hit | Miss |
| Signal absent | False alarm | Correct rejection |
A higher hit rate can come from better sensitivity or simply from lowering the criterion. False alarms reveal which interpretation is plausible.
3. Sensitivity Is Different From Response Bias
Sensitivity asks how far apart the signal and noise evidence distributions are. Response bias asks where the observer places the threshold for acting as though the signal is present.
Changing incentives can move the criterion without changing sensory or evidential sensitivity.
4. Costs and Base Rates Move the Rational Criterion
If missing a real threat is catastrophic, a lower criterion may be rational even though it produces more false alarms. If false alarms are costly, the criterion may move upward.
The correct criterion depends on prior probability and the relative consequences of each error type.
5. Signal Detection in Mathematics and Statistics
Signal detection theory formalises overlapping evidence distributions, sensitivity measures and decision criteria. Receiver operating characteristic curves show how hit and false-alarm rates trade off as the criterion moves.
The curve helps separate detector quality from one chosen operating point.
6. Signal Detection in Learning
A student deciding whether a grammar rule applies is performing a detection task. A teacher deciding whether a hesitation signals misunderstanding is also performing one.
Good teaching distinguishes low sensitivity—poor discrimination between cases—from a criterion problem such as chronic over-caution or over-guessing.
7. Repeated Exposure Can Change Sensitivity and Criterion Differently
Practice may genuinely improve separation between signal and noise. Feedback may instead teach the learner to shift the decision threshold.
Both can improve task performance, but they are different learning mechanisms.
8. Vigilance Problems Can Masquerade as Sensitivity Problems
Over long tasks, attention can fluctuate. A detector may become more conservative, less sensitive, or both.
Time-on-task analysis is useful because one overall accuracy figure can hide changing internal states.
9. Signal-Detection Failure Atlas
| Failure | What happens | Repair |
|---|---|---|
| Accuracy-only reading | Sensitivity and criterion are collapsed into one score | Inspect hits and false alarms |
| Hit-rate worship | More detections are assumed to mean better detection | Check false alarms |
| False-alarm panic | A cautious threshold is demanded regardless of miss cost | Make error costs explicit |
| Base-rate neglect | Criterion ignores how common the signal is | Use relevant priors |
| Threshold moralisation | Liberal or conservative responding is treated as character | Separate strategy from sensitivity |
| Stationarity assumption | One criterion is assumed across fatigue and context | Track criterion over time |
| Feedback confusion | Performance gains are attributed to sensitivity when criterion shifted | Analyse both components |
10. Signal Detection and Evidence Accumulation Are Different
Evidence accumulation asks how support builds over time toward a decision. Signal detection reasoning asks how separable the states are and where the response criterion sits.
Accumulation can feed a detection decision, but the concepts answer different questions.
11. Teams Should Name Which Error They Fear
Security, quality control, medicine, moderation and education all contain asymmetric error costs.
A team cannot choose a sensible threshold until it says whether a miss or a false alarm is more costly—and by how much.
12. Institutions Need Operating Points, Not One Universal Threshold
Different contexts can justify different criteria even with the same detector. Emergency screening may run at a liberal threshold; confirmatory review may run at a stricter one.
Layered systems can therefore combine high sensitivity early with high specificity later.
13. Artificial Intelligence and Detection Criteria
AI classifiers often output scores that are converted into actions by a threshold. The model’s discrimination ability and the deployment threshold should be evaluated separately.
A stronger system adjusts thresholds to base rates, consequence and human review capacity without claiming that a threshold change improved the underlying model.
14. The Signal Detection Audit
- States: What counts as signal present and signal absent?
- Evidence: What observation is used to distinguish them?
- Overlap: How much do the evidence distributions overlap?
- Hits: How often are real signals detected?
- False alarms: How often is noise called signal?
- Miss cost: What happens when a real signal is missed?
- False-alarm cost: What happens when noise triggers action?
- Base rate: How common is the signal in this environment?
- Criterion: Where should the action threshold sit?
- Learning: Did performance improve through sensitivity, criterion or both?
15. CivDJ Reading: Better Ears and a Hotter Trigger Are Not the Same
In the CivDJ frame, one operator may truly separate wanted signal from noise better. Another may simply open the gate more readily.
Do not praise the gate for hearing better when it merely opens more often.
16. Return to Sensitivity and Criterion
Signal detection reasoning prevents intelligence from confusing a decision policy with perceptual or evidential ability.
It keeps hits, misses, false alarms and correct rejections inside one geometry.
The mature mind asks both: how well can I tell the states apart, and where should I place the line for acting?