Entrust Identity Verification Services AI Transparency Notice

Use & Processing · Context

AI Transparency Notice scope

Entrust presents this Notice as transparent information about the Identity Verification AI system it provides, covering the system’s capabilities, limitations, biases, data usage practices and safety measures so readers can make informed decisions about interacting with the technology. The notice describes the AI generally; it does not itself grant rights or promise particular outcomes.

Supporting quotations

AI Transparency Notice

This Notice provides transparent information about the AI system you are using, outlining the system’s capabilities, limitations, biases, data usage practices, and safety measures.

Qualifications & exceptions

• Descriptive transparency notice, not a contractual commitment or a guarantee of AI performance.

Not stated here

• The supplied text does not state separate regional versions or a different effective version.

Use & Processing · Company practice

IDV inputs and accept/consider signal

Entrust’s IDV product lets clients remotely process, verify and onboard users. Its inputs are a photo or video of an identity document and a photo or video of the applicant. The decision engine returns an accept/consider signal with context and a detailed breakdown of how the result was produced.

Supporting quotations

Entrust’s IDV (Identity Verification) product enables clients to remotely process, verify and onboard users.

The inputs to this system are a photo/video of an identity document and a photo/video of the applicant.

Entrust’s decision engine uses these inputs to return an accept/consider signal, together with appropriate context and detailed breakdowns on how that result was produced.

Qualifications & exceptions

• The notice describes a product used by client organizations; it does not state their decision rules or outcomes.

Not stated here

• Accept/consider thresholds and the review path for a “consider” result are not stated in this section.

Use & Processing · Company practice

Four AI-driven IDV technology segments

Entrust says the IDV product is driven by four core technology segments that each rely on AI systems: facial similarity, biometric anti-spoofing and liveness detection on facial images, automated data extraction from documents, and anti-fraud technology on identity documents. Together these perform data extraction, biometric matching, document validation and biometric validation.

Supporting quotations

The IDV product is driven by four core technology segments, each of which relies on AI systems.

The four segments are (1) facial similarity; (2) biometric anti-spoofing and liveness detection (facial images); (3) automated data extraction from documents; and (4) anti-fraud technology on identity documents.

Qualifications & exceptions

• The notice does not map each output to a single model or disclose model versions.

Cross-references

• qualified by (model-proposed): IDV combines AI systems with non-AI algorithms — The four AI-driven segments operate within an application that also contains a substantial amount of non-AI algorithms.

Use & Processing · Context

IDV combines AI systems with non-AI algorithms

Entrust describes the IDV application as a combination of multiple AI systems and a substantial amount of non-AI algorithms, described as traditional computer code. The notice does not attribute individual outputs to AI versus conventional logic.

Supporting quotations

the IDV application contains a substantial amount of non-AI based algorithms as well

Qualifications & exceptions

• This passage does not quantify the AI/non-AI split or identify which decisions rely on conventional code.

Not stated here

• The contribution of non-AI code to the accept/consider signal is not disclosed.

Cross-references

• qualifies (model-proposed): Four AI-driven IDV technology segments — The four AI-driven segments operate within an application that also contains a substantial amount of non-AI algorithms.

Use & Processing · Company practice

Design aims to minimize human review of personal data

Entrust says responsible use of AI enables faster, more consistent data processing and a smoother, more secure end-user experience. It says the systems are designed to minimize human review of personal data wherever possible, helping reduce unnecessary exposure of end-user information.

Supporting quotations

Entrust’s responsible use of AI enables faster and more consistent data processing, resulting in a smoother and more secure experience for end users.

Our systems are designed to minimize human review of personal data wherever possible, helping to reduce unnecessary exposure of end user information.

Qualifications & exceptions

• Stated as a design aim (“wherever possible”); it does not promise that personal data is never reviewed by people.
• The notice separately describes Human-in-the-Loop review of model outputs, analyzed later in this document.

Not stated here

• The notice does not quantify how often human review occurs or which data is reviewed.

Cross-references

• qualified by (model-proposed): Human-in-the-loop review of model outputs — Human-in-the-loop review of model outputs is a stated human-oversight practice that qualifies the earlier design aim to minimize human review of personal data.

Use & Processing · Company practice

AI extracts information from the identity document

The AI extracts relevant information from the document image or video, such as name or date of birth.

Supporting quotations

Extract relevant information from the document image or video, such as name or date of birth

Not stated here

• The full set of extracted fields is not listed.

Use & Processing · Company practice

Deep-learning models, pre-trained and fine-tuned

Entrust says all applications use deep-learning models, with specific paradigms depending on the application. The models are typically pre-trained and then fine-tuned over Entrust’s own datasets to optimize accuracy and fairness.

Supporting quotations

All applications use deep learning models, with specific paradigms depending on the application.

These models are typically pre-trained then fine-tuned over our own datasets to optimize for performance (both accuracy and fairness).

Qualifications & exceptions

• The notice does not identify the datasets or provide performance metrics.
• Training and benchmarking practice, including human review of outputs, is described later in the notice and analyzed separately.

Not stated here

• Model architectures, versions and training-data sources are not disclosed.

Use & Processing · Company practice

Vision-language model for document extraction

Document extraction uses a vision-language model (VLM) for document intelligence.

Supporting quotations

Document extraction uses a vision-language model (VLM) for document intelligence

Qualifications & exceptions

• The notice does not name the VLM or describe its outputs beyond extraction.

Use & Processing · Company practice

Training and benchmarking on end-user datasets

Entrust says its AI models are optimized for accuracy and fairness through a systematic benchmarking protocol. The models are trained and fine-tuned, then benchmarked against datasets collected from end-user data.

Supporting quotations

Our AI models are optimized towards performance (both accuracy and fairness) using a systematic benchmarking protocol.

They are trained and fine-tuned, then thoroughly benchmarked against datasets collected from end user data.

Qualifications & exceptions

• The notice does not give dataset sizes, performance metrics or the specific training objective.

Not stated here

• Whether all models follow the same protocol is not stated.

Use & Processing · Company practice

Quality-control sampling and analyst feedback

Entrust uses Quality Check (QC) data uniformly sampled from production to estimate the true false acceptance and false rejection rates of a model. It also runs a Human in the Loop (HITL) feedback process in which human analysts evaluate rate outputs to improve model performance under its responsible AI principles.

Supporting quotations

We use Quality Check (QC) data uniformly sampled from production to obtain unbiased estimates of the true false acceptance/rejection rates of a given model.

Additionally, we employ a feedback process called Human in the Loop (HITL), whereby human analysts evaluate rate outputs to continuously improve model performance in accordance with Entrust’s responsible AI principles.

Qualifications & exceptions

• The QC sampling rate and the measured error rates are not disclosed.

Not stated here

• How analysts' evaluations translate into model changes is not described.

Use & Processing · Company commitment

Personal data processed only with a valid legal basis

Entrust says its use of AI involves processing end-user personal data, which it processes only where it has a valid legal basis. It directs readers to a separate IDV Product Privacy Notice for more detail on its use of personal data.

Supporting quotations

Our use of AI involves the processing of end user personal data.

We process such personal data only where we have a valid legal basis for doing so.

To learn more about our use of personal data, visit the IDV Product Privacy Notice.

Qualifications & exceptions

• The notice does not identify the legal bases or the controller/processor roles; the linked privacy notice is not part of the supplied source.

Not stated here

• Which legal basis applies to each processing purpose is not stated here.
• The linked IDV Product Privacy Notice’s text is not supplied here, so this card cannot describe its legal-basis details or data-handling terms.

Use & Processing · Company commitment

Bias-mitigation methods

Entrust says its AI systems are developed and continuously refined to minimize unintended bias and promote fair outcomes across demographic and geographic groups. It lists balanced and adaptive sampling during training, regular performance evaluations across countries, continents and genders, dynamic sampling that emphasizes underrepresented groups, and fine-tuning on real-world domain-specific data such as selfie-to-document comparisons.

Supporting quotations

Entrust’s AI systems are developed and continuously refined to minimize unintended bias and promote fair outcomes across demographic and geographic groups.

To mitigate bias, we apply:

Balanced and adaptive sampling techniques during model training to ensure diverse regional and demographic representation.

Regular performance evaluations across countries, continents, and genders to identify and address disparities.

Dynamic sampling adjustments that increase training emphasis on underrepresented groups to improve accuracy.

Fine tuning on real-world domain specific data (such as Selfie to document comparisons) to strengthen model generalization across different image types and conditions.

Qualifications & exceptions

• The methods are described without results, targets or independent validation.

Not stated here

• How the methods interact and their measured effect are not disclosed.

Cross-references

• qualified by (model-proposed): Measures reduce but do not eliminate bias — The mitigation methods are stated to reduce performance differentials but not to eliminate bias.

Use & Processing · Context

Reference to an ICO Sandbox Report

Entrust points readers to an ICO Sandbox Report for more information about its commitment to developing technology responsibly.

Supporting quotations

For more information on our commitment to developing innovative technology responsibly, you can refer to our ICO Sandbox Report.

Qualifications & exceptions

• The external report is not supplied and its findings are not summarized in this notice.

Not stated here

• The linked ICO Sandbox Report’s text is not supplied here, so this card cannot describe its findings or conclusions.

Use & Processing · Company practice

Human-in-the-loop review of model outputs

Entrust says it provides identity verification services using a Human in the Loop (HITL) model that is central to its AI operations. Through this process, trained analysts regularly review, validate and refine model outputs to improve performance and alignment with Entrust's responsible AI principles.

Supporting quotations

Entrust provides identity verification services for customers utilizing a Human in the Loop (HITL) model and is central to how our AI systems operate.

Through this process, trained analysts regularly review, validate, and refine model outputs to improve its performance and alignment in accordance with Entrust’s responsible AI principles.

Qualifications & exceptions

• The notice describes review of model outputs; it does not say every individual check is reviewed by a person.

Not stated here

• Review frequency, sample selection and whether review can change an individual decision are not stated.

Cross-references

• qualifies (model-proposed): Design aims to minimize human review of personal data — Human-in-the-loop review of model outputs is a stated human-oversight practice that qualifies the earlier design aim to minimize human review of personal data.

Use & Processing · Company practice

Pre-release testing and human oversight

Before models are released to production, Entrust says experts evaluate performance using analytic dashboards and controlled A/B testing to confirm results meet its quality and fairness standards. Human reviewers also oversee data labeling and verification so training data is accurate and representative. In certain products, human specialists remain available in real time to assess results when the AI detects unusual or uncertain inputs.

Supporting quotations

Before models are released to production, experts evaluate its performance using analytic dashboards and controlled A/B testing to confirm that results meet our standards for quality and fairness.

Human reviewers also oversee data labeling and verification, ensuring that the training data used to improve our systems is accurate and representative.

In certain products, human specialists remain available in real time to assess results when the AI detects unusual or uncertain inputs.

Qualifications & exceptions

• Real-time specialist availability applies only “in certain products”.

Not stated here

• Which products include real-time specialist review is not stated.

Use & Processing · Procedure

Feedback channel for AI inaccuracies

Entrust welcomes feedback to improve its AI systems and invites details of potential inaccuracies or user experience by email to AI@Entrust.com.

Supporting quotations

We welcome your feedback for the purpose of improving our AI systems.

To provide details on potential inaccuracies or to share your experience, email us at AI@Entrust.com.

Qualifications & exceptions

• No response time, remedy or outcome is promised.

Not stated here

• How feedback is used or retained is not stated.

Use & Processing · Company commitment

Customer visibility and control over data

Entrust says customers should have full visibility and control over their data, helping customers stay transparent with their end users about how their data contributes to improving Entrust’s services.

Supporting quotations

We believe customers should have full visibility and control over their data, helping them stay transparent with their end users about how their data contributes to improving our services.

Qualifications & exceptions

• Stated as a belief or aim (“should have”); the notice does not create an enforcement mechanism.

Not stated here

• How visibility and control are provided is not described here.

Use & Processing · User right or choice

Customer opt-out from machine-learning use

Entrust says its IDV systems use machine learning to enhance accuracy, detect anomalies and improve performance, but participation is entirely at the customer’s discretion: the IDV operates on an opt-out basis. When a client opts out, its checks are excluded from machine-learning activities, including model training. Customers control whether end-user data may be used to develop and refine the services and can change that setting at any time through account configuration.

Supporting quotations

Our identity verification (IDV) systems use machine learning to enhance accuracy, detect anomalies, and improve overall performance.

Participation in these processes is entirely at the customer’s discretion.

Our IDV operates on an opt-out basis.

When a client opts out, their checks are excluded from any machine learning activities, including model training.

Customers have full control over whether end user data may be used to help develop and refine our services and can customize at any time through account configuration.

Qualifications & exceptions

• The choice belongs to the client organization; the notice does not give individual end users a corresponding opt-out.

Not stated here

• How end users are informed of the choice, or whether they can object, is not stated.

Cross-references

• qualified by (model-proposed): Machine-learning opt-out also excludes checks from QC — Opting out of machine learning automatically carries exclusion from the QC process, linking the customer’s machine-learning choice to quality-control review.
• has exception (model-proposed): Pseudonymized and aggregated data after opt-out — Even when a customer opts out of machine learning and QC, Entrust may still process pseudonymized and aggregated data for the stated business purposes.

Use & Processing · Company practice

Machine-learning opt-out also excludes checks from QC

Entrust says a customer that opts out of machine learning is automatically excluded from its Quality Control (QC) process. In that process, Entrust samples production data and approximately three percent of all checks are randomly reviewed each month to validate accuracy and support improvements. Opting out removes the customer’s checks from this feedback cycle.

Supporting quotations

When a customer opts out of machine learning, they are also automatically excluded from our Quality Control (QC) process.

As part of our standard practice, approximately three (3%) percent of all checks are randomly reviewed each month to validate accuracy and support ongoing improvements.

These QC insights help strengthen the reliability of our models, but opting out removes the customers’ checks from this feedback cycle.

We use Quality Check (QC) data uniformly sampled from production to obtain unbiased estimates of the true false acceptance/rejection rates of a given model.

Qualifications & exceptions

• The QC sample is approximate and monthly; opting out removes the checks from the QC feedback cycle.

Not stated here

• Whether the three-percent sample is drawn from all checks before or after opt-outs is not stated.
• The validation criteria applied in QC are not described.

Cross-references

• qualifies (model-proposed): Customer opt-out from machine-learning use — Opting out of machine learning automatically carries exclusion from the QC process, linking the customer’s machine-learning choice to quality-control review.

Use & Processing · Company practice

Pseudonymized and aggregated data after opt-out

Even when customers choose not to participate in machine learning or QC, Entrust says its systems may process pseudonymized and aggregated data for legitimate business purposes such as analytics, benchmarking and statistical reporting. It says aggregated data does not include personally identifiable information (PII).

Supporting quotations

Even when customers choose not to participate in machine learning or QC, our systems may process pseudonymized and aggregated data for legitimate business purposes such as analytics, benchmarking, and statistical reporting.

Aggregated data does not include any personally identifiable information (PII).

Qualifications & exceptions

• “May” indicates a possibility, not a commitment to process; the listed purposes are examples (“such as”).

Not stated here

• “Pseudonymized” is not defined and the notice does not describe re-identification safeguards.
• What counts as legitimate business purposes beyond the examples is not stated.

Cross-references

• exception to (model-proposed): Customer opt-out from machine-learning use — Even when a customer opts out of machine learning and QC, Entrust may still process pseudonymized and aggregated data for the stated business purposes.

Use & Processing · Company commitment

Continuous bias monitoring and model updates

Entrust treats bias monitoring and fairness evaluation as a continuous process. It says its AI models are regularly reviewed and updated, and that when disparities are identified it adjusts training data, sampling strategy or calibration parameters to improve balance and accuracy.

Supporting quotations

We treat bias monitoring and fairness evaluation as a continuous process.

Our AI models are regularly reviewed and updated to enhance performance and address issues.

When disparities are identified, we adjust our training data, sampling strategy, or calibration parameters to improve balance and accuracy.

Qualifications & exceptions

• No review cadence, metrics or disparity thresholds are stated.

Not stated here

• How disparities are detected and verified is not described.