Fighting Fraud with an AI Grandma - Ben Hopkins and Morten Legarth from faith @ VCCP

To fight phone scammers, they created an AI grandma. Her secret weapon? Turning technical lag into the charming, natural pauses of an elderly person.

#1about 4 minutes

Embracing generative AI with an optimistic outlook

Faith @ VCCP was created to proactively explore generative AI as an accelerator for human creativity rather than succumbing to industry fear.

#2about 2 minutes

Creating an AI grandma to fight phone scammers

The project Daisy was inspired by scam baiting YouTubers to create an AI that could engage thousands of scammers simultaneously, preventing them from targeting real people.

#3about 1 minute

Shifting from scam awareness to direct disruption

Traditional educational campaigns about scams lack cut-through, prompting a shift to proactive "story doing" to disrupt scammers directly.

#4about 3 minutes

Blending technical development with creative world building

The bot was developed in-house by a team that combined technical skills for the workflow with creative direction for world-building and character design.

#5about 4 minutes

Choosing a grandma persona for creative and technical reasons

A grandma persona was chosen to create a compelling story by flipping stereotypes and to technically mask AI response latency as natural pauses.

#6about 2 minutes

Navigating GDPR and the ethics of scam baiting

The team consulted with legal experts to navigate GDPR and the ethical rights of scammers, ultimately proceeding due to the criminal nature of the activity.

#7about 2 minutes

Designing a believable persona beyond generic AI

To make the character believable, the team used detailed world-building prompts and iterative training, even incorporating emergent traits like a pet cat named Fluffy.

#8about 7 minutes

Cloning a voice from natural conversational data

Instead of using generic voice libraries trained on narration, the team cloned a real grandmother's voice from hours of natural conversation to achieve a believable tonality.

#9about 3 minutes

The successful first live test with a real scammer

The first live test was a major success, with the AI keeping a scammer engaged for 15 minutes by seamlessly pivoting the conversation to a bird in her garden.

#10about 4 minutes

Considering the energy consumption of a voice-based AI

The project's energy consumption was deemed relatively low as it is voice-based and primarily involves fine-tuning models on consumer-grade hardware, not large-scale pre-training.

#11about 3 minutes

Measuring success through awareness and disruption

The project's success was measured by its ability to waste thousands of scammers' time and by driving a 44% increase in the use of the 7726 scam reporting service.

#12about 4 minutes

Future applications and positive public reception

Future plans involve using the AI to gather intelligence on scammer tactics, building on the project's positive reception from both the public and security professionals.