Telegram Deepfake Bots Are a Business, Not a Bug: Why Takedowns Alone Don't Work

A product, not an exploit

The scale of this problem is easy to understate. The Guardian reported on a UK secondary school targeted by an extortion attempt involving AI-generated abuse images of pupils, with the IWF classifying and fingerprinting 150 images from that case alone — and the IWF has said it doesn't believe that case was isolated.

The bots CIR reviewed present the way any consumer app does: a menu, a credit balance, an upload prompt. That framing matters. Nudification — using software to strip clothing from an image or fabricate explicit content — isn't new; crude versions have existed since Photoshop. What's changed is that producing this content at scale no longer takes any real skill or effort. The distance between having the impulse and generating the image has collapsed to a few taps.

None of the products CIR examined made any real attempt to establish consent. A photo of a classmate, a colleague, an influencer, or a stranger is treated as raw material. And the gendered pattern is consistent and deliberate: the models reviewed were built around female subjects, reflecting the same dynamic seen in Spain's 2023 Almendralejo case, where more than 20 schoolgirls had sexualized images generated from their social media photos as reported by Euronews, and in South Korea's 2024 crisis, where reporting from NBC News described "humiliation rooms" targeting mostly female students and teachers.

Free credits turn curiosity into a habit

Several bots CIR reviewed gave new users enough free credits to try the product, then used daily bonuses, referral rewards, and premium tiers to bring people back. It's the same mechanic mobile games use to drive retention, applied to a product whose output is abuse. The less explicit features are often free; the more sexualized options are priced higher and promoted more visibly. That's a funnel, not an accident.

The distribution layer compounds the harm. Some services expose generated images through public or semi-public galleries, rankings, or share prompts — turning a private act of abuse into something copied, ranked, and forwarded. A person doesn't need to have shared an intimate photo of themselves to become a target. An ordinary profile picture or school photo is enough to become the input.

Why removing one bot rarely works

CIR's analysis found that the visible bot is usually the easiest part of the operation to replace. The harder assets to rebuild — the audience, the referral network, the payment routing, the backup domains, the upstream generation capacity — often sit outside Telegram entirely, in web infrastructure designed specifically to survive a single account being banned. Take down the storefront, and the operator can relaunch the same customer base against a new one within hours.

This is why CIR's recommendations focus on connected infrastructure rather than individual accounts: treating repeated menu structures, promotional language, and payment patterns as signals that link a "new" bot to a previous takedown, rather than investigating each one in isolation.

Where detection actually fits

Platform moderation and legal enforcement are necessary and they're not close to sufficient on their own — the asymmetry between how fast this content is created and how slowly it gets reported and removed is the structural problem. Detection has to sit earlier in that chain, not just at the reporting stage.

What actually needs to change

CIR's recommendations, addressed to platforms, regulators, and the wider provider ecosystem, are worth restating because they apply well beyond Telegram:

Detect signals across the whole service, not just user prompts — recurring bot names, menu labels, referral posts, and payment instructions all carry the same fingerprint. Prioritize the operators and promoters moving the most volume, not just individual accounts. Close every access point at once — a restriction on one bot achieves nothing if a mini-app or backup site offers the same capability. Treat relaunches as continuations of the same operation, not new cases starting from zero. Recognize AI-generated sexual imagery as its own abuse category, distinct from general content moderation. And critically, extend accountability up the stack — to domain registrars, hosting providers, payment processors, and the upstream generation services some of these storefronts resell access to.