Implementation guide
OWASP AISVS v1.01
Training Data Integrity & Traceability · Training Data Quality and Security Assurance
Check auto-generated labels
Verify that automatically generated labels are subject to confidence thresholds and consistency checks to detect misleading or low-confidence labels.
OWASP AISVS v1.01, requirement 1.3.2
In plain English
When labels come from a model or a script, keep the ones it is confident about and send the doubtful or inconsistent ones to a person. Automatic labels carry the errors of whatever produced them, and those errors compound when you train the next model on them.
What good looks like
- Each automatic label is stored with a confidence score.
- Labels below a set threshold are left out or sent for human review.
- A sample of automatic labels is regularly compared with human labels on the same items.
- The threshold was chosen from measured error rates and is revisited when the labelling model changes.
How teams usually do it
- Keep a "gold" set labelled carefully by people, and measure the automatic labeller against it on every change.
- Label each item with two different methods (two models, or a model and a rule) and review the disagreements.
- Weak-supervision frameworks such as Snorkel, which combine several noisy labelling sources and estimate how reliable each one is.
Suggested evidence
- Threshold settings and the measurements behind them.
- Recent agreement figures between automatic and human labels.
- The review queue, showing low-confidence items were looked at.
Common mistakes
- Treating model-generated labels as ground truth for the next model.
- Trusting raw confidence scores. Models are often overconfident, so calibrate the scores against real error rates first.
- Leaving the threshold alone after swapping the labelling model.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.