Selfie-and-Document Checks at Account Opening: Everywhere, but Bank-Level Accuracy Goes Unpublished

The use case

In November 2025, India's identity authority recorded 28.29 crore (282.9 million) face authentication transactions in one month, against 12.04 crore in November 2024 (official release, 2 December 2025)⁶. Those checks match a live face to a national record across many services, not only banking, but they show the scale at which face matching has become routine.

At account opening the same idea works in three steps. The customer photographs an identity document; the system tests whether the document looks genuine; and the customer takes a selfie or short video, which software compares with the photo on the document. A "liveness" check tries to confirm that a real person is present and that the image is not a printout, a replayed video or a synthetic fake.

Scope note. AI here means machine-learning components: face matching, liveness (presentation attack) detection and document-forgery detection. The sources read describe "algorithms" and do not always state the method, so the machine-learning share of any deployment is not established. Older technology and human work do similar jobs and are not credited as AI: document inspection by staff, database lookups, rule-based checks and face-to-face verification. Login authentication and fraud monitoring after the account is open are out of scope.

Where it is running

In the EU, the European Banking Authority's guidelines on remote onboarding have applied since 2 October 2023 to credit and financial institutions². No named bank with published onboarding volumes or results was found, so the table shows national and sector-level scale only. US deployments are older: a $28.5 million settlement over an identity verifier's collection of face data was sought for approval in 2023 (plaintiffs' law firm report, 2 May 2023)¹⁰.

What the evidence shows

Independent testing covers components, not whole onboarding. The strongest evidence is the US National Institute of Standards and Technology (NIST) evaluation of passive liveness detection, published in September 2023. It tested 82 algorithms from 45 developers on about 20,000 attack attempts and 21,000 genuine presentations across nine attack categories (measured in a controlled offline test)³. At a setting where 1% of genuine users are wrongly flagged, four algorithms missed none of the photo print or replay attacks in the category reported. In two categories of evasion attack, where a person tries to avoid being recognised, the best algorithm still missed 13% and 62% of attacks³. NIST found that accuracy "varies widely" across algorithms, use cases and attack types, and that combining algorithms beat the best single one in all impersonation categories³.

The evaluation did not cover digital injection attacks, in which a fake image or video is fed directly into the software instead of being shown to the camera. It also excluded hardware approaches and challenge-response checks, where the user is asked to perform an action³. A test of camera-facing fakes therefore says little about the attack route that vendors describe as growing.

Threat figures come from regulators without numbers and from vendors. The US Financial Crimes Enforcement Network (FinCEN) reported on 13 November 2024 that suspicious activity reports about deepfake media had increased from 2023 into 2024, and that accounts had been opened with identities suspected to be produced with generative AI; the alert gives no counts (regulator observation)¹. Entrust reports that deepfakes made up 1 in 5 biometric fraud attempts, that injection attacks rose 40% year on year and that deepfaked selfie attempts rose 58% in 2025 (vendor-reported, from its own verification events; periods for the first two figures not stated)⁸. LexisNexis Risk Solutions reported on 14 July 2026 that 1 in 100 failed identity checks involves a deepfake, with attacks up 180% year on year (vendor-reported; the release gives no method or period)⁹. The two sets cannot be compared or added: one divides by biometric fraud attempts, the other by failed checks, and neither isolates account opening.

Lab attacks on liveness checks. A 2022 preprint, older than three years and not peer-reviewed in the version read, reported deepfake evasion rates above 95% on some image-based liveness services it tested (measured in a research test, 2022)¹¹. No newer replication was found, so the figure shows that the weakness has been demonstrated, not how common it is in production today.

What it takes: cost, integration, governance

Cost. No official per-check price, contract value or total cost of ownership was found in the sources read. Cost is a gap.

Process. The EBA guidelines of 22 November 2022 set out what a supervised institution must do². For unattended solutions they require liveness detection and "strong and reliable algorithms" to match the photo or video with the picture on the document. They require tests for impersonation fraud and end-to-end testing before launch, and ongoing monitoring, including sample testing and manual reviews, that the institution can demonstrate to its supervisor. A policy must state which steps are fully automated and which need a person, and where evidence is of insufficient quality the process must be restarted or redirected to face-to-face verification². The guidelines are technology-neutral and, in the text read, do not use the word "deepfake"².

Attack defences. FinCEN lists warning signs for remote checks, including use of a third-party webcam plugin during a live verification and device or location data inconsistent with the identity documents. It recommends re-reviewing account-opening documents, reverse image searches, deepfake detection software and multifactor authentication¹. A webcam plugin matters because it is one way of feeding a fake video into a live check.

Law. Recitals 15 and 54 of the EU AI Act state that systems whose sole purpose is to confirm that a person is who they claim to be are excluded from the biometric identification category treated as high-risk⁵. The articles and annexes could not be read in this run, so the full classification of onboarding systems and the application dates were not verified. In the US, class actions under Illinois biometric privacy law add a separate exposure: the $28.5 million settlement sought in 2023 was proposed in a case against the identity-verification firm Onfido (plaintiffs' law firm report)¹⁰.

Where it falls short

The challenge question is what would make this use case look less successful than it is presented. The answer is that the benefit usually claimed, less fraud and faster onboarding, is not measured anywhere in the sources read.

No outcome data. None of the sources read measures fraud losses avoided or onboarding time saved by AI checks. The NIST tests, the regulator guidance and the vendor statistics describe parts of the system, not its result.

Vendor sources. Both attack-trend datasets are published by firms that sell identity or fraud products⁸⁹. The figures may be accurate, but no method is disclosed that would let a reader test them.

Uneven effects. NIST's 2019 demographic study, older than three years and covering general face recognition on government image sets rather than selfie-to-document matching, found that false positive rates "often vary by factors of 10 to beyond 100 times" across demographic groups (measured in a controlled test, 2019)⁴. No newer figure for selfie-to-document matching was found. In the UK, 41% of surveyed certified organisations hold accuracy information across demographic groups, up from 30%, and 52% of those using biometrics offer a non-biometric route (self-reported, 2025); the report publishes no pass rates by group and does not assess individual services⁷.

Dependence on people and process. The EBA requires defined human intervention, trained staff in attended solutions and a face-to-face fallback², so results depend on process design and supervision as well as on the matching software.

Failures and withdrawals. Searches found litigation¹⁰, regulator warnings¹ and lab bypasses¹¹, but no regulator ban and no withdrawn account-opening programme. A trade-press report of a Spanish data-protection fine against an identity-verification provider was not confirmed against the regulator's decision and is not relied on.

Phase: In early use — why

The hypothesis was "In wide use". Identity checking at this scale is established: national face authentication in the hundreds of millions of transactions a month⁶, UK services above one million checks a month⁷ and a vendor reporting over a billion events⁸. But only confirmed machine-learning use counts towards the phase, and none of those figures separates machine learning from other methods, or onboarding from other uses. Where the scale of confirmed AI use cannot be established, the lower plausible phase is chosen. A re-run with bank-level data could support "In wide use".

What to watch next

• Final court approval of the Onfido settlement, which was sought in 2023 and not confirmed as approved in the source read¹⁰.
• Any FinCEN publication of the counts behind its reported rise in deepfake-related suspicious activity reports, which were tagged with the key term FIN-2024-DEEPFAKEFRAUD¹.
• The next UK inclusion monitoring report, following the 16 July 2025 edition, to show whether the 41% of organisations holding demographic accuracy information rises⁷.
• The next editions of the Entrust and LexisNexis reports, to see whether methods and periods are stated for the deepfake figures⁸⁹.
• Publication in the Official Journal of any act amending the application dates of the AI Act's high-risk rules.

Questions to ask before adopting

• What error rates, for wrongly rejected genuine customers and wrongly accepted impostors, has the vendor measured on this institution's own customer mix, split by demographic group?
• Does the system defend against injection attacks as well as camera-facing fakes, and under which independent test?
• Which steps are automated, which go to a trained person, and what route exists for customers who fail?
• How will results be monitored after launch through sample testing, manual review and fraud outcomes, and what can be shown to the supervisor?
• Where is facial data stored, for how long and on what legal basis, and who bears liability under the vendor contract?