Facial recognition has evolved from Woodrow Bledsoe’s RAND tablets in the 1960s to today’s deep learning systems, which now protect digital identities against fraud. We look at the key milestones in this evolution, how the technology works, how accurate it has become, its main limitations, and why multimodal biometrics and liveness detection are already fundamental components of digital identity verification.

What is facial recognition?

This type of technology extracts discriminative biometric features from a facial image and models them as a unique numerical pattern to distinguish between individuals.

Facial recognition is based on extracting biometric features with sufficient discriminative power to distinguish one person from another.

Each facial image is modelled using a series of numerical values which, when grouped together, form what is commonly known as a feature pattern.

These patterns can subsequently be used to perform different operations: calculating the similarity between two faces (matching), identifying a person within a group (identification) or verifying their identity, that is, checking whether someone really is who they claim to be (verification).

The technology has been widely adopted because of its robustness, reliability and low-intrusion nature, as the process requires little more than an image.

Unlike other biometric methods such as fingerprint or iris recognition, facial recognition requires neither physical contact nor specialised hardware, only a standard camera, which helps explain its rapid adoption.

Who was the first to develop facial recognition technology?

The mathematician Woodrow Wilson Bledsoe developed the first facial recognition system in the 1960s, digitising facial features using RAND tablets and electromagnetic coordinates.

Woodrow Wilson Bledsoe, a mathematician working in the 1960s, created a prototype using RAND tablets that made it possible to digitise facial features through electromagnetic coordinates. The system compared new images against a pre-existing database, laying the foundations for modern facial biometrics.

Bledsoe did not work alone

Between 1964 and 1966, at Panoramic Research in Palo Alto, the project was completed by Helen Chan Wolf and Charles Bisson, co-authors of the technical report “A Man-Machine Facial Recognition System, Some Preliminary Results” (1965). The system was semi-automated: a person manually marked points such as the centre of the pupils or the corners of the eyes on a graphics tablet, at a rate of around 40 photographs per hour, and the computer calculated 20 normalised facial distances to compensate for head rotation, tilt and size. In subsequent tests involving more than 2,000 photographs, the computer consistently outperformed human operators.

When Bledsoe left the project in 1966, Peter Hart continued the work at the Stanford Research Institute (SRI). Wolf also joined SRI, where she contributed to image processing and coordinate extraction for Shakey, the world’s first autonomous mobile robot, a project recognised with an IEEE Milestone in 2017, the same distinction awarded, for example, to Marconi’s first transatlantic radio transmission.

How has facial recognition evolved to the present day?

Eigenfaces and object detection

From Bledsoe’s manual system, facial recognition progressed to Eigenfaces in 1991, the Viola-Jones algorithm in 2001 and deep learning in 2014, when DeepFace reached accuracy levels comparable to the human eye.

Following Bledsoe’s work, anthropometric methods based on facial landmarks emerged. In 1971, researchers Goldstein, Harmon and Lesk proposed identifying faces using 21 subjective markers, such as lip thickness or hair colour, although these measurements still had to be calculated manually, with no real automation.

In 1991, MIT researchers Matthew Turk and Alex Pentland introduced the Eigenfaces method, based on Principal Component Analysis (PCA). It represented any face as a weighted combination of a reduced set of eigenfaces, making real-time facial recognition viable for the first time at a reasonable computational cost.

A decisive milestone came in 2001, when Paul Viola and Michael Jones proposed a framework for object detection using Haar features, enabling real-time face detection on modest hardware for the first time.

The creation of Mobbeel

Alongside these global advances, Mobbeel was founded in Spain. In 2009, it succeeded in recognising people through their irises using the mobile phones of the time, becoming a global pioneer in iris biometrics on mobile devices, before Google launched Face Unlock in 2011 or Apple introduced Touch ID in 2013. That same year, the project became a global finalist in the first Android Developer Challenge organised by Google, giving the Extremadura-based company an international profile from the outset and making it part of the history of facial recognition.

What happened after 2014

The next major leap came in 2014, when the Facebook AI team (Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato and Lior Wolf) presented DeepFace at the CVPR conference: a system based on deep neural networks that achieved 97.35% accuracy on the LFW benchmark dataset, almost matching the 97.53% achieved by humans.

Google responded in 2015 with FaceNet, which, rather than directly classifying faces, learned to map them into a 128-dimensional space using a function known as triplet loss: bringing images of the same person closer together and separating those of different people. The result was a leap in accuracy to 99.63% on the same LFW benchmark.

In 2019 came ArcFace, developed by the InsightFace team. It introduced an additive angular margin loss function to better separate identities in feature space. It became the de facto standard for training facial recognition models, with more than 8,000 academic citations and widespread adoption in commercial frameworks.

Today, deep learning algorithms derived from this lineage — DeepFace, FaceNet and ArcFace — predominate, creating their own feature extractors by analysing complex relationships within an image, without relying on descriptors predefined by humans.

This technical progress coincides with an expanding market. The global facial recognition market is estimated at $10.13 billion in 2026, with a forecast to reach $30.52 billion by 2034, according to Fortune Business Insights.

The EU AI Act as a regulatory milestone in the history of facial recognition

The regulatory framework has also become clearer. Since February 2025, the EU AI Act recognises 1:1 identity verification with consent, the model used in onboarding processes, as a permitted use, while its restrictions focus on real-time biometric identification in public spaces.

How accurate is facial recognition today and what are its limitations?

Today’s most accurate algorithms have low error rates across demographic groups, although real-world accuracy depends on the provider, image quality and protection against deepfakes.

NIST evaluation of facial recognition algorithms

NIST (the US National Institute of Standards and Technology) has maintained the FRVT/FRTE programme (Face Recognition Vendor Test) since 2000. It is an independent, ongoing evaluation of the sector, operating indefinitely with no fixed end date rather than being a one-off report. Its demographic effects studies alone have evaluated nearly 200 algorithms from around 100 different developers, using a database of more than 18 million images of more than 8 million people.

Any specific figure from this evaluation should be viewed as a snapshot in time rather than a fixed value. Because the programme is continuous, rankings and error rates change with each new round, meaning that figures that are relevant today may become outdated. It is therefore advisable to consult the current comparison on the NIST website rather than quoting a frozen figure.

The challenge of facial identity spoofing

Another ongoing challenge is facial identity spoofing through deepfakes. A University of Florida study published in March 2026 found that, when presented with AI-generated facial images, people performed no better than chance, while detection algorithms achieved up to 97% accuracy. With videos, the pattern was reversed: people correctly identified authentic rather than fake videos around 67% of the time, while the algorithms performed at chance level, apparently because humans can detect subtle inconsistencies in movement. Neither approach is sufficient on its own. This is why no serious facial verification system relies solely on face matching; it is combined with liveness detection to check that a real person is present in front of the camera, rather than a photograph, video or mask.

Why is facial recognition combined with other biometric modalities?

Multimodal biometrics combines facial, voice, fingerprint and document verification to compensate for the limitations of each individual modality and make fraud more difficult.

Layering multiple biometric factors mitigates the weaknesses of each one individually. Changes in lighting can affect facial recognition, for example, while background noise can affect voice recognition. It can also improve the user experience when liveness detection runs passively, without noticeable friction. A landmark NIST study on biometric fusion showed that combining fingerprint and facial recognition reduces the false rejection rate by between 50% and 90% compared with using a single modality, precisely because the scores from the two modalities are largely independent of one another.

This is not a future trend. It is already here.

How can you verify that there is a real person behind an AI agent?

As AI agents become increasingly autonomous, buying, booking and negotiating on behalf of a person, two distinct but related questions arise. The first is an identity problem concerning the agent itself: is it who it claims to be, and does it continue to operate within the limits it has been given, when there is no face, fingerprint or passport to verify? Mobbeel has already addressed this question in detail in KYC for robots: how to verify the identity of an AI agent. The second question, which is more closely related to facial biometrics and the one we are concerned with here, is different: when an agent acts on behalf of a specific person, how can we prove that a real person is authorising that action, rather than a script or a synthetic identity?

This second concept is increasingly being referred to as proof of human, and the face is emerging as the most practical way to address it.

The technical foundation already exists and is standardised: facial liveness detection, which distinguishes a real person present in front of the camera from a photograph, video or mask. It is governed by the international ISO/IEC 30107 framework, which defines how the resistance of a biometric system to spoofing attempts is evaluated and certified.

The same liveness detection technology that already protects digital onboarding can also safeguard the “human” half of this problem. Recognising a face is not enough; it is necessary to demonstrate that a real person is physically present behind that face at the very moment the action takes place. This is likely to become one of the next major application areas for facial recognition.

What is facial biometrics currently used for?

Today, facial biometrics is used primarily for digital onboarding regulated by KYC (Know Your Customer) and AML (Anti-Money Laundering) requirements. Sectors such as banking, insurance and telecommunications need to verify customers’ identities remotely before opening an account or activating a service. BFSI (Banking, Financial Services and Insurance) is the sector with the greatest economic weight, accounting for 23.55% of the global facial recognition market in 2026, driven precisely by these compliance requirements.

Border control is another area where the technology is already operating at scale. On a smaller scale, it is also used for physical access control, criminal identification, finding missing persons and unlocking devices — established applications, although they represent a smaller share of the market.

Healthcare, for example to streamline patient registration, and security are in fact the sectors expected to experience the strongest growth through to 2034, with annual growth rates of 19.9% and 19%, respectively. This is even faster than the banking sector, which already starts from a much larger base.

Behind this adoption lies a specific economic problem: traditional identity fraud cost US consumers $23 billion in 2024. Against this backdrop, biometrics is already gaining ground over traditional passwords. In a FIDO Alliance survey of 10,000 consumers across ten countries, 29% preferred fingerprint or facial scanning to log in, compared with just 19% who preferred entering a password manually.