Why Benchmarking and Explainability are so Important in Identity Verification

A recent blog from Microblink CEO Hartley Thompson explored how generative AI is fundamentally changing the economics of fraud by making sophisticated attacks cheaper, faster, and dramatically more scalable. In many ways, the advancements made in Gen AI to fight fraud  are also used to create more sophisticated attacks at scale. In fraud, volume is everything: even low-success, low-payout attacks become profitable when they can be automated massively.”

That’s why at the Microblink Fraud Lab, we continuously test emerging fraud techniques ranging from injected video streams and face swaps to AI-generated documents and synthetic onboarding flows. What stands out is not just how realistic these attacks are becoming, but how accessible they have become. Many of the same tools powering entertainment, content creation, and consumer AI experiences can now be repurposed for fraud with very little technical knowledge. Fraud is no longer just becoming more sophisticated. It is becoming easier to produce, easier to scale, and easier to automate.

Fraud is Becoming Multi-Modal

A major shift is that fraud is no longer isolated to a single attack vector. Modern attacks increasingly combine multiple techniques simultaneously across documents, biometrics, devices, and behavioral manipulation within the same operation. A fraudster might generate a synthetic identity document, inject an AI-generated selfie stream into a verification flow, spoof device signals, and automate portions of the attack chain simultaneously. In other words, fraud is becoming multimodal.

This is important because many identity systems were originally designed to evaluate isolated signals at specific stages of the user journey. A document check happens at onboarding. A biometric check happens during authentication. A device signal gets evaluated during login. But modern attacks increasingly blur those boundaries. Fraudsters are combining AI-generated media, automation, behavioral simulation, and impersonation techniques into layered attack flows designed to exploit gaps between systems rather than a single point of failure.

Traditional Verification Models Were Built for a Different Era

Many existing identity systems still operate on assumptions that are increasingly outdated. Historically, documents forgery required highly specialized skills, biometrics were difficult to spoof, attackers were mostly human-operated, and onboarding represented the primary point of risk. Generative AI is rapidly breaking down those assumptions.

The challenge is no longer simply determining whether a document appears authentic or whether a face matches an ID photo. Organizations now need to determine whether an image itself was synthetically generated, whether a verification session is using a live camera feed or an injected stream, whether a voice has been cloned, whether behavioral patterns are automated, and whether a trusted identity remains trustworthy over time. This requires layered systems capable of evaluating multiple signals continuously rather than relying on isolated verification events.

Benchmarking Matters More Than Ever

All of this is why benchmarking and independent evaluation are becoming critically important across the identity industry. One of the biggest concerns with many emerging AI-powered trust systems is opacity. Organizations are increasingly being asked to trust black-box models without clear visibility into what attack vectors were tested, how systems were evaluated, which scenarios remain out of scope, how models behave under adversarial pressure, or how reliability changes over time. Unlike many traditional ML domains, fraud detection operates in a continuously shifting adversarial environment where out-of-distribution scenarios are the norm rather than the exception.

This creates real operational risk. AI systems often perform impressively in controlled demonstrations while behaving very differently in real-world fraud environments where attackers continuously adapt, probe weaknesses, and iterate against defenses. Identity systems therefore need to evolve continuously as well. That is why adversarial testing, reproducible evaluation, and independent benchmarking are becoming essential parts of modern fraud defense. Organizations need to understand not only where systems succeed, but where they fail, because attackers will inevitably identify those gaps themselves.

Another major challenge emerging alongside AI adoption is explainability. As organizations integrate larger AI systems into identity and fraud workflows, fraud decisions risk becoming increasingly opaque. In regulated industries such as financial services, healthcare, marketplaces, and digital commerce, explainability is not optional.

Organizations increasingly need visibility into why a verification passed or failed, which signals contributed to risk, what changed between decisions, and how authority or trust were evaluated at a given moment. This becomes even more important as AI agents and automated systems begin interacting directly with financial and identity infrastructure. Without explainability and auditability, organizations gradually lose visibility into trust itself.

The Future of Fraud Defense Is Continuous

The broader shift underway is that fraud is no longer static. It is adaptive, automated, AI-assisted, and continuously evolving. Defensive systems must evolve in the same way.

There will not be a single model or technology capable of solving this problem independently. Modern fraud defense increasingly requires layered identity intelligence spanning document verification, biometrics, liveness detection, device intelligence, behavioral analysis, contextual risk evaluation, and continuous monitoring. Most importantly, these systems must continuously adapt as attack methodologies evolve, because the fraud landscape is no longer changing at a human pace. AI is accelerating the speed, scale and automation of attacks.

At Microblink, this is one of the reasons we launched the FREUID Challenge: to engage the research community on realistic out-of-distribution fraud detection problems operating under adversarial conditions.
If you’d like to chat more about Microblink’s capabilities in detecting Gen AI fraud, contact us today.