Anti-Fraud Projects Are Teaching AI Bots to Play Victims and Wear Scammers Down

Scammers have spent years cold-calling from scripts, counting on a live, confused person at the other end. Increasingly, they now reach a neural network instead. Anti-scam initiatives are deploying AI bots with realistic voices and speech patterns so that criminals take them for real victims. One of the best-known early examples is Daisy from Virgin Media O2: a "granny" who, according to the operator, keeps a caller on the line for up to 40 minutes. Every one of those minutes is lost to the scammer, and for defenders it turns into data.

How the Trap Works

A typical decoy is built from three components: speech recognition, a language model and voice synthesis. The model is given a role, for example an elderly, talkative and slightly absent-minded person. It neither refuses right away nor agrees right away. The bot asks people to repeat themselves, mixes up digits, wanders off into stories about grandchildren and the cat, "looks for the card" and "can't find its glasses." The goal is not to win an argument but to keep the conversation going as long as possible and make it look convincing.

The main advantage over a human volunteer is that a bot doesn't get tired or irritated. It also doesn't give itself away through emotions, which are hard to hide in the fiftieth minute of a pointless conversation. A human can't do that.

Who Is Already Doing This

The idea isn't new, but language models have made it cheap and scalable. Below are projects known from open sources. It is best to check the specific figures for each against primary sources.

• Daisy (Virgin Media O2, UK): a voice-based AI "granny" that, according to the operator, keeps scammers on the line for up to 40 minutes.
• Apate (Macquarie University, Australia): a research project with AI personas that answer scammers' calls and messages.
• Re:scam (Netsafe, New Zealand): a user forwards a suspicious email, and the bot corresponds with the sender as the "victim."
• Jolly Roger Telephone Company (US): a paid service with bots aimed at telemarketers and scam calls.

All of them share one core metric: contact duration. The longer it is, the more the attack costs. The second metric is the quality of the collected data: phone numbers, names of "banks," scripts and payment details for transfers.

Economics and Limitations

A scam call center runs on an expensive resource: operators' time. A bot that handles hundreds of conversations in parallel for the cost of compute breaks that economic model. A scammer can't quickly tell that they're wasting time until their own filter kicks in.

The approach has weak spots. First, criminals themselves use generative AI, so in some cases a bot ends up talking to a bot, and the value of stalling drops. Second, the operator on the other end may not be a willing criminal: schemes in Southeast Asia often rely on people who were taken there by force. Wearing down such "employees" is ethically harder than going after the organizers. Third, legal questions remain: recording conversations, storing voices and data, and false positives that hit real people.

What to Do in Practice

For ordinary users, these traps are no substitute for basic hygiene. If someone calls "from the bank" or "from the police," end the call and call back on the official number. Suspicious emails and messages are worth forwarding to the relevant services in your country, if they exist: some of them feed exactly these systems.

Telecom operators and banks should treat decoys as a source of threat intelligence. Numbers, scripts and wallets collected by bots can be quickly passed into blocklists and anti-fraud rules. Security teams should think through the rules in advance: what the bot is allowed to say, what data to retain and when to hand material over to law enforcement. Without this, the trap will quickly become a legal problem for its owner.