Europe labelled the deepfakes that want to be seen. Bank identity checks sit on the other side

Last week the EU's transparency rules for synthetic content came into force, requiring that AI-generated material carry a machine-readable tag ^[3][4]. The obligations that would hold the systems banks use to check a customer's face or voice to a standard of manipulation resistance did not arrive with them, and that gap is where the fraud actually lives ^[9][10].

The argument comes from an opinion piece in American Banker by a deepfake forensics practitioner based in Amsterdam, who founded a detection company after a close friend's likeness was used in a romance scam ^[1][12]. Their point about Article 50 is structural rather than technical: the tagging duty catches synthetic content that wants to be noticed, while fraud runs on the opposite logic and needs discretion ^[5]. Nobody attacking a bank's onboarding flow is going to attach the marker that identifies them, and the author is explicit that this is a boundary of the rule rather than a defect in it ^[5][15].

The consequential detail is a definitional one. Biometric identification, matching a face or voice against a database of many, appears on the AI Act's high-risk list ^[6]. Biometric verification, checking one person against their own record, does not ^[6]. Verification is the cheaper, faster one-to-one check, and it is increasingly the mechanism for onboarding remote customers ^[7]. The author cites a trial in the Netherlands earlier this year in which the check the bank relied on was precisely that unregulated category ^[8].

Compliance teams may read an exemption as relief. The author argues the opposite: high-risk classification would eventually require providers to demonstrate greater resilience against manipulation, and verification vendors will not be held to that ^[9][16]. The industry has built its identity layer on the assumption that a voice or a face is sufficient evidence of who someone is, and AI has broken that assumption ^[2]. What has not happened, in the author's view, is any requirement that verification platforms carry liability for their own system errors ^[14].

The timing compounds it. The high-risk obligations were originally meant to land alongside Article 50 and were pushed back, so for roughly the next year banks operate under a regime that tells them to label AI-generated material while leaving the systems they depend on unaccountable ^[10][17]. The proposed fix is narrow and worth noting for its modesty: add biometric verification to Annex III ^[11]. That is a drafting change, not a new regulatory architecture.

For US operators this is not a foreign story. AI-driven fraud in banking is occurring on both sides of the Atlantic, banks in both markets are buying the same categories of defensive technology, and the author expects US rules to follow Europe's ^[13]. The EU version is simply the version you can read now.

Watch three things. Whether biometric verification gets pulled into Annex III, which would change vendor procurement questions from accuracy claims to manipulation-resistance evidence ^[11][9]. When the delayed high-risk obligations actually apply, because the interim is the exposure window ^[10]. And whether liability for verification failures starts moving from the bank to the vendor in contracts before it moves in statute ^[14].