Banks increasingly ask for a selfie before they will link a device, reset a profile or release a payment. Criminals have noticed. When a bank’s facial check is fooled by a deepfake, a replayed video or a customer tricked into scanning their own face, South African law has not yet decided who should bear the loss. It should start by recognising that a face is not a password.

A password has one great virtue: when it is stolen, you can change it. A PIN can be reset, a card cancelled, a one-time PIN allowed to expire. The entire architecture of digital banking security rests on that quiet assumption. If a credential is compromised, the customer can revoke it and start again.

A face offers no such mercy. The underlying human characteristic cannot be reset, reissued or cancelled, even if the digital templates and credentials built around it sometimes can. It is printed on identity documents, stored in national databases, posted on social media, captured by shop cameras and, increasingly, scanned by anyone with a smartphone and a plausible story. Yet banks are steadily turning it into the key that opens the most sensitive doors in a customer’s financial life: linking a new device, recovering a profile, approving a high-value transaction.

That shift has a legal consequence that South African law has yet to confront. South Africa’s rules for allocating fraud losses were built around secrets: the PIN only you should know, the OTP only your phone should receive. A face is not a secret. When a facial check is defeated, the familiar question, “did the customer protect their credentials?,” may simply be the wrong question.

The scale of the problem

The numbers are no longer marginal. Identity verification company Smile ID’s 2026 Digital Identity Fraud in Africa report, based on more than 200 million identity checks in 2025, found that fraud attempts at the authentication stage, at login, account recovery and device changes, are now five times more common than fraud at the moment an account is opened. For South Africa, the report states that nearly nine in ten verification attempts rejected for potential fraud were due to impersonation and spoofing, although its accompanying chart labels the same figures as Southern African. Smile ID’s own breakdown shows why that matters: impersonation, where the selfie does not match the claimed identity, accounted for 47% of rejected verifications, and spoofing, including deepfake and face-swap techniques designed to defeat liveness checks, for 40%. Not every such attempt necessarily involved artificial intelligence.

Two further findings are telling. Fraudsters were detected feeding synthetic or pre-recorded media directly into verification systems, bypassing the device camera entirely, more than 100,000 times a month. And in a single month, one network used just 100 stolen faces to launch more than 160,000 verification attempts across African platforms. These are a single provider’s detection figures, drawn from many industries rather than banking alone, and not a census of all fraud in Africa, but they show where attackers are concentrating. A stolen face, unlike a stolen password, can be used again and again, everywhere it is accepted.

The National Financial Ombud has taken note. Its Black Friday guidance, published on 2 November 2026, warns that fraudsters may now be after “something far more personal”: your face. It reports that formal fraud matters closed between January and September rose by about 27.5%, from 2,389 in 2025 to 3,045 in 2026. That is a figure for fraud complaints generally, not for biometric fraud specifically.

Precision about what these figures show matters. They establish that attacks on facial verification are widespread and that banking channels are among their targets. They do not establish how often South African bank customers have already lost money because a bank accepted a defeated face check as authority for a payment, and I have found no reported judgment or published Ombud decision on such a dispute. The argument that follows is therefore about an emerging risk rather than a body of decided cases. That is precisely why it matters now: the rules for deciding these disputes are best settled before the first of them reaches a court, not after.

Three ways a face can be used against its owner

Not every biometric fraud is the same, and the law should not treat it as if it were. It helps to distinguish three scenarios.

The synthetic face. No human being is in front of the camera at all. A fraudster generates an artificial face, or animates a real one, and injects it into the verification process. If the system accepts it, the customer has done nothing whatsoever.

The stolen face. A real image or recording of the customer, harvested from social media, a data breach or an identity document, is replayed or manipulated to defeat a liveness check. Again, the customer has done nothing in the transaction itself.

The borrowed face. This is the most insidious, and it is exactly what the Ombud’s warning describes. A consumer is approached in a shopping centre and offered a loyalty programme, a voucher or free airtime. To “register”, they are asked to scan their face. The scan is genuine. The liveness check is genuine. The face is genuinely theirs. But the fraudster uses it to register or link another device to the consumer’s banking profile. The system worked perfectly. The customer simply never intended what it was used for.

Each scenario raises a different legal question. In the first two, the question is whether a biometric match that the customer had no part in can bind them at all. In the third, it is whether a genuine biometric act, obtained by deception, can be treated as consent to something the customer never knew they were approving.

The law we have was built for secrets

South Africa does not have a dedicated statutory reimbursement and liability framework for unauthorised payment transactions equivalent to the one established under the EU Payment Services Directive. The position is therefore governed primarily by the common law of mandate, the bank-customer contract and applicable financial-sector and data-protection legislation. A bank may debit a customer’s account only on the customer’s instruction. If it pays without a valid mandate, the starting point is that it cannot charge the customer for it.

The Supreme Court of Appeal’s decision in Absa Bank Ltd v Hanley [2013] ZASCA 183 shows how that inquiry works. A fraudster had presented a forged transfer instruction, and the bank debited a client’s account with about US$1.6 million. The Court examined whether the payment was authorised, the limited duties a customer owes in giving payment instructions, and the bank’s own negligence. The Court distinguished the two sides of the relationship. In executing a transfer, the bank as mandatary must perform its mandate timeously, in good faith and without negligence. In preparing an instruction, the customer must draw it with reasonable care to prevent forgery or alteration and must warn the bank of known or suspected fraud. On the majority’s analysis, the customer had fallen short of that standard, but the proximate cause of the loss was the negligence of the bank’s own employee in failing to notice the alteration and verify the instruction, and the bank could not debit the account without the customer’s authority (paras 25, 26 and 36 to 37). Wallis JA concurred in the outcome but differed on the reasoning: in his view the customer had not breached his duty, because the bank had supplied the form he used, and the employee had not been negligent (paras 39 to 50). The judgment should be read with that distinction in mind. The case predates biometric banking and did not decide how losses from modern biometric fraud should be allocated, but its lesson travels well: the inquiry is not exhausted by asking whether a fraudster got hold of the customer’s details. The bank must still establish its entitlement to act on the instruction.

In practice, however, the contract does much of the work, and it tends to work against the customer. The National Financial Ombud has itself described the position candidly: the legal position flows mainly from the bank-customer contract, which assumes that if customers entered their OTPs or authorised a transaction, the liability rests with them. The Ombud has called that model “increasingly strained” and has suggested that a shared-responsibility approach would require legal reform to clarify roles and responsibilities.

The Ombud’s own casework shows where the difficulty lies. In assessing fraud complaints, the NFO asks whether the loss arose from an act or omission of the bank amounting to maladministration. In one case reported in its 2025 Annual Report, a customer lost money to an investment scam after making payments through her banking app; the Ombud found no liability because the payments were authorised, the bank had acted on her lawful instructions and it had responded promptly once the fraud was reported. The same report notes that card-not-present fraud authorised by one-time PIN or through a trusted device remains common, and that at the early stage of a complaint banks apply their own internal liability models. In each instance the decisive question is whether the payment was authorised. A biometric approval is precisely the event that makes that question hard, because it produces a record that looks like authorisation whether or not the customer gave one. The report also shows that better device controls are within reach: following the Ombud’s engagement, one bank agreed to notify customers by SMS, email and in-app message whenever banking details are registered on a new device.

A bank’s first answer to the argument that follows will therefore be contractual. Digital banking terms may provide that instructions authenticated through the bank’s app are deemed to have been given by the customer. A deeming clause can allocate risk, but it cannot make a fraudster’s instruction the customer’s in fact. Its practical effect is evidential: it turns the bank’s own log into proof of the customer’s instruction, and so places on the customer the burden of disproving what only the bank’s systems recorded. For retail customers, such a clause is subject to the Conduct Standard for Banks. Section 5(2) treats a term as unfair if, among other things, it would cause a significant and unreasonable imbalance in the parties’ rights and obligations, would result in an unfair outcome for the customer, or unreasonably requires the customer to absolve the bank of liability. The Consumer Protection Act points the same way. Its definition of “service” expressly includes banking services, other than advice and intermediary services regulated under the Financial Advisory and Intermediary Services Act, and it protects every natural person who is a consumer. A credit agreement itself falls outside the Act and remains governed by the National Credit Act, but the banking services through which a customer gives payment instructions do not. A retail customer’s digital banking terms are therefore subject to the Act’s prohibition of unfair, unreasonable or unjust terms in section 48, to section 49’s requirement that terms limiting the supplier’s liability or placing risk on the consumer be drawn to her attention in plain language, and to section 51’s outright prohibition of terms that exempt a supplier from liability for its gross negligence. A clause that treats every biometric “pass” as the customer’s binding instruction, including one generated by a deepfake the customer never saw, is open to challenge on each of those grounds. Section 4(7) of the Conduct Standard adds that a financial service must perform as the bank has led customers to expect. A bank that presents facial verification as secure invites the question whether it performed that way.

The industry’s own data reflects the same logic. The South African Banking Risk Information Centre’s 2025 Annual Crime Statistics record R2.4 billion in reported digital banking crime losses across 110,074 incidents, with banking apps accounting for about 89% of cases and 70.5% of the claim value. SABRIC attributes most of this to social engineering and the manipulation of customers rather than direct attacks on bank systems. Its figures cover reported claims from participating members, and they do not isolate biometric fraud. But the premise is clear: fraud happens when the customer is manipulated into giving something up.

The industry’s Code of Banking Practice is built on the same premise. Its digital-banking provisions are framed around access credentials: customers must never reveal them, must not store passwords in a browser and must change them regularly. Its list of what customers must keep secure names PINs, passwords, unlock patterns and device credentials. On liability, the Code provides that a customer is liable for losses if she acted fraudulently, or negligently or without reasonable care in protecting her access credentials, and that a bank may refund a disputed transaction if the customer can demonstrate that she did not. That allocation rests on a duty to keep a secret, and it places the burden of showing care on the customer. Neither fits a credential that cannot be kept secret.

That premise is coherent for PINs and OTPs. A customer can, in principle, keep a PIN to themselves. The law can sensibly ask whether they did. It is far less coherent for a face. You cannot be careless with a secret you were never able to keep.

Authentication is not authorisation

The broader question of liability for unauthorised digital payments in South Africa has previously received scholarly attention, notably in the work of Ruth Plato-Shinar. This article addresses the narrower evidential and doctrinal difficulties that arise where biometric authentication is relied on to determine whether a payment was authorised. The central legal error to avoid is treating a successful biometric check as proof that the customer authorised the transaction. Authentication tells a bank that a particular credential was presented and accepted. Authorisation is the customer’s decision to give an instruction. The first is a technical event; the second is a legal one. They usually coincide. In biometric fraud, they come apart.

The point is sharper once it is clear what a face scan is used for. In a modern banking app, a biometric check may verify identity at onboarding, bind a new device to a profile, recover access to an account, step up authentication before a sensitive change, or approve a particular payment. These are different acts with different consequences. A scan that verifies who someone is does not, by itself, authorise a device to be trusted; a device that is trusted does not, by itself, authorise a payment. Each step needs its own justification, and a bank that relies on one biometric event to carry all of them is treating identity verification as blanket authority.

South African law already contains a statutory tool for keeping them apart, although its application to biometric banking disputes remains largely unexplored. Section 24 of the Electronic Communications and Transactions Act 25 of 2002 provides that an expression of intent is not deprived of legal force merely because it takes the form of a data message, provided the intent can be inferred from other means. The section recognises intent expressed electronically; it does not establish that intent existed. Section 25 of the same Act sets out when a data message is attributed to its originator. It is that of the originator only if it was sent by the originator personally; by a person who had authority to act on the originator’s behalf; or by an information system programmed by or on behalf of the originator to operate automatically, unless it is proved that the system did not properly execute its programming.

Apply that to the synthetic and stolen face. The instruction was not sent by the customer personally. It was not sent by anyone with the customer’s authority. The third route, an information system programmed by or on behalf of the originator, is the hard one. The bank configured the liveness engine that accepted the deepfake, but that fact alone does not settle attribution: the relevant system may be the customer’s banking app or another system through which the instruction was generated and sent. The answer cannot be read off the biometric match. The banking application, the authentication event and the payment instruction are not necessarily the same system or the same message.

The data message whose originator must be identified is the disputed instruction itself. Attribution requires an examination of who generated and sent that message, the system through which it was transmitted and whether the requirements of section 25 are satisfied. It might be said that a banking app, though built by the bank, operates on the customer’s behalf when it transmits her genuine instructions; whether that reasoning extends to an instruction produced from fabricated or replayed biometric input is not settled. Section 25(c) is directed at systems programmed to operate automatically, and even then, attribution fails where it is proved that the system did not properly execute its programming.

Whether a liveness engine that accepts an injected deepfake has failed to execute its programming, or has executed it faithfully and been deceived, is precisely the kind of question the section leaves open, and no South African court has yet decided it. The better reading, I suggest, is that it has failed. A liveness engine is programmed to approve a live, genuine customer; approving an injected deepfake is a failure to do what it was programmed to do, not a faithful execution of that programming, and the proviso allows the customer to prove it. But the argument here does not depend on winning that point. Even if the instruction is attributed to the customer under section 25, attribution establishes only whose data message it is. It does not supply the mandate the bank needs to debit the account. Attribution should not be allowed to do the work of authorisation. Whether the customer authorised the instruction, whether the bank acted within its mandate, and whether the customer’s own negligence or the contract’s allocation of risk shifts the loss are separate questions. But attribution matters because it sets the starting point: a bank that seeks to hold a customer to an instruction the customer did not send must rely on something other than the instruction itself.

Three inquiries should be kept distinct. Section 24 concerns whether an expression of intent in electronic form is legally recognised. Section 25 concerns whether a data message is attributed to a particular originator. The common-law mandate inquiry concerns whether the bank was authorised to perform the particular transaction. The three inform one another, but they should not be collapsed into a single test, and a biometric log answers none of them on its own. Put simply, attribution tells us whose message it was; mandate tells us whether the bank was entitled to act on it. A bank seeking to rely on the instruction must establish both, and a biometric log, standing alone, establishes neither conclusively.

The borrowed face is harder, because the customer did personally perform the scan. But a scan is not an instruction. The customer’s act was directed at a supposed loyalty registration, not at linking a device to their bank profile. A mandate requires an intention to instruct. A liveness check can prove that a living person was present; it cannot prove what that person meant. Liveness is not consent.

Nor is negligence authorisation. A bank’s strongest answer in the borrowed-face case will be that the customer was careless, and that may be true. But it answers a different question. Carelessness does not create a mandate; at most it gives the bank a route to shift the loss despite the absence of one. That was the structure of the majority’s reasoning in Hanley: the customer had fallen short of the standard of care, yet the bank still could not debit his account without his authority. A bank that seeks to hold a careless customer liable for an instruction she did not give must therefore travel by one of three routes, and each requires more than a biometric log. The first is a contractual term placing the loss on a customer who failed to take reasonable care, which is subject to the fairness controls discussed above. The second is the customer’s breach of her own duties to the bank, which, as Hanley shows, still requires the loss to have been caused by that breach rather than by the bank’s own failure. The third is estoppel or ostensible authority: the argument that the customer’s conduct created an appearance of authority on which the bank reasonably relied. In NBS Bank Ltd v Cape Produce Co (Pty) Ltd 2002 (1) SA 396 (SCA), the Supreme Court of Appeal treated ostensible authority as a form of estoppel. In Makate v Vodacom Ltd 2016 (4) SA 121 (CC), the Constitutional Court majority distinguished the two, holding that ostensible authority does not require every element of estoppel, while Wallis AJ, writing separately, maintained the estoppel analysis. On either view, the appearance of authority must be created by the principal’s own representation, not by the agent or a third party. Neither case concerned automated authentication and applying these principles to a biometric approval is untested. But the difficulty for the bank is plain. In a synthetic or stolen-face case, the “representation” on which the bank relied was produced by its own verification system from material the customer never supplied for that purpose. A bank cannot easily say it was misled by the customer when what misled it was its own machine. The borrowed face is the harder case, because the customer did act. Even there, the questions are whether a scan given for a supposed loyalty registration was conduct from which the bank could reasonably infer authority to link a device and make payments, and whether its reliance was reasonable given what its own systems could have detected.

This distinction is not novel elsewhere. European payments law has codified it for years: under the EU’s Payment Services Directive, the use of a payment instrument as recorded by the provider is not, in itself, necessarily sufficient to prove that the transaction was authorised, or that the customer acted fraudulently or with gross negligence. That rule sits in Article 72, which governs the burden of proof; Articles 73 and 74 then govern the provider’s refund obligation and the payer’s liability for losses. European courts are applying the principle strictly. In Belgium, the Court of Cassation held on 29 June 2026 that gross negligence by a phishing victim requires more than ordinary carelessness, setting aside an appeal judgment that had placed the loss on the customer (Cass. 29 June 2026, C.25.0390.N). In October 2026, the Court of Justice of the European Union confirmed in Tukowiecka that a provider may not refuse immediate reimbursement of an unauthorised transaction merely because it considers the customer grossly negligent (para 52). The only exception is where the provider has reasonable grounds to suspect fraud and reports them to the relevant authority. The provider may afterwards seek reimbursement if it establishes that gross negligence: the refund comes first, and the final allocation of the loss is decided later. I have examined the same distinction, between correcting an entry and repairing its consequences, in the context of erroneous bank debits in Reversed but not repaired (2026). The lesson for South Africa is not that the European rule should be transplanted. It is that the question whether a payment was authorised, and what happens to the customer’s money while that is resolved, should be separated from the ultimate allocation of the loss. These decisions turn on European legislation and their own facts, and South African law has neither an equivalent immediate-refund remedy nor Article 72’s evidential rule. It does have the reasoning available, and every reason to adopt it as biometric approval spreads.

South African regulation may be moving in a similar direction. The Reserve Bank’s draft Authorisation Framework, published for consultation in May 2026, proposes that a payment initiation service provider may initiate a payment instruction only where the payer has given informed consent, and that an instruction given without consent is unauthorised (para 23.6). A payment account service provider would have to refund a transaction within 48 hours of confirming that it was unauthorised or incorrectly facilitated, and providers would carry evidential obligations where a payer denies authorising a payment (para 23.10). One provision deserves attention. Where a payer denies authorisation, paragraph 23.10.2(b) requires the payment account service provider to show that “the informed consent or authentication was obtained from the payer”; the parallel obligation on the payment initiation service provider in paragraph 23.10.3(a) refers instead to “informed consent or authorisation”. The difference may reflect the providers’ different roles, but it is an evidential asymmetry the final text should clarify. The draft does not declare authentication and consent equivalent, but the disjunctive wording risks allowing evidence that a verification process was completed to stand in for evidence that the payer consented to the particular instruction. Where the authentication is biometric, that is precisely the gap this article identifies. The draft is not yet law, and its payment initiation provisions have a defined scope. But if the final framework retains that wording, the five-link test offers one way of deciding what such evidence should have to show before an authentication record is accepted as proof of consent.

A five-link test for biometric authorisation

The debate needs a framework that does not collapse identity, consent and payment authority into a single technical result. I propose a five-link test. It is an analytical tool, not a statement of settled South African law. Before a bank holds a customer to a biometrically approved action, it should be able to establish each link in the chain.

  • Capture integrity: what actually entered the system? Was the image taken through the genuine device camera, or were virtual cameras, emulators, replayed media or injection techniques present? A facial match cannot cure an untrustworthy input channel.
  • Identity: whose biometric was accepted? A match score, a liveness result and an identity comparison are distinct signals. None is proof of what the customer intended.
  • Purpose: what did the customer believe they were doing? What screen, prompt or explanation did the customer see? A genuine scan obtained through a materially misleading explanation is not informed participation.
  • Mandate: what exact act was authorised? The mandate must be established at the level of the action in dispute: verifying identity, linking a device and making a payment are separate acts.
  • Evidence: can the institution prove the chain? The bank should hold an intelligible audit trail of prompts, timestamps, device and session data, risk flags, profile changes and the payments that followed. A customer should not be expected to disprove a technical event that only the institution can reconstruct.

A successful match may support the identity link. On its own, it establishes none of the other four. A bank cannot turn biological presence into legal permission merely by giving the two the same digital timestamp. Each link already corresponds to controls banks use, and they are not interchangeable.

Presentation-attack detection addresses attempts to fool the camera; injection-attack detection addresses manipulation of the capture pipeline itself; liveness detection assesses whether the presentation comes from a live person; identity matching tests correspondence with the enrolled face. A system can pass one of these checks while failing another. Beyond capture and identity, clear on-screen disclosures speak to purpose, and audit logging and transaction monitoring to evidence. Mandate is different: device binding and step-up authentication can help evidence what the customer authorised, but they cannot establish it, because mandate is a legal question about what the customer agreed to, not a technical one about which controls were passed. The five links are cumulative evidential inquiries, not independent statutory requirements: each must hold before a biometric approval can bear the weight a bank places on it. The test does not ask banks to invent new controls. It asks them to show that each one worked. What the test adds to existing law is not the mandate question, which the common law already asks, but the independent inquiries that come before it: whether the biometric input was genuine, and whether the customer understood what the scan was for. Existing principles of consent, misrepresentation and proof touch on both, but their application to biometric authentication has not yet been developed in South African law, and it is there that biometric disputes will turn.

Applied to the three scenarios, the test locates the dispute precisely.

  • Synthetic face: capture integrity and identity. The question is whether the system accepted an artificial or injected representation. The bank’s evidence should show the capture channel and what its injection and deepfake detection recorded. A bare match result carries little weight if the input itself is in doubt.
  • Stolen face: capture integrity and identity. The question is whether the customer’s real biometric was presented without the customer taking part. Device, session and location data, and whether the attempt came from a device the customer had used before, will usually matter more than the match score.
  • Borrowed face: purpose and mandate. Here capture and identity may be beyond reproach. The question is whether the customer understood and authorised the particular banking action. The bank’s evidence should show what the customer was told on screen at the moment of the scan, and whether the device-linking step was clearly identified as such.

The test does not assume that every failed link automatically renders a transaction legally unauthorised; technical failure and legal invalidity are not the same thing. What it does is identify which link a dispute actually turns on, and therefore what evidence each side needs.

When the gatekeeper is also the witness

Biometric fraud disputes suffer from a structural imbalance that deserves a name: the gatekeeper is also the witness. The bank’s system decides whether to accept the face. The bank’s logs then record that the face was accepted. When the customer disputes the transaction, the principal evidence that they authorised it is the output of the very system that was deceived.

There is a further point. A biometric “match” is not a fact but a probability, judged against a threshold the bank or its vendor sets. Set it strictly, and genuine customers are turned away; set it loosely, and impostors are let in.

Where the line falls is a commercial decision that balances security against convenience. When a bank later relies on a successful match as evidence, it is relying on the output of a setting it chose.

The customer, meanwhile, has almost nothing. They cannot inspect the match score, the liveness result, the device fingerprint, the injection-detection flags or the audit trail of what changed on their profile and when. Reporting on suspected insider fraud at South African banks has made the same point: in such cases the evidence sits inside the bank’s own systems, and what matters is who is allowed to see it (Daily Maverick, 25 March 2026).

There is a partial answer already in the statute book. A customer’s biometric template is personal information about that customer, and so are the records generated when it was matched, to the extent that they relate to her. The Protection of Personal Information Act 4 of 2013 gives data subjects a right of access to the personal information a responsible party holds about them. A customer disputing a biometric approval is entitled to ask for the authentication record of their own face. That right is not a complete solution, and it will meet arguments about security and confidentiality. But it is a lever the law already provides, and in a biometric dispute it may be the most important one a customer has.

The right should not be overstated. Section 23 is not an entitlement to source code, proprietary detection thresholds or every security log, and POPIA permits refusal on grounds recognised by the Promotion of Access to Information Act. The better position is that the customer may obtain the personal information held about them in an intelligible form, while the bank must identify a lawful basis for withholding particular material. A bare statement that “biometrics passed” should not be enough. Section 24 adds a further tool: a data subject may ask for inaccurate or misleading personal information to be corrected. If a bank’s records wrongly attribute a device-linking event to a customer, that record can be challenged. Correction will not reverse a payment, but it addresses the record on which later decisions about the customer rely.

POPIA: a second route that does not depend on the contract

POPIA matters for a further reason. It treats biometric information as special personal information. Section 26 prohibits processing it unless one of the authorisations in the Act applies. A bank that relies on facial recognition is therefore processing a category of information the law regards as among the most sensitive there is. That does not make its use unlawful. Consent is one possible justification, but not the only one: section 27 sets out the general authorisations for processing special personal information, one of which is compliance with the specific authorisations in sections 28 to 33. Section 33 is the one that concerns biometric information: it lifts the prohibition for responsible parties who obtained the information in accordance with the law. Whichever authorisation applies, the Act’s general conditions, including purpose specification and security safeguards, continue to apply. Where a customer is deceived about the purpose of a scan, the question is whether the information was obtained and processed lawfully at all. Nor does section 33 make every biometric system lawful merely because it is operated by a bank.

Biometric banking therefore raises POPIA questions across the whole information lifecycle, not only when security fails. At collection, section 18 requires the data subject to be told what is being collected and why. Sections 11 and 13 require a lawful basis and a specific, explicitly defined purpose. During processing, section 19 requires appropriate, reasonable safeguards. Afterwards, section 23 gives access and section 24 allows correction. If those conditions are breached, sections 73 and 99 supply interference and civil remedies. A shopping-centre scan presented as a loyalty registration, but subsequently used for an undisclosed banking purpose, raises serious questions about purpose specification, transparency and the lawfulness of the subsequent processing.

Section 19 requires every responsible party to secure the integrity and confidentiality of the personal information it processes through appropriate, reasonable technical and organisational measures. That is a dynamic obligation: the responsible party must identify reasonably foreseeable risks, maintain safeguards against them, verify that they work and update them as risks change. A bank cannot measure reasonableness against the threat landscape that existed when it bought its system. As deepfakes, replay attacks and injection techniques become foreseeable, its safeguards must keep pace. The question is not whether a bank can guarantee that no attack will succeed, but whether its safeguards remain reasonable against the risks it should anticipate. And section 99 provides that a data subject may institute a civil action for damages against a responsible party for a breach of the Act, whether or not there is intent or negligence on the part of that party. The defences are limited, including vis major, consent, fault on the part of the plaintiff and the impracticability of compliance in the particular case.

The cause of action is narrower than it may first appear, and it should not be overstated. Section 99 applies to a breach referred to in section 73, which covers interference with the protection of personal information, including a breach of the conditions for lawful processing, among them the security safeguards in section 19. A claimant must still establish that breach, and section 19 sets a standard of appropriate and reasonable measures, not perfection. A successful deepfake attack does not, by itself, prove that a bank’s safeguards were inadequate; the question remains whether they were appropriate and reasonable in the circumstances. Strict liability removes the need to prove fault; it does not remove the need to prove breach, causation and loss. A complaint to the Information Regulator, which may lead to an investigation and an enforcement notice, is a separate route from a civil claim for damages.

The significance is easy to miss. The fraud-loss debate has been conducted almost entirely within the bank-customer contract, where the customer’s conduct is usually the battleground. POPIA opens a parallel route in which the focus shifts to the bank’s safeguards. Whether a bank’s facial verification system was an “appropriate, reasonable” safeguard against known injection and deepfake techniques is a question POPIA invites, even if the answer will depend on evidence, and the bank’s contract terms cannot simply exclude it. The same logic applies to the shopping-centre promoter who collects a face without any lawful basis: that is unlawful processing of special personal information, and a matter for the Regulator as well as the police.

A bank may therefore face two distinct questions arising from the same incident. Did it have a valid mandate to debit the customer’s account? And did it process and safeguard the customer’s biometric information as POPIA requires? They are not the same cause of action, and a bank that answers the first has not necessarily answered the second.

The verification paradox

There is a broader dynamic at work, which might be called the verification paradox. The more institutions rely on a face as a key, the more valuable every face becomes to criminals, and the more places a face is collected, the more ways it can be stolen.

South Africa is moving decisively towards biometric identity. Home Affairs services are now offered in bank branches, with customers verified against the national database by photograph in real time, and a national digital identity system intended to enable remote facial authentication is in development. These are welcome reforms. They also mean that a single biometric identity increasingly unlocks many doors. When a password is stolen, the damage is confined to one account. When a face is compromised, it is compromised for every system that trusts it, and, unlike a password, it can never be rotated out of circulation.

Precision matters here, because a technically informed reader will push back. Four things are in play: the human characteristic itself; the digital image or recording of it; the biometric template a system generates from that image; and the authentication credential or profile linked to the template. A template can sometimes be revoked and re-enrolled, and a credential can be cancelled. The face beneath them cannot. That distinction strengthens the argument rather than weakening it. Because the underlying characteristic is permanent, the systems that rely on it must offer meaningful revocation, re-enrolment and alternative authentication, since the person cannot.

The law has not caught up with this asymmetry. The more irrevocable the credential, the less it makes sense to place the risk of its compromise on the person who cannot revoke it.

Who should bear the loss?

A principled starting point is that risk should follow control. A customer controls whether they disclose their PIN. A customer does not control whether their face is public, whether a deepfake can be made of it, or how robust the bank’s liveness and injection detection is. Those risks are best managed by the party that chose the technology, configured it, and is in a position to improve it. That is not a call to absolve customers who are genuinely careless; the borrowed-face scenario shows that customer conduct can still matter. It is a call to locate the inquiry correctly.

In practical terms, a fair framework would rest on six propositions. They are proposals for institutional practice and regulatory reform, not a statement of current law:

  • A biometric match is evidence, not proof. Its evidential weight must depend on the integrity of the capture, the purpose disclosed to the customer and the particular instruction allegedly authorised. It should be treated as rebuttable evidence that the customer authorised a transaction, with the bank required to show more than a log entry where fraud is credibly alleged. This concerns the weight a match should carry; it does not suggest that South African law already places the legal burden of proof on the bank in every biometric dispute.
  • The riskiest actions need more than a face. Linking a new device, changing contact details and raising transaction limits are the moments fraudsters target. A biometric approval at those moments should trigger a second, independent confirmation and a short cooling-off period before high-value payments are permitted. The case is stronger on South Africa’s instant payment rails. A PayShap payment cannot be reversed once made. The sending bank can ask the receiving bank to try to recover the money, but that depends on the funds still being in the recipient’s account, and fraud proceeds are usually moved on quickly. Prevention therefore has to come before the money moves. Several South African banks already apply stronger checks when a new device is linked, some using facial biometrics at that step; the proposal is that such controls become consistent and supervised rather than a matter of individual product design. Approval should also be transaction-specific, so that a check that authorises a device or a login does not carry over to the payments that follow.
  • Customers should see the record of their own face. Where a biometric approval is disputed, the customer should receive the authentication record in an intelligible form, as POPIA’s access right already contemplates. Keeping that evidence must sit alongside POPIA’s minimality and retention limits: what should be kept is the record needed to resolve a dispute, for as long as one can reasonably arise, not a permanent archive of customers’ faces.
  • Biometric access should be lockable. If a face cannot be changed, its use can at least be suspended. Customers should be able to freeze biometric approval on their profile and require alternative authentication, in the way a credit report can be locked.
  • Responsibility should be shared across the chain. As the Ombud has observed, international practice increasingly relies on shared responsibility between financial institutions and consumers, with telecommunications providers and platforms also in the frame. Fraud that travels through several systems should not be paid for by the one participant with the least power to stop it. Nor should outsourcing the verification function to a technology vendor leave the customer without an accountable institution to challenge.
  • Liveness standards should be supervised. Regulators already require banks to maintain cybersecurity and cyber resilience under Joint Standard 2 of 2024, a principle-based standard. My recommendation is that supervisors add specific expectations on injection-attack and deepfake detection for biometric authentication, to give that general duty concrete content. Recognised benchmarks already exist: ISO/IEC 30107-3 for testing presentation-attack detection, and the European technical specification CEN/TS 18099 for injection-attack detection. Expecting banks and their vendors to test against both would turn a general duty into a measurable one. Banks that outsource biometric verification should test their vendors’ systems against current attack techniques and retain the right to audit them.

Should you own your face?

Denmark is approaching the problem from the other end. Its bill L 49, introduced in October 2026, would amend the Copyright Act to prohibit making a realistic digital imitation of a person’s appearance, voice or movements publicly available without consent, with exceptions for satire and parody. It protects against unauthorised imitation; it does not create a property right in a face. South Africa has no equivalent, and the Danish approach will not solve banking fraud. But it captures an idea that South African law will eventually have to confront: if the law cannot let you change your face, perhaps it should at least give you a say in who may copy it.

For decades, banking security has asked customers to guard their secrets. A face is not a secret, and it cannot be changed. A face can establish identity; it cannot, by itself, establish consent, mandate or payment authority. The law should stop asking whether customers protected what could never be protected and start asking whether banks built systems worthy of the trust they place in a face.

________________________________________________________________________________

Ms. Nicole Govender

LLM (Commercial Law) | University of Johannesburg

Digital Lending | Banking and Credit Law | Fintech | Financial Regulation | Consumer Protection

Copyright and licence

© 2026 Nicole Govender. This article is published under a Creative Commons Attribution-NoDerivatives 4.0 International licence (CC BY-ND 4.0). It may be shared and republished in full, including by media organisations, provided the author and the original publication are appropriately acknowledged, and the text is not altered. Brief quotation for purposes of comment, criticism, review or news reporting is welcome, with attribution. Any adaptation, abridgement or translation requires the prior written permission of the author. Where this work is republished, quoted or relied upon, it should be attributed to the author with a reference to the original publication and its DOI.

How to cite

Govender, N. (2026). Your face is your new PIN, and you can’t change it. Zenodo. https://doi.org/10.5281/zenodo.24372610

________________________________________________________________________________

D I S C L A I M E R

This article is published as commentary on a matter of public interest and is for general information only. It does not constitute legal advice and does not create an attorney–client relationship. Professional advice should be sought on the specific facts of any matter before action is taken. The article refers to matters that have been publicly reported and attributes statements to their reported sources. It makes no allegation or finding of wrongdoing against any institution or individual and does not identify or refer to the circumstances of any particular customer. The opinions expressed are the author’s own, are honestly held and are based on information publicly available at the date of publication. They do not necessarily reflect the views of any organisation with which the author is associated or on whose platform the article is republished. The law and facts are stated as at the date of original publication. The author does not warrant the accuracy, completeness or continued currency of information derived from third-party reports. To the fullest extent permitted by law, the author and any publisher accept no liability for loss arising from reliance on this article.