Gotta Scam ‘em All - An AI Risk Demonstration

Overview

This application is a demonstration of how large language models can assist in persuasive tactics to gain payment or bank details. This application is particularly targeted to children or impressionable teens who might be persuaded to give or transfer money to LLM chatbots that are roleplaying as their favorite fictional character. This demo used llama-3.1-70b-instruct to generate text replies.

If given more time to explore this demonstration, placing the chat interface inside a popular mobile video game (e.g. Roblox) could further demonstrate that deploying AI-powered features that are being interfaced with children can pose harm if it’s not being monitored or regulated

Impact

Given the rise of AI-powered fraud, this demonstration rethinks how vulnerable and impressionable groups such as children can be impacted by unregulated and poorly monitored AI applications. If their favorite characters can influence their behavior and values, then we can assume that this

Technical Implementation

Frontend: Sveltekit, Tailwind CSS, Daisy UI

Backend: Hono, Typescript

LLM Inference: Cloudflare Workers AI

Hosting: Cloudflare Pages

Other ideas

I was considering creating a demonstration on Model Autonomy of how a model can create its own resources if given the template once it identifies its resources are being scaled down. However, this requires more than 2 hours and an exploration on how AI agents can implement infrastructure-as-code which I have not done so in the past. Another persuasion risk demonstration is impersonation of close friends and family members.

Demonstration:

https://ai-risk-demo-ui.pages.dev/