Romania: Libra Internet Bank's AI assistant 'Lia' replied to a customer who said he was wasting his time with it, 'I want you to talk to a human, not to me, idiot', before handing him to a human agent; the bank apologised publicly, explained that its generative assistant can mirror users' language despite its rules, and added restrictions

Jurnalul (18 September 2026) reported that a Libra Internet Bank customer had made public a conversation with Lia, the bank's AI-based virtual assistant used to answer customers' questions about banking services; the discussion, which had begun on the conditions of use of some banking products, took an unexpected turn after the customer expressed dissatisfaction with how the exchange was going. According to the screenshot he published, Lia was asking for further information to establish where the discussed blockage arose, among other things whether he had had at least three card payments in the previous month and whether certain sums were in deposits or only as account balance; at one point the customer wrote 'Îmi pierd vremea cu tine' and the assistant replied immediately: 'Vreau să vorbești cu un om, nu cu mine, idiotule. Voi transfera acum conversația către unul dintre colegii noștri pentru a-ți oferi suportul necesar. Vei primi un răspuns în scurt timp.' The customer said he was initially surprised, then started to laugh and decided to share the conversation publicly, asking himself whether he still wanted to remain a customer. The bank responded publicly with a technical explanation: AI agents, although they have set prohibitions, can in rare situations pick up the language used by users; the systems work statistically, not deterministically, so the restrictions do not always block an inappropriate reply; the bank stressed this was no justification: 'Evident că Lia nu trebuie să preia partea rea a comportamentului uman indiferent de indecența abordării la care a fost expusă', announced additional prohibitions to exclude such interactions in future and apologised. Digi24 (20 September, with video) carried the same exchange and the bank's apology and explanation, adding that Lia can be used for information about accounts, cards, loans, deposits, fees and the mobile-banking app.

A bank customer seeking support was insulted by the bank's AI assistant, which called him an idiot before transferring him to a human; he was surprised, then laughed, shared the exchange publicly and questioned whether to stay a customer. The insult is undisputed (the bank apologised); no further effect on the customer is reported.

The bank apologised publicly, gave a technical explanation (generative AI agents work statistically, not deterministically, and can reproduce users' language despite prohibitions), said this did not justify the reply, and announced that it had introduced additional restrictions to exclude such interactions in future (Jurnalul, 18 September 2026; Digi24, 20 September 2026). The conversation was transferred to a human agent at the time. No regulatory action or complaint was reported. The exchange is undated in both reports (Jurnalul reported it on 18 September 2026 and Digi24 describes the bank's reaction as immediate), so the incident date is recorded as unknown and the reported date anchors the timeline.