Age Estimation API: Selfie Age Verification

Didit Age Estimation predicts a person’s age from a selfie, with passive liveness included, so you can gate age-restricted services without asking every user for an ID document. The model is most precise where age gating decides: its mean absolute error is 1.1 years for 16 and 17-year-olds and 1.9 years for 18 to 20-year-olds, and it correctly classifies 96% of selfies as under 18 or 18 and over (see Model performance).

Age Estimation Methods

Our platform implements age estimation in conjunction with different liveness verification technologies:

Method	Description	Security Level	Best For
3D Action & Flash	• Combines multi-factor biometric verification with a randomized action sequence and dynamic light pattern analysis. • At the start, the user is prompted to perform a simple action—like blinking or nodding—ensuring real-time interaction. • Simultaneously, the system projects a sequence of light patterns onto the face, analyzing the reflections to confirm the face’s three-dimensional structure. • Deep learning algorithms examine micro-expressions and the light reflection responses to verify the presence of a live person. • Offers the highest security by integrating behavioral (action) and physical (light-based depth) cues, making it nearly impossible to spoof with static images, videos, or even advanced masks.	Highest	Banking, healthcare, government applications
3D Flash	• Uses dynamic light pattern analysis to validate facial topology without requiring user interaction. • Projects a series of light patterns onto the face at over 30 frames per second, analyzing the reflections to create a depth map. • This depth map confirms the face’s three-dimensional structure, distinguishing it from flat images or 2D spoofs. • Provides a seamless experience while maintaining high security against presentation attacks like photos or screens.	High	Financial services, account access, identity verification
Passive Liveness	• Relies on single-frame deep learning analysis to detect signs of liveness. • For privacy, the user’s face appears blurry in the interface, assuring them that their image is being analyzed for age estimation only, not for identification. • Examines the image for artifacts, texture patterns, and other subtle indicators that differentiate a real face from a spoof. • A convolutional neural network (CNN) validates facial features and identifies anomalies, such as those from printed photos or digital screens. • Offers fast and convenient verification but provides standard security, suitable for low-risk use cases.	Standard	Low-friction scenarios, consumer applications

Each method generates a precise age estimate along with confidence scores and supplementary gender estimation data.

Configurable Thresholds

You can customize security levels by setting different thresholds for age estimation. For example:

Per-Country Age Restrictions

Age restrictions can be configured on a per-country basis, reflecting the fact that the legal age of majority varies across jurisdictions. Instead of setting a single global minimum or maximum age, you can define specific age limits for each country — and even for individual states or regions within a country.

How It Works

• Country-level configuration: Set a minimum and/or maximum age for each country (identified by the document’s issuing state). For example, you might require a minimum age of 18 in the United States, 19 in South Korea, and 21 in the United Arab Emirates.
• State/Region overrides: For countries with sub-national variation (such as the United States or Mexico), you can configure overrides per state. For example, Mississippi may require a minimum age of 21 while the US default is 18. The system matches the state using the region extracted from the document by OCR.
• Age of majority defaults: The console provides a one-click “Apply age of majority” button that auto-populates each country’s minimum age based on the known legal age of majority worldwide. You can then customize individual countries or states as needed.
• Configurable actions: When a user’s age falls below the minimum or exceeds the maximum for their document’s country, you can choose the action to take: Decline or Review.

Example Configuration

How It Works

Adaptive Age Estimation

For scenarios where precise age verification is critical, our platform offers adaptive age estimation with ID verification fallback. This approach provides a balance between user convenience and regulatory compliance.

How Adaptive Age Estimation Works

What the fallback returns

The fallback face match runs without a Face Match node in the workflow, so it shows up in the session decision in a way that is easy to mistake for a check you never configured:

• features lists FACE_MATCH in addition to AGE_ESTIMATION and ID_VERIFICATION once the fallback has run.
• face_matches[] carries one entry with the score, the source_image (document portrait) and the target_image (age estimation selfie). Its node_id is the age estimation node (feature_age_estimation on a native workflow) because the match belongs to that step, not to a Face Match step.
• A low score raises the usual face match warnings and can send the session to review or decline, exactly like a configured Face Match node would.
• It is not billed as a Face Match: the session is charged for the age estimation and for the ID verification only.

Sessions that pass or fail on the selfie alone never run the fallback, so they return face_matches: null. To keep the images out of the decision payload, limit the face_matches fields with the workflow’s response attributes.

Benefits of Adaptive Age Estimation

• Reduced Friction: Most users complete verification with just a selfie
• Enhanced Compliance: Uncertain cases receive thorough document verification with country-specific age rules
• Cost Efficiency: ID verification is only used when necessary
• Customizable Risk Tolerance: Adjust thresholds based on your regulatory requirements
• Jurisdiction-Aware: Automatically applies the correct age of majority based on the user’s document issuing country and region

This approach is particularly valuable for age-gated services like online gaming, alcohol delivery, and adult content platforms where balancing user experience with regulatory compliance is essential.

Age assurance (ISO/IEC 27566-1)

A workflow can declare that it follows the ISO/IEC 27566-1 age assurance standard. Create it from the console templates Age Estimation (ISO/IEC 27566-1) or Age Verification (ISO/IEC 27566-1). The declaration is made when the workflow is created and cannot be changed afterwards. While a workflow declares the standard, Didit pins it to a one-shot yes or no answer about age. The pinned settings apply to every save, in the console and through the API:

• You receive the age decision and nothing else. The response attributes are locked, so the decision carries no personal data.
• 13 is the lowest minimum age in every country, in every regional override and in the adaptive age estimation thresholds. A lower value is rejected.
• No recognition across sessions. Duplicate document and duplicate face checks are off, and no biometric template is stored.
• Limited capture. Documents are captured with the camera only, with no file upload, no silent selfie capture while the document is photographed and no audio recording.
• No extra data checks. The expected details comparison and the data review step are off, and ID verification reads documents only, with no database lookups or digital ID wallets.
• A threshold is required. A workflow without an age threshold is rejected on save, so it can never approve everyone by default.

Model Performance and Statistics

The figures below describe the age model in production since 4 November 2026. They come from its locked evaluation set: 2,399 selfies of people aged 13 and over, set aside before training, with each person’s age taken from their verified identity document. No model was trained on these selfies, or on these people. The overall figures are for an evaluation mix of 10% under-18s, with adults in the age distribution of live traffic; each age range is reported on its own below.

Accuracy Metrics

Metric	Value	Description
Mean Absolute Error (MAE)	2.6 years	Average difference between estimated and document age
Standard Deviation	2.3 years	How much the error varies from one estimate to another
Correct at an 18 threshold	96.0%	Selfies correctly classified as under 18 or 18 and over, with no fallback band; at least 95.3% with 95% confidence

Performance Across Age Ranges

The error is smallest around the ages where age gating decides, and grows with age.

Age range	MAE (years)	Standard deviation (years)
13-15	1.70	1.17
16-17	1.14	1.37
18-20	1.86	1.74
21-25	2.21	1.78
26-40	2.77	2.20
41-60	3.03	2.62
Over 60	4.33	3.64