Last updated: 6 November 2026
No, face-to-face checks are not automatically safer, and neither is fully automated AI. Today the most secure way to verify identity is remotely: read the cryptographic chip in the ID document, confirm that a live person is in front of the camera, and send every case that falls short of certain to a trained human reviewer. This article is for compliance, risk and procurement teams choosing how customers or signers are identified. It compares the three methods against what EU law now expects.
The hybrid model is the only one that combines chip-level document checks, human judgement on uncertain cases and a full evidence record.
| Face-to-face | Fully automated AI | Hybrid (AI + human) | |
|---|---|---|---|
| Document check | Visual check, rarely with tools | Chip validation; image analysis without a chip | Chip validation plus a human check of security features |
| Right person? | Staff judgement; a lookalike can pass | Automated face match and liveness | Face match score; unclear cases go to an officer |
| Deepfakes and injection | Not exposed: no camera | Main exposure; depends on tested detection | Tested detection plus human review |
| New fraud tactics | Depends on the person | Missed until the model is retrained | Spotted by a human and fed back |
| Evidence afterwards | A copy and a register entry | Logs of the automated decision | Full evidence package |
| Speed and reach | Travel and opening hours | Minutes, from anywhere | Minutes; longer if reviewed |
| eIDAS Art. 24(1a) route | Point (d), physical presence | Point (c), if conformity confirmed | Point (c), if conformity confirmed |
Face-to-face feels safest because a person is in the room, but a counter check is only as good as the person doing it and the tools on the desk. In most branches, offices and notary rooms, the check is a look at the photo and a glance at the document. Nobody shines a UV light on it, compares security features against a reference database, or reads the chip. A good forgery, or a genuine document held by someone who looks like its owner, passes that check.
The second weakness shows up later. An in-person check usually leaves a photocopy and an entry in a register. If someone challenges the identity years afterwards, there is little to show how the decision was made.
EU law does not rank physical presence above remote methods. Article 24(1a) of eIDAS, as amended in 2024, lists physical presence as one route to a qualified certificate, next to remote methods that "ensure the identification of the person with a high level of confidence". The AMLR, Annex III treats a non-face-to-face relationship as a risk factor to weigh, not as high risk by default. The FATF Guidance on Digital Identity goes further: non-face-to-face onboarding that relies on reliable, independent digital ID systems "may present a standard level of risk, and may even be lower-risk."
Automated verification is fast and consistent. It can validate a chip signature, run a liveness check and match a face to a document photo in seconds, to the same standard every time. Where it stops is new fraud: a model recognises the attacks it was trained on, and fraudsters keep inventing others.
"Anyone who tells you their system cannot be fooled is describing a system that doesn't exist," says Tomas Zuoza, CEO of ZealiD. On deepfakes, his answer is the same: "It shouldn't work on our process. ... But I won't say 'can't.' It's a cat-and-mouse game."
The standard that EU trust services are measured against accepts automation, with conditions. ETSI TS 119 461 V2.1.1 allows fully automated, unattended identity proofing at the Extended level, the level suitable for qualified certificates. It also requires presentation attack detection and injection attack detection. For the Extended level, injection attack detection must be lab-tested to CEN/TS 18099 level High before the end of 2026 and retested every two years. Automation is allowed when it can prove it resists attack, not by default.
The AI Act is often mentioned as a brake here, but it targets a different use. Annex III, point 1(a) of the AI Act makes remote biometric identification high-risk, and explicitly excludes biometric verification whose only purpose is to confirm that a person is who they claim to be. A one-to-one face match during onboarding falls in that excluded category. Article 50 obliges deepfake creators to label their content, which is no help to a verifier: fraudsters do not label.
Hybrid verification lets the machine do what it does well and gives the uncertain cases to a person. The machine provides speed and consistency. The person handles judgement and spots tactics the model has not seen yet. For more on how attacks get past automated checks, see injection attacks vs. liveness checks.
Every case goes through the same three automated checks. Only the highest confidence level is decided without a person, and every route ends in the same evidence package.
This is how it works in ZealiD's own process, as set out in its public Trusted Registration Authority Practice Statement:
The standard also closes the opposite door. ETSI TS 119 461 states that purely manual remote proofing, a person checking a person over video, does not reach the Extended level. A human alone is not enough remotely, and neither is a machine that cannot hand off.
"The point of layered defenses isn't that no layer ever fails. It's that you have enough time to respond when one does," Tomas says.
When someone claims a signature or an account was never theirs, the provider must prove it was. "The burden of proof lands on us," Tomas says. "We are guilty until proven innocent, the reverse of how courts normally work."
The evidence decides that challenge. ZealiD's practice statement lists what is logged for each identification: the reviewing officer, any fraud reason, the face map from the liveness check, pictures and video of the person, and the chip signature and photo. These form the evidence package, and audit logs are kept for 12 years. A face-to-face check that left a photocopy cannot answer the same challenge.
Two deadlines raise the cost of choosing the wrong method. Commission Implementing Regulation (EU) 2025/1566 makes ETSI TS 119 461 the reference standard for remote identity verification under eIDAS Article 24(1a)(c) from 19 August 2027. And Article 22(6) of the AMLR, which applies from 10 July 2027, allows identity verification through electronic identification at assurance level substantial or high and relevant qualified trust services. A process that cannot show conformity by then will have to be rebuilt.
Ask any provider four questions before you sign:
"If a vendor ever tells you their system is 100% secure, that's not a security claim. That's a sales claim," Tomas says.
As a Qualified Trust Service Provider on the EU Trusted List, ZealiD identifies people with a hybrid process whose conformity with eIDAS Article 24 and ETSI TS 119 461 at the Extended level has been confirmed by a conformity assessment body. That means every qualified signature traces back to an evidence package that holds up when an identity is challenged. See how ZealiD identity verification works.
It is safe when it is tested and when it does not decide alone. Automated checks are fast and consistent, but they only recognise fraud they were trained on, so the safest setups send every uncertain case to a trained human reviewer.
Is video identification as secure as face-to-face verification?It depends on how the video is checked. A person checking someone over a video call alone does not reach the Extended level of ETSI TS 119 461, which qualified certificates require. Remote verification that reads the document chip, runs tested liveness detection and sends unclear cases to an officer does.
Do I need to meet someone in person to issue a qualified electronic signature?No. Article 24(1a) of eIDAS allows physical presence or a remote method that ensures a high level of confidence, confirmed by a conformity assessment body. Both are valid routes to a qualified certificate.
How does AI identity verification handle deepfakes?Through liveness and injection attack detection, plus human review. ETSI TS 119 461 requires both detection layers, and for the Extended level the injection attack detection must be tested to CEN/TS 18099 level High. Cases the system is unsure about go to a trained officer.
Does the EU AI Act ban AI identity verification?No. The AI Act treats remote biometric identification as high-risk, but it explicitly excludes one-to-one verification whose only purpose is to confirm that a person is who they claim to be. A face match during onboarding falls in that excluded category.
Will AI replace human reviewers in identity verification?Not for cases that are unclear. ETSI TS 119 461 allows fully automated decisions, but providers that keep a human for uncertain cases catch new fraud patterns the model has not seen yet.
Regulation (EU) No 910/2014 (eIDAS) as amended by Regulation (EU) 2024/1183, Article 24(1a). European Union, 2024.
Commission Implementing Regulation (EU) 2025/1566 of 29 July 2025, reference standards for the verification of identity for qualified certificates. European Union, 2025.
Artificial Intelligence Act, Regulation (EU) 2024/1689, Article 3(36), Article 50 and Annex III point 1(a). European Union, 2024.
Anti-Money Laundering Regulation (AMLR), Regulation (EU) 2024/1624, Article 22(6) and Annex III. European Union, 2024.
ETSI. "TS 119 461 V2.1.1: Policy and security requirements for trust service components providing identity proofing of trust service subjects." February 2025.
FATF. "Guidance on Digital Identity." March 2020.
ZealiD. "Trusted Registration Authority Practice Statement (TSPS) v19." January 2026.
Learn more