Entrust Identity Verification Services AI Transparency Notice

Inference & Profiling · Procedure

Every validation step must pass for an accept signal

The notice says data extraction, biometric matching, document validation and biometric validation all require a positive result for the decision engine to return an accept signal. Under the described flow, a non-positive result in any of those tasks prevents an accept signal.

Supporting quotations

all requiring a positive result for an accept signal to be returned

Qualifications & exceptions

• This describes the composition of the accept signal; thresholds, error rates and any manual override are not stated.

Not stated here

• Whether a client can override a non-accept result is not addressed in this notice.

Cross-references

• qualified by (model-proposed): Results are risk indicators, not definitive determinations — The statement that every validation task must pass is qualified by the direction that results are risk indicators, not definitive determinations.

Inference & Profiling · Company practice

AI flags document manipulation or fraud

One listed AI use is to indicate whether a photo or video of a government ID appears genuine or may show signs of manipulation and/or fraud.

Supporting quotations

Indicate whether a photo or video of a government ID appears genuine or may show signs of manipulation and/or fraud.

Qualifications & exceptions

• Framed as an indicator (“may show signs”), not a definitive finding.

Inference & Profiling · Company practice

AI classifies the identity document type

The AI classifies the document type, for example a passport or driver’s license.

Supporting quotations

Classify the document type, for instance a passport or driver’s license.

Not stated here

• The notice does not list every supported document type.

Inference & Profiling · Restriction or exception

Results are risk indicators, not definitive determinations

Entrust says its solutions identify patterns that may indicate authenticity or potential fraud to help organizations make trust and safety decisions. It directs that results be interpreted as risk indicators rather than definitive determinations and used together with other controls and verification procedures. This qualifies the accept/consider signal described earlier.

Supporting quotations

Our solutions are designed to identify patterns that may indicate authenticity or potential fraud, helping organizations make informed trust and safety decisions.

Results should be interpreted as risk indicators rather than definitive determinations and should be used in conjunction with other controls and verification procedures.

Qualifications & exceptions

• Applies to client organizations’ use of the results and limits how much weight the output should carry.

Cross-references

• qualifies (model-proposed): Every validation step must pass for an accept signal — The statement that every validation task must pass is qualified by the direction that results are risk indicators, not definitive determinations.

Inference & Profiling · Company practice

Binary classifiers for fraud and liveness detection

Document fraud detection and biometric liveness detection use binary classification models to assess the probability that an input image or video is genuine, as opposed to being altered, deepfaked or deliberately obstructed.

Supporting quotations

Document fraud detection and biometrics liveness detection use binary classification models to assess the probability of the input image/video if being genuine (as opposed to being altered in any fraudulent manner, e.g. being a deepfake, or tampered with some kind of deliberate obstruction).

Qualifications & exceptions

• Describes probability assessment; thresholds are not stated.
• The source wording reads “probability of the input image/video if being genuine”; the card paraphrases the evident meaning, and the exact wording is preserved in the anchor.

Not stated here

• False-accept and false-reject rates are not given in this section.

Inference & Profiling · Company practice

Multi-class model classifies document type

Document classification uses a multi-class classification model that returns the most probable document type for an input document image.

Supporting quotations

Document classification uses a multi-class classification model to return the most probable document type given an input document image.

Qualifications & exceptions

• “Most probable” indicates a ranked prediction, not certainty.

Inference & Profiling · Company practice

Models may inherit bias from real-world training data

Entrust says that, like most AI systems trained on real-world data, its models may reflect inherited biases from the datasets used to train and improve them. Training-data diversity varies across countries and demographic groups, which may lead to performance differences for some populations or document types. Entrust says it monitors these factors and applies data rebalancing, retraining and fairness testing to reduce unintended differences.

Supporting quotations

Like most AI systems trained on real-world data, our models may reflect inherited biases from the datasets used to train and improve it.

The diversity of training data varies across countries and demographic groups, which may lead to differences in performance for some populations or document types.

We actively monitor these factors and apply measures such as data rebalancing, retraining, and fairness testing to reduce unintended differences in performance across user groups.

Qualifications & exceptions

• Framed as possibility (“may reflect”, “may lead”), not as a quantified finding.

Not stated here

• The magnitude or direction of any performance difference is not disclosed.

Inference & Profiling · Context

Examples of inherited-bias sources

The notice lists potential sources of inherited bias: training-data composition, where public datasets used in early development may overrepresent certain regions, facial characteristics or imaging conditions; image quality and document design, including photo quality, printing standards, holograms and security features; and capture conditions such as lighting, camera devices and how images are collected.

Supporting quotations

Training Data Composition: Publicly available datasets used in early model development may overrepresent certain regions, facial characteristics, or imaging conditions.

Image Quality & Document Design: Variations in photo quality, printing standards, holograms, or security features may affect performance across document types.

Capture Conditions: Differences in lighting, camera devices, or how images are collected can influence model outcomes across environments.

Qualifications & exceptions

• The list is introduced as examples (“include”), not as an exhaustive or ranked account.

Not stated here

• The notice does not quantify each source's effect on performance.

Inference & Profiling · Company commitment

Measures reduce but do not eliminate bias

Entrust says these measures significantly reduce performance differentials, but states that no AI model is entirely free from bias. It remains committed to transparency, continuous evaluation and ongoing improvement as part of responsible AI development.

Supporting quotations

These measures significantly reduce performance differentials; however, no AI model is entirely free from bias.

We remain committed to transparency, continuous evaluation, and ongoing improvements as part of our responsible AI development process.

Qualifications & exceptions

• The notice acknowledges residual bias but does not quantify it.

Not stated here

• Any remaining disparity by group or document type is not disclosed.

Cross-references

• qualifies (model-proposed): Bias-mitigation methods — The mitigation methods are stated to reduce performance differentials but not to eliminate bias.

Inference & Profiling · Restriction or exception

Stated scope, language and context limitations

The notice states that the AI is specifically designed for fraud prevention, supports multilingual inputs based on languages that use Latin and Arabic scripts, and evaluates inputs as either fraudulent or genuine.

Supporting quotations

Scope: Our AI is specifically designed for fraud prevention

Language: Our AI supports multilingual inputs based on languages that use Latin and Arabic scripts

Context: Our AI evaluates inputs as either fraudulent or genuine

Qualifications & exceptions

• These are the notice's own stated scope limits; other scripts, uses or result categories are not addressed here.

Not stated here

• The notice does not list every supported language or excluded use case.

Inference & Profiling · Restriction or exception

Inconsistent results in limited cases

The notice says that in limited cases the AI may generate inconsistent results if information is inaccurately extracted from a source document.

Supporting quotations

Hallucinations: In limited cases, our AI may generate inconsistent results if information is inaccurately extracted from a source document.

Qualifications & exceptions

• “Limited cases” is not quantified.

Not stated here

• What safeguards detect or correct inconsistent results is not stated.