Two years ago, deepfakes—AI-generated videos or images that impersonate real people—were a novelty. Today, they are standard within a fraudster’s arsenal.
Generative AI has transformed once-specialized attacks into assembly-line operations, including account takeovers (stealing access to existing accounts), phishing (impersonating trusted institutions to steal credentials), and synthetic identities (fabricating entirely new people).
Criminals can now use large language models (LLMs) to hijack a victim’s phone number to intercept verification codes, and run software that mimics real devices to run hundreds of fake accounts. They operate dark web marketplaces where stolen data is bought and sold.
In our work at Incognia, we’ve watched manual scams that once required teams and days now run solo in seconds. Call centers have given way to automation, and media manipulation is now a one-click workflow. This is what an industrial revolution in fraud looks like.
Global financial fraud losses hit $442 billion in 2025, with generative AI identified as a primary accelerant. The lesson for leaders is brutally simple: Fraudsters don’t wait for permission or quarterly reviews. Institutions must match their machine speed while anchoring digital trust to persistent, real-world physical behavior.
3 examples of the new fraud factory floor
Here are three examples of how AI fraud is ramping up.
1. Job onboarding fraud
Job onboarding was once a reliable gatekeeper. Users scan government IDs and submit selfies with head turns, blinks and on-screen prompts.
Fraudsters can sidestep this entirely. They harvest breached personal data, feed it into image generators for flawless forgeries, then use video models to mimic exact liveness motions. App-tampering technology injects pre-recorded media directly into an app’s data pipeline, bypassing the camera entirely and convincing the system it’s seeing a live feed.
2. Financial fraud
Deepfake-related losses in banking and financial technology topped $410 million in 2025 alone. Generic phishing emails evolved into hyper-personalized social engineering campaigns scripted by language models to nail institutional tone, weave in scraped account details and spawn polished replica sites, all in minutes.
Voice scams reduce call-center overhead by using synthetic audio to convincingly impersonate branch managers or executives and field questions with contextual memory.
Research projects that annual fraud losses could hit as high as $40 billion by 2027 as these attacks scale without human limits. Most fraud victims authorize payments confident they’re dealing with legitimate support.
3. Large-scale consumer fraud
Automation completes the triad, and the scale is staggering. The FTC reported $15.9 billion in consumer fraud losses in 2025, up 430% since 2020. AI coding agents now orchestrate device farms: racks of phones or virtual emulators (software that mimics real devices) that spin up parallel apps, auto-reset hardware IDs, and test stolen credentials across thousands of accounts hourly.
One physical location can generate hundreds of unique identities to evade rules designed for individual bad actors. These fraud farms are difficult to detect despite being geographically concentrated. Most fraud detection tools can flag a suspicious cluster, but cannot distinguish the fraud operation from legitimate users in the same building.
From that altitude, a fraud farm and its neighbors look identical. Detection at a more granular level, down to the specific device patterns within a location, is what turns signals into actionable enforcement.
Trust anchored in physical reality
These attacks thrive in purely digital environments where content rules. Physical reality provides an antidote, as AI struggles to fabricate real human traces at volume. Does this device show months of continuity at the claimed home address? Or is it one of dozens suddenly appearing from the same apartment block? Are high-dollar-value transfers to new payees occurring mid-phone call from an unfamiliar location?
Location intelligence, paired with device integrity (whether a device shows signs of tampering or resets) and behavioral continuity (a consistent pattern of movement over time), reveals patterns invisible to content analysis.
A deepfake can pass a liveness check; the fact that the same phone has visited the same home, office, and gym for months cannot be faked. Fraud-motivated coordinated rings reveal themselves through an unnatural number of devices at a single physical location, repeated factory resets from identical geographic coordinates, and login spikes from a single location.
Federal Reserve data show that 48% of consumers’ hesitancy toward instant payments stems from fears of fraud. That fear is solvable. Unlike fraudsters, who operate from warehouses of devices or rotate through the same handful of locations, real customers move through the world in consistent, recognizable patterns. Systems that verify those patterns can protect consumers invisibly, without adding friction to every transaction.
Anchor consumer trust in the physical frontier
Fraudsters move at machine speed, unconstrained by policy, ethics, or headcount. Companies must weaponize AI with surgical guardrails. Release security teams to harness language models for stress-testing systems against simulated attacks, synthesizing real-time threat across channels, and flagging anomalies the moment they emerge—tasks that previously required weeks of manual analysis.
As consumer AI agents proliferate for bill pay, transfers, investments and daily finance, it’s critical to verify human intent through context that cannot be faked. This includes real-world location patterns, device history, and behavioral signals. This is currently the only way to link digital actions to verified humans in the physical world.
Fraudsters may have written the AI playbook first and continue to execute without remorse, but smart leaders know that they can burn that script by rooting digital trust in physical reality.
André Ferraz is the cofounder and CEO of Incognia.