AI detection limitations

Purpose and when to use it

Use a detector only after defining the question and checking that its tested use conditions are relevant. Detection of a synthetic signal is different from identifying a model, proving deception, or authenticating a depicted event.

Evidence required

Obtain the detector version, score definition, decision threshold, validation population, labeled controls, input preprocessing, and error results. If these are unavailable, report that gap before treating the output as consequential.

Principles and procedure

• Preserve the input and record every preprocessing step, detector version, and configuration.
• Test relevant known positives, known negatives, and unknown-generator objects separately.
• Evaluate compression, resizing, screenshots, transcoding, post-processing, and adversarial alteration under authorized conditions.
• Report false positives and false negatives by condition; test drift and calibration rather than assuming a score’s probability meaning.
• Compare independent evidence and examine whether an ensemble shares data, architecture, or errors before counting its votes.

What to document

Publish the decision rule, denominators, exclusions, input quality, source labels, and evaluation date. Distinguish threshold accuracy from calibration and explain when the evaluation does not represent the case population.

Failure modes

Unknown generators, low-quality inputs, domain shifts, and updates can alter errors. Ordinary processing may remove signals; unrelated content may resemble them. Repeated correlated detectors can amplify the same mistake.

Limits on the conclusion

Base rates matter. In an illustrative population of 1,000 objects with 10 synthetic objects, 90% sensitivity and a 1% false-positive rate produce about 9 true alerts and 10 false alerts. Those invented teaching counts show why alert precision differs from sensitivity; they are not a measured product result.

Independent checks and corroboration

Pair detector findings with source-system records, provenance, and validated content examination. A negative result cannot establish non-synthetic origin; a positive result cannot by itself establish a creator or intent.

How to report the result

Name the signal observed, tested conditions, error information, and uncertainty. If evaluation evidence is insufficient, say the score’s forensic meaning remains unresolved.

Related methods

• The complete evaluation method
• Making bounded attribution claims
• Writing calibrated confidence statements

Related glossary terms

• false positive
• false negative

What does an alert mean in this population?

Inspect the target label, threshold, evaluated inputs, and known error conditions. Consider the prevalence of the positive class in the actual population. Ordinary processing and unknown generators may differ from the detector’s validation corpus.

Report a result as a method-dependent measurement alongside its conditions. A positive alert needs supporting evidence, and a negative result cannot establish that no AI generation or editing occurred.