Banks can use AI to fight the dangers of AI

What if AI is used against you? BankBeat spoke with Jessica Gonzalez, general manager and vice president of customer success for Informed.IQ, a fintech founded in 2016 that offers AI-powered loan verification and fraud prevention..How can AI be used against community banks, both as an internal tool and as a criminal’s tool?Jessica Gonzalez: The same AI functionality community banks are adopting for efficiency and customer experience, which are proving to be highly valuable, are also being leveraged for fraud or increasing risk.On the criminal side, generative AI has fundamentally changed the economics of fraud. Tools widely available to consumers—not sophisticated crime rings—can now produce convincing paystubs, bank statements and employment records with realistic formatting, accurate employer details, and even subtle imperfections designed to appear genuine. Adding to this challenge, our recent survey found that 80 percent of lenders expressed little to no confidence their current verification systems could detect a GenAI-created counterfeit paystub, or AI fake. Meanwhile, roughly one in five paystubs submitted in auto loan applications are now forged.On the internal side, the risks are subtler but serious. Generic, off-the-shelf AI models can "hallucinate,” generating plausible but incorrect information. Fifty-two percent of lending executives we surveyed identified hallucinations as their top concern when implementing standard Large Language Models in document processing. A bank using a generic AI model to review loan documents could approve falsified submissions it was never designed to catch. The threat is compounded by the fact that AI-enabled fraud moves faster than traditional risk models, meaning losses often don’t surface until three to six months after funding, when recovery options are limited.When it comes to manual, human review of this scenario, what role does the human banker play?JG: As fraud detection increasingly becomes a machine-speed competition, it’s tempting to assume the human banker is being squeezed out. The reality is more nuanced and more important. Human judgment remains essential, but the role must evolve from document inspector to pattern analyst and exception handler.AI systems are well-suited to flag anomalies at scale: Metadata irregularities, circular fund movement through peer-to-peer apps designed to manufacture healthy cash flow, and cross-document inconsistencies that a busy underwriter might miss. What AI cannot do well—at least not yet—is apply the kind of contextual, relationship-based judgment a community banker is uniquely positioned to bring. A long-standing member with a 20-year account history and a sudden application inconsistency warrants a different response than an unknown borrower with the same red flag.What kind of training should banks be using when it comes to AI and how to spot AI when it’s coming for you instead of helping you?JG: Training programs at most community banks struggle to keep pace with how quickly the fraud environment changes. Annual refresher courses were not designed for a world where a convincing fake paystub takes less than five minutes to generate. Banks need training that is both conceptual and tactical.On the conceptual side, every customer-facing employee—not just underwriters—should understand that document fraud is no longer the province of organized crime. It’s accessible, fast and often undetectable to the human eye. That psychological shift matters. Complacency is a fraud vector.On the tactical side, training should focus on behavioral and systemic red flags rather than visual document inspection, which AI fakes are specifically engineered to pass. Staff should be trained to look for inconsistencies across the full application: Does the employer on the paystub match the deposit history? Does the stated income align with the broader financial picture? They should also understand when to escalate to automated verification tools rather than relying on visual review. Banks with AI fraud detection tools need staff training specific to those platforms: How to interpret flags, when to override,and how to document decisions for examiners. Training should also be held regularly given how rapidly the threat landscape shifts.Regulation is always a concern. What’s your snapshot of compliance at the moment?JG: The regulatory environment is shifting quickly, but the direction is clear. Explainability, auditability, and bias documentation are no longer optional considerations. They are table stakes for any AI use in lending or verification.Before deploying any AI model, banks should require vendors to produce model documentation that satisfies examiner expectations, including training data provenance, bias testing results, validation methodology, and governance protocols for model updates.Banks should also establish internal AI governance policies that define who owns the model, how often it’s reviewed, and how overrides are documented. Regulators are not opposed to AI, but they are opposed to black-box AI that cannot be explained and audited. Banks that treat governance as a precondition rather than an afterthought will be in a far stronger position when examiners arrive.