A.I.P.D.

Anti-scam tools are catching criminals in real time

A scammer dials what appears to be an elderly woman’s number. Over the next 40 minutes, the caller is led through meandering stories about a cat named Fluffy, detailed accounts of a knitting project, and a set of bank details that correspond to no real account. The voice on the other end belongs to dAIsy, a conversational AI deployed by UK carrier O2 and designed for a single purpose: every minute spent on the decoy is a minute unavailable for real victims.

dAIsy is one entry in a broader wave of AI-powered anti-scam tools that have arrived in early 2026, spanning social platforms, phone operating systems, and the financial sector. The tools differ in scope and method, but they share a design principle: each intervenes while a fraud is in progress rather than flagging it after the damage has occurred.

The digital heist

Scammers extracted over $1 trillion from victims worldwide in 2024, according to the Global Anti-Scam Alliance and Feedzai. The sum rivals the GDP of the Netherlands. In the United States, the FTC reported $12.5 billion in fraud losses for the same year, a 25% increase over 2023, driven by a sharp rise in the share of people who reported losing money rather than by a rise in the number of reports. GASA’s research found that nearly half of consumers globally now encounter a scam attempt at least once a week, a frequency that has turned fraud from an occasional misfortune into a routine feature of digital life.

Generative AI has handed scammers a production toolkit. Convincing deepfakes, cloned voices, and forged identity documents can now be manufactured at volume, enabling fraud operations to scale in ways that human impersonation never could. Deloitte has projected that gen AI could fuel $40 billion in US fraud losses by 2027, roughly triple the $12.3 billion recorded in 2023. For decades, fraud detection relied on rule-based systems that flagged transactions after money had already moved. Analysts set manual thresholds on transaction amounts or purchase locations and reviewed flagged matches after the fact. Against scams that AI has accelerated to execute in minutes, retrospective detection arrives too late.

The banking sector’s response has moved toward speed. Feedzai’s 2025 report found that 90% of financial institutions now use AI for fraud detection, with two-thirds having adopted it in the past two years. Instead of reviewing transactions days after execution, these models analyze them during execution, flagging suspicious patterns as money moves and reducing the false positive rates that plagued rule-based systems. American Banker’s 2026 survey found that 53% of bankers ranked fraud detection as the highest-impact application of AI, ahead of back-office automation and customer service. The financial sector has proven the approach at institutional scale. The consumer-facing tools arriving this month extend the same principle to the phone in a user’s pocket and the platforms on which scammers make first contact.

Always on guard

Meta launched new anti-scam tools on March 11 across WhatsApp, Facebook, and Messenger, each designed to interrupt a scam before the user engages. On WhatsApp, the update warns users when someone attempts to link their account to an unfamiliar device, a tactic that scammers use to hijack accounts by tricking victims into sharing a QR code or device-linking code. On Facebook, new alerts flag friend requests that exhibit patterns associated with fraudulent accounts. The tools supplement broader enforcement efforts. Meta says that it removed over 159 million scam ads globally in 2025, catching 93% proactively, and that it aims to have 90% of ad revenue flowing from verified advertisers by the end of 2026.

British carrier O2 has aimed its AI directly at the scammer. dAIsy, launched in late 2024 and trained with the help of YouTuber and scambaiter Jim Browning, is a conversational AI that mimics an elderly woman. When scammers call, dAIsy engages autonomously, rambling about her cat Fluffy, her knitting, and her family, while feeding callers fabricated bank details. O2 says that the bot has kept scammers on the line for over 40 minutes at a stretch. After more than 1,000 calls, dAIsy has also yielded intelligence about how scam operations function. Most callers follow scripts from organized call centers, impersonate banks or government agencies, and share knowledge across teams.

Meta’s tools and O2’s bot operate at different points in the scam’s lifecycle, but both activate while the fraud is in progress. Meta shields the target before engagement begins. O2 lets the engagement happen and redirects it toward a decoy, converting the scammer’s time into a wasted resource. Every minute that dAIsy occupies is a minute unavailable for targeting real victims.

Caught in the act

Google’s Scam Detection activates during a live phone call or text conversation already underway. The feature runs Gemini Nano, Google’s on-device AI model, on the phone’s own hardware, analyzing conversation patterns in real time without sending audio to any external server. If a caller claims to represent a bank and urges the user to transfer funds, the feature displays a warning that banks do not make such requests, accompanied by audio and haptic alerts. Google Messages has expanded the detection to cover subtler conversational scams, including job-offer fraud and romance-baiting schemes that build trust over days or weeks of messaging before making a fraudulent request.

The March 2026 Pixel Drop expanded Scam Detection to France, Italy, Spain, Mexico, Germany, and Japan, adding to earlier availability in the US, UK, Australia, Canada, India, and Ireland. A more consequential expansion arrived in February, when the feature appeared on Samsung Galaxy S26 phones for the first time. Scam Detection had been exclusive to Pixel devices since its launch. The move to Samsung, which ships far more handsets globally, shifted the feature from a Pixel-exclusive experiment to a consumer tool with meaningful reach. Google has signaled that it plans to bring the capability to additional manufacturers.

Google’s privacy architecture places all audio processing on the device itself. No call audio or transcription is stored on the phone, sent to Google’s servers, or shared with third parties, and the feature is off by default. Users must choose to enable it, either through the Phone app settings or on a per-call basis. The design did not prevent criticism when the feature was first announced in 2024. Signal president Meredith Whittaker called it “incredibly dangerous,” arguing that the precedent of AI monitoring live calls could extend to surveillance of reproductive care, LGBTQ resources, or whistleblowing. The on-device architecture addresses the data-transmission concern directly. How live-call analysis develops as the technology matures will depend on policy and oversight choices that extend beyond any single company’s default settings.

The eternal arms race

The defensive tools arriving this spring all depend on the same generative AI capabilities that made the scam crisis worse. The models that recognize fraud patterns in a live phone call share architectural DNA with the models that generate convincing deepfakes and cloned voices. Scammers are adapting in parallel, deploying AI to generate novel scripts, rotate synthetic identities, and probe platform defenses, creating pressure on every detection system to keep pace with an adversary that has access to the same underlying technology.

The convergence on real-time intervention across banking, platforms, and devices marks a structural shift in where fraud defense operates. The point of interception has moved from post-hoc cleanup to the moment of the attempt. Early results are encouraging: banks report reduced losses and fewer false positives, and Google and Meta are expanding their tools aggressively. The durability of the approach will depend on iteration speed. Every detection pattern that works today becomes a pattern for scammers to learn to evade tomorrow.